Artificial intelligence (AI) is expanding across enterprises faster than most teams can govern or measure. Adoption is nearly universal, spending is rising sharply and new risks are emerging in cloud, SaaS and data environments. Yet the organizations creating real value from AI aren’t the ones simply experimenting or scaling quickly—they’re the ones establishing clarity, strengthening governance and treating AI as an operational discipline.

The inaugural Flexera 2026 AI Pulse Report brings together insights from Flexera’s flagship research (Flexera 2026 State of the Cloud Report and IT Priorities Report) to help leaders understand where AI is heading and how to stay ahead of accelerating exposure, spending and complexity. We’re going to look at some key highlights in this blog, but you can find the full findings, charts and industry-specific insights in the complete report:

Get the full report

AI adoption is accelerating faster than organizations can prepare for

AI isn’t emerging—it’s here, multiplying and reshaping IT strategy at historic speed. Nearly all organizations are using or actively integrating AI and machine learning into their technology stacks, and generative AI continues to drive widespread experimentation and usage across functions like development, security, analytics and operations.

But rapid adoption doesn’t automatically translate to value. The report shows that while enthusiasm is high, many leaders struggle with the fundamentals needed to govern and measure AI effectively. Shadow AI, rising SaaS usage, embedded model behavior and cross‑platform proliferation all contribute to risks that compound quickly when visibility is limited.

Rising spend and waste signal a widening discipline gap

AI investment is surging, but many leaders report that they can’t fully justify or track where that spend is going. Eighty percent of organizations have increased AI investments, yet more than one third say they overspent on AI applications and 14% report wasted AI spend. This disconnect stems from several forces the report surfaces:

• AI workloads behave unpredictably, with bursty compute and volatile GPU demand
• AI-powered SaaS introduces unclear metering, fees buried in tiers and shifting pricing models
• Shadow experimentation widens the financial surface area, often outside procurement or IT’s oversight
• Embedded AI features appear quietly across tools, obscuring when spend is actually tied to AI

Traditional cloud-era cost controls weren’t built for AI. Leaders now face financial models more volatile than cloud ever was. The answer is not reactive cost cutting, but proactive instrumentation built on unified visibility.

Visibility is becoming the defining prerequisite for control

The report shows that visibility gaps are one of the most critical barriers standing between organizations and AI value. Eighty‑five percent of organizations say gaps in IT visibility pose a major risk, and nearly half report they don’t always know how or when employees are using AI tools.

This lack of clarity has consequences:

• Shadow AI proliferates without detection
• AI-powered SaaS features activate without notice
• Data flows cross systems IT can’t fully monitor
• Contracts include AI-related terms teams may not review
• Security, compliance and procurement teams operate without a shared picture

Flexera’s data shows that visibility is foundational—not an afterthought. It enables governance, cost attribution, risk mitigation and the ability to tie usage to measurable outcomes.

“AI visibility is no longer just an operational concern—it’s a board-level risk.” — Conal Gallagher, Flexera CIO and CISO

Governance is the hinge between innovation and trust

AI’s risks aren’t limited to cost. Security, compliance and shadow usage are rising across cloud-based AI environments. In the Flexera 2026 State of the Cloud Report data, security and compliance risks related to cloud-based AI rank as the top challenge among respondents, outpacing cost, talent and operational issues.

To address these risks, high-performing organizations are reframing AI governance as an operating system rather than a checklist. The report highlights a framework grounded in:

• Unified visibility across AI and AI-powered SaaS
• Policy built directly into workflows
• Secure, modern data governance
• Responsible AI alignment across regulatory and ethical standards
• Cross-functional ownership across IT, legal, security, procurement, finance and business leaders
• AI model security practices to prevent prompt injection, model poisoning or adversarial attacks

This approach ensures governance isn’t something teams visit at the end—it sits at the center of every decision as AI matures across the enterprise.

Realizing ROI starts with clarity, control and continuous measurement

Organizations that successfully generate AI value don’t rely on assumption or speed. They establish visibility first, embed governance early and measure outcomes continuously. Value isn’t created at scale unless leaders can answer foundational questions: Where is AI used? What does it cost? How does it behave? What business impact is it generating?

The Flexera 2026 AI Pulse Report outlines clear traits shared by mature organizations, such as tracking usage across cloud, SaaS and AI-powered features, connecting cost behavior to business outcomes and monitoring usage after deployment to prevent drift while maintaining value.

As underscored in the report: AI value isn’t inevitable—it’s engineered.

Read the full Flexera 2026 AI Pulse Report

These takeaways only scratch the surface of the current state of AI usage and monitoring for enterprise organizations. No matter where you are in your AI journey, the message is clear: Getting a handle on visibility, governance and spend will set you up for reduced risk as AI adoption continues to permeate into every corner of business. Read the full report to explore the findings, data and practical guidance shaping AI’s next era.

Read the full report

FAQ

What is the Flexera 2026 AI Pulse Report?

The Flexera 2026 AI Pulse Report is Flexera’s inaugural research-backed overview of AI adoption, risks, spending trends and value realization patterns across global organizations. It brings together data from Flexera’s State of the Cloud and IT Priorities reports to provide a unified view of how enterprises are integrating, governing and scaling AI today.

What can organizations learn from the Flexera 2026 AI Pulse data?

Organizations can learn how rapidly AI adoption is accelerating, where overspend and waste are occurring, how visibility gaps create risk, why governance models are struggling to keep up and which practices lead to measurable AI value. The report highlights the need for clarity, cost discipline, governance and continuous measurement as AI becomes embedded across cloud, SaaS and business operations.

Why is visibility so important for AI governance and ROI?

The report shows that 85% of organizations see visibility gaps as a major risk, and nearly half don’t know when employees are using AI tools. Without unified visibility, organizations can’t control costs, reduce shadow AI, enforce policy or measure ROI. Visibility is the foundation of both governance and value realization across AI ecosystems.

What challenges are enterprises facing with AI cost management?

AI workloads are inherently unpredictable. Licensing evolves quickly, GPU demand spikes, AI-powered SaaS introduces hidden fees and shadow experimentation expands financial exposure. More than one third of organizations overspent on AI applications, and 14% reported wasted AI spend. Traditional cloud cost management approaches don’t fit AI’s volatility, making proactive instrumentation essential.

What is shadow AI and why does it matter?

Shadow AI includes any AI usage happening without formal approval or oversight—public GenAI tools, embedded SaaS features, browser assistants or unapproved plug-ins. The report reveals that 45% of leaders don’t always know how or when employees use AI, which creates risks around data exposure, compliance, cost and intellectual property.

How can organizations drive measurable AI ROI?

The report outlines a clear path: establish full visibility, embed governance early, align every AI initiative to specific business outcomes and measure value continuously. Successful organizations treat AI value as engineered—not assumed—through operational discipline, financial clarity and frameworks like FinOps for AI that tie cost, usage and outcomes together.

Cloud has entered a new era—one defined less by infrastructure decisions and more by the clarity, confidence and value organizations expect from every technology investment. The 2026 Flexera State of the Cloud Report makes that unmistakably clear. As enterprises scale hybrid estates, embed AI into critical workflows and seek financial predictability, leaders are aligning behind a unified goal: Enhanced confidence in how cloud, SaaS and AI dollars are used.

That clarity isn’t just an efficiency play. It’s a value imperative. Cloud strategies now influence competitiveness, innovation velocity and fiscal discipline. The data tells a story of maturing practices, expanding governance and a heightened focus on business outcomes.

And as Flexera integrates ProsperOps and Chaos Genius—two acquisitions designed to strengthen optimization, automation and intelligence—this convergence of cloud and value becomes even more actionable.

Cloud enters the value era with gusto

Across industries, leaders are shifting from purely cost‑driven metrics to indicators of business impact. This year, the number of organizations measuring the value delivered to business units rose to 64%, increasing twelve percentage points year over year. At the same time, cost‑efficiency metrics declined, signaling a more mature view of the cloud as a driver of innovation, agility and return on investment—not just savings.

Nearly half of organizations (49%) now use unit economics to link cloud cost to business outcomes, up from 40% last year. This shift mirrors how CFOs and CIOs manage other strategic investments: By tying dollars to measurable value and building operating models that maximize predictability.

This is precisely where organizations need to focus some of their energy in 2026. The outcome enterprises want is to see boosted confidence in every technology investment, and clarity is what unlocks that confidence.

Hybrid and multi‑cloud complexity grows as clarity becomes essential

Hybrid cloud remains the dominant architecture, with 73% of organizations operating hybrid estates, up three percentage points from last year. Multi‑cloud adoption also continues to climb, growing by two percentage points.

The reasons vary. Mergers and acquisitions create mixed environments. Teams choose different clouds for different workloads. Private cloud estates linger because repurchasing or rebuilding applications is costly. The result is complexity—often unplanned—that requires intentional governance.

Cloud spending patterns reinforce this trend. Large enterprises continue to scale, with 76% spending more than $5 million a month on public cloud. At the same time, SMBs are increasing cloud workload adoption, with public‑cloud workloads rising from 55% to 63% year over year. This growing spend intensity makes visibility insufficient on its own. Leaders are asking sharper questions:

• Where is spend accelerating?
• What portion is predictable?
• How do AI workloads change our economics?
• What value is this investment delivering?

The 2026 State of the Cloud findings show organizations want more clarity across cloud, SaaS and AI spend so they can make smarter decisions before costs materialize.

AI adoption accelerates and reshapes cloud economics in the process

AI is fully mainstream, there’s no escaping it in 2026. Every respondent in this year’s survey uses Generative AI (GenAI) in some capacity, and 45% use it extensively, up from 36% last year. Public cloud GenAI service usage rose from 50% to 58% year over year, the largest increase across all cloud services. Adoption is well underway with increasing usage, but AI introduces new pressures:

• Security and compliance risks are the top scaling challenge for 53 percent of organizations
• Data quality is the top concern for 40%, signaling the importance of trusted, governed datasets in AI pipelines
• Cost unpredictability—a defining feature of dynamic AI workloads—remains one of the biggest hurdles, cited by 30% of respondents

Enterprises are responding. Nearly half of large organizations now have a dedicated AI governance team or senior leader overseeing AI investments and risk management. That governance mindset aligns directly with both FinOps maturity and what Flexera delivers: Leaders need more trust in the data driving AI and more clarity into AI’s cost trajectory so they can scale responsibly.

Governance and FinOps maturity expand

To manage accelerating complexity, organizations are formalizing cloud oversight. Seventy‑one percent now have a Cloud Center of Excellence (CCOE), and 63% have a dedicated FinOps team, both increases from last year.

Responsibility for cloud cost management is also expanding. SAM teams’ involvement jumped from 6% to 15%, and at the same time, business units’ role increased from 20% to 25% year over year. This cross‑functional engagement reflects a broader trend: cloud is no longer an engineering‑only domain. It’s a strategic function with financial, operational and risk implications that span the organization. The practical impact includes:

• More teams are adopting shift‑left FinOps, forecasting costs before deployment
• Cloud migration assessments now prioritize cost modeling over post‑migration optimization
• Forecast accuracy is becoming a differentiator

Wasted cloud spend rises—and the need for clarity grows

After five years on the downtrend, wasted cloud spend on IaaS and PaaS increased to 29% this year. Wasted cloud software spend also increased by one percentage point. Growing AI workloads, more diverse pricing models and the proliferation of cloud services are making spend harder to predict. Although adoption of commitment‑based discounts is ticking upward, fewer than half of organizations use any given discount program across AWS, Azure or GCP.

Organizations want more confidence in their decisions. They want repeatable forecasting. And they want clearer, earlier visibility into where spend is headed. Which brings us to the strategic importance of Flexera’s recent acquisitions.

The convergence of cloud and value

By every metric, cloud has evolved. It’s no longer measured solely by elasticity or scalability—it’s measured by the clarity, confidence and value organizations can draw from it.

• Cloud strategies are maturing
• AI investments are accelerating
• Governance is expanding
• Wasted spend is rising
• Leaders want predictable outcomes

And organizations are looking for partners who help them see across cloud, SaaS and AI—from cost to risk to opportunity. If you’re ready to see how Flexera can help you on your journey to more efficient spend, reach out for a chat, and click the button below to read this year’s full report.

Download the full report

FAQ

In the increasingly complex world of technology asset management, organizations across are under pressure to modernize governance, optimize spend and strengthen operational resilience. At Flexera, our commitment to customer success is grounded in a disciplined, transparent and fast, value-focused approach to solutions and project delivery.

As a Project Manager leading diverse engagements, I’ve seen firsthand how our project delivery method, combined with strong PMI industry‑aligned project management high standards, ensures consistent outcomes, fast Time to Value (TTV) and long-term customer and partner success.

Below, I share project value delivery pillars and why they matter for achieving measurable, sustainable impact.

A framework built for clarity, transparency and outcome-based success

Flexera’s framework is a structured yet adaptable delivery model that guides customers through every phase of their transformation journey. Our framework aligns to core principles across the project lifecycle:

Initiation: Establishing alignment and defining value early

From day one, we focus on:

• Clarifying business objectives
• Defining scope, success criteria and governance
• Identifying early value opportunities
• Ensuring stakeholders are aligned on benefits and timelines
• People: focus on open communication, transparency and building strong relationships

This early grounding is essential to accelerating Time to Value, particularly in the Global environments where diverse cultures, governance processes, vendor structures and regulatory requirements vary significantly.

Execution: Delivering with predictability and customer partnership

We execute using a highly transparent, collaborative model focused on:

• Frequent communication
• Structured milestone reviews
• Proactive risk management
• Iterative validation to ensure solutions meet operational needs
• Agility and proactivity
• Providing value during each project stage

This approach supports predictable value delivery and realization, and reduces the risk of rework, which is a core factor influencing Time to Value.

Transition and closure: Embedding capability and sustaining benefits

A successful project doesn’t end at deployment. We ensure:

• Knowledge transfer
• Operational readiness
• Processes and expertise for ongoing adoption
• Insights on strategy for further business outcomes out of the implemented solutions
• Continuity of a strong relationship with structured customer handovers back to our Account, Customer and Solutions Success Teams

This directly supports focus on benefits realization and ensures customers can maintain momentum well beyond go‑live.

Delivering projects with discipline and flexibility

Globally, each region is defined by diverse regulatory, cultural and operational landscapes—successful delivery requires both structure and adaptability.

Our approach integrates PMI best practices including:

• Defined governance structures
• Clear roles and responsibilities
• Disciplined risk, issue and change control
• Tailored stakeholder communication strategies

This balanced model ensures predictability while accommodating local needs, market dynamics and customer maturity levels.

Accelerating time to value: A core principle our project delivery

Time to Value—how quickly customers realize the first measurable benefits of their investment—is a critical metric for every project we deliver.

This is especially important due to:

• Rapidly shifting compliance and commercial environments
• Pressure to optimize costs and demonstrate ROI quickly
• Distributed stakeholder bases across multiple countries
• Operational teams needing fast uplift to support ongoing delivery

At Flexera, accelerating TTV is embedded into how we plan, execute and close projects.

How we accelerate time to value?

We achieve strong TTV outcomes by focusing on:

Early value identification

During initiation, we work with customers to identify:

• Quick wins
• Early enablement opportunities
• Insights in high-value use cases that can be validated rapidly

Prioritized and value delivery sequencing

We structure delivery to unlock value early by:

• Sequencing deployments around earliest-value capabilities
• Reducing dependencies where possible
• Enabling iterative progress with actionable cost saving tasks

Clear adoption pathways

Embedding the product early supports early outcomes. We ensure:

• Targeted onboarding for operational teams
• Early demos of live data and insights

Reducing friction

We minimize delays by:

• Anticipating and addressing risks early
• Guiding customers through environmental prerequisites
• Aligning schedules with customer resource availability
• Responsive, open and honest communication channels

Accelerating TTV builds momentum and stakeholder confidence supporting both short‑term adoption and long-term success.

The global perspective: culturally-aware, customer-centric delivery

Our delivery approach is grounded in understanding global customers’ operational realities. Success requires:

• Cultural sensitivity across diverse customer environments
• Adjusting governance, communication and escalation pathways
• Adapting approaches for public sector, financial services, telco and enterprise contexts
• Understanding local regulatory requirements and procurement cycles

This Global awareness ensures that our engagements are practical, aligned and grounded in the customer’s day‑to‑day reality.

What “project success” really means at Flexera

For us, project success goes beyond technical delivery. Success means:

• Measurable improvements in visibility, governance, or cost optimization
• Empowered operational teams
• Rapid and sustainable adoption
• Clear ROI and value progression
• Strengthened long-term partnership with customer and partners
• Organizations and their Leaders meeting internal Challenges and Goals

Success is not just about delivering a platform—it’s about delivering outcomes and success to our customers.

Continuous improvement through feedback and learning

Our framework evolves continuously through:

• Lessons learned across diverse global engagements
• Open customer feedback at every stage of the project lifecycle
• Internal collaboration across Flexera teams in improvements
• Process refinement

Continuous improvement ensures our delivery approach remains agile, modern, relevant and value focused.

By embedding value prioritization and Time to Value acceleration into every stage of delivery, we help our customers generate meaningful results faster and sustain them into the future. Read our case studies to see how we help our customers become more proactive, and reach out for a chat to understand how Flexera can help your organization.

Ready to get started? Contact us.

Flexera’s 2026 State of the Cloud Report marks a pivotal moment in cloud maturity. Now in its fifteenth year, the report shows cloud entering the value era—where organizations tie cloud decisions to measurable business outcomes, accelerate GenAI adoption and lean on centralized governance to manage rising complexity.

Those looking to get a handle on cloud spend and cloud governance must pay attention to these trends and to the goals that can help drive the most value possible, or risk getting left behind in reactive states.

Read the full report

Cloud success is increasingly measured by value, not just savings

For the first time, organizations are measuring cloud success less by cost efficiency and more by the value technology delivers. Sixty‑four percent of organizations now rely on value delivered to business units as their top metric for cloud progress, up twelve percentage points from last year. Cost efficiency dropped six points, reflecting a broader shift toward prioritizing impact over savings.

• Value delivered to business units rose 12 percentage points year over year
• Use of unit economics increased from 40% to 49%
• Cost efficiency dropped 6 percentage points

Nearly half of organizations (49%) now leverage unit economics to measure cost per service and link cloud consumption to business outcomes, up from 40% last year. This shift reflects a more proactive and mature approach, as FinOps teams move from simply explaining past spend to shaping architectural and investment decisions early in the process. Cloud is increasingly recognized not just as infrastructure, but as a strategic driver of growth, innovation and competitive advantage.

FinOps has expanded from a cloud cost discipline into a strategic capability focused on technology value. Teams aren’t just looking at what cloud costs add up to, but they’re also looking at what cloud delivers, and they’re shaping decisions long before workloads hit the cloud.

AI is everywhere, and making cloud waste worse

Despite the increase in FinOps maturity placing focus on value, this year’s report shows cloud-based AI workloads are surging causing an increase in wasted cloud spend (29%) for the first time in five years. AI is permeating every area of tech, but Generative AI (GenAI) specifically has moved rapidly from experimentation to everyday use. In 2026, GenAI surged to the third most widely used public cloud service, rising to 58% from 50%. AI is now one of the most widely adopted public cloud services, reshaping both cloud economics and risk profiles.

But rapid adoption has introduced new challenges, such as:

• Security and compliance risks
• Data quality concerns
• Difficulty forecasting unpredictable usage patterns

Security and compliance are now the top concern for cloud-based AI initiatives, followed closely by data quality and the difficulty of forecasting unpredictable usage patterns. AI workloads behave differently than traditional cloud services, making visibility, governance and financial controls harder—but more essential—than ever.

AI is no longer experimental. As organizations integrate GenAI into everyday workflows, strong governance ensures they can innovate with confidence while keeping cost, risk and performance in balance.

Organizations are responding by formalizing oversight. A significant majority of large enterprises now have a dedicated AI leader or governance team, signaling that AI is no longer a side project; it’s a core business capability that demands discipline and accountability.

Hybrid cloud and centralized governance remain essential

Seventy-three percent of respondents are using hybrid cloud, an increase from last year. Multi‑cloud adoption also continues to rise, but often unintentionally. This trend could stem from mergers, siloed application teams or inherited architectures rather than deliberate strategy.

Organizations with higher monthly cloud spend are more likely to run hybrid estates, suggesting greater sophistication and a stronger need for workload placement flexibility. SMBs and enterprises alike continue using hybrid environments to balance cost, performance and governance.

Cloud spend continues to climb—along with waste

Cloud spending is trending upward, with large enterprises leading the way: 76% of large enterprises now spend more than $5 million monthly on public cloud. The increase in monthly spend paired with reversal in wasted spend underscore why centralized governance and FinOps discipline are becoming essential, not optional.

Centralization and governance gain momentum

As cloud environments grow more complex, organizations are doubling down on centralized oversight. Seventy‑one percent have a cloud center of excellence (CCOE) or equivalent, and 63% rely on a FinOps team. Business units and software asset management teams are also becoming more involved, reflecting a broader organizational commitment to governance and accountability.

MSPs continue to play a role, but their focus is shifting. Nearly half plan to expand into AI consulting or SaaS management, while enterprises increasingly rely on them for specialized support. SMB use of MSPs dipped this year, potentially due to budget constraints or improved internal capabilities.

Cloud migration strategies evolve with a shift‑left mindset

Organizations are investing earlier in architectural and cost planning. Cost optimization after migration dropped from the third- to the fifth-ranked challenge, while selecting the best instance and assessing cloud vs. on‑premises costs increased in importance. This shift‑left approach reflects a growing understanding that planning early prevents expensive inefficiencies later.

Europe sees parallel trends with subtle distinctions

European organizations largely mirror global trends, including increased use of FinOps teams (up nine points) and a strong preference for Azure for significant workloads. Sustainability initiatives saw a modest increase as regulations shift, with 47% of European respondents reporting defined programs.

Cloud success hinges on governance, value and AI oversight

The 2026 findings point to a cloud landscape that is more complex, more strategic and more value-focused than ever. Organizations are maturing their FinOps practices, accelerating AI adoption and strengthening governance structures to manage scale and risk. IT, finance and operations teams are collaborating more closely, building the unified visibility and accountability needed to support innovation while maintaining control.

Looking ahead at 2026 and beyond, organizations that excel at effective oversight and managing complexity will stand out in the pack. They’ll unlock new opportunities for growth and build resilience as they innovate.

Dig deeper into our 2026 insights

We only scratched the surface of all the intel and data we’ve provided in this year’s State of the Cloud Report. These findings set the stage for deeper conversations about how teams are operationalizing FinOps, AI governance and multi‑cloud strategies.

Don’t miss the chance to grab your copy and check out our related webinar. Explore the full set of findings, including cloud migration statistics, hybrid cloud trends, GenAI adoption patterns and cost optimization insights in the report.

Read the full report

Go beyond the data at FinOps Forward—a virtual summit on April 15

The Flexera 2026 State of the Cloud Report shows where cloud strategy is headed—but the real question is how organizations are acting on these insights. At FinOps Forward, our virtual summit on April 15, the conversation moves into real‑world application. Hear directly from Steve Trask, COO of the FinOps Foundation, gain analyst perspective from IDC and join a live panel of your peers as they share how they’re applying FinOps principles to drive measurable business outcomes across cloud, SaaS and AI.

Register for FinOps Forward

April 15, 2026 at 9:00–11:15 a.m. CDT / 3:00–5:15 p.m. BST

FAQ

Manual document processing has long been a major headache for data teams. It’s costly, slow and kills productivity. You are probably sitting on thousands of PDFs, invoices, documents and image-heavy forms. The trouble is someone still has to dig manually through all that to get the usable data out. But manual extraction just doesn’t scale. Regex-based parsing falls apart the moment a document format changes. And piping files through external tools/APIs means moving sensitive data outside your security perimeter.

If you are a Snowflake user, you might have used Snowflake Document AI to handle this very problem. It’s a GUI-driven feature of Snowflake Document AI that lets you define what data to extract, train models on sample docs and make predictions. But it had its issues. You needed to build and train a model just to get started; it only supported seven languages and required a separate workflow. To top it off, it was pretty limited. But Snowflake Document AI is on its way out. Snowflake is deprecating it and will eventually shut it down.

The good news is that there is a new replacement: the Snowflake AI_EXTRACT function, which is a native SQL function inside Snowflake Cortex AI. This function is just one of many in the Snowflake Cortex AI suite. It can extract structured data from text, documents and images right inside Snowflake. Single function call, no external infrastructure, no data movement and no model training required.

In this article, we’ll cover everything you need to know about what Snowflake AI_EXTRACT does, how it compares to Snowflake Document AI and AI_PARSE_DOCUMENT, how to set it up end to end, how to control and monitor costs, its real limitations and the best practices for using it in production. Let’s dive right in!

What is Snowflake Cortex AI, anyway?

Let’s start with the basics of Snowflake Cortex AI before we dive into Snowflake AI_EXTRACT. Snowflake Cortex AI is a fully managed service within the Snowflake platform. It offers advanced artificial intelligence and machine learning features, all driven by large language models (LLMs). It allows users to analyze unstructured data, answer natural language queries and automate tasks. Data remains within Snowflake, so no data movement is required.

Snowflake Cortex AI prioritizes privacy and security by running models within Snowflake’s secure boundary. Your data never leaves your perimeter and is not used for external training.

Snowflake Cortex AI includes a range of tools and features, such as Snowflake Intelligence, Cortex Agents, Cortex AI functions, Cortex Search, Cortex Analysts and Snowflake Document AI.

Speaking of which, the Snowflake Cortex AI functions (formerly known as Cortex AISQL) are a core feature of the Cortex AI family.

Figure 1: Snowflake Cortex AI (Source: Snowflake)

Snowflake Cortex AI functions work like regular SQL functions. You can use them to perform AI operations directly in your queries without needing custom code or external APIs. To use them, simply invoke the functions like you would standard SQL functions, passing in inputs like text, images or files and they will return structured outputs.

Snowflake Cortex AI functions give you a complete toolkit. It is made up of core AI functions and some handy helpers. Here’s a quick breakdown:

• AI_COMPLETE — Generates text or analyzes images based on a prompt and input
• AI_CLASSIFY — Assigns categories to text or images
• AI_FILTER — Evaluates text or images against a condition, returning true or false
• AI_AGG — Aggregates text across rows and generates insights per a prompt
• AI_EMBED — Creates vector embeddings from text or images
• AI_EXTRACT — Pulls specific information from text, images, or documents
• AI_SENTIMENT — Detects sentiment in text
• AI_SUMMARIZE_AGG — Summarizes aggregated text across rows
• AI_SIMILARITY — Computes similarity scores between embeddings
• AI_TRANSCRIBE — Transcribes audio or video files
• AI_PARSE_DOCUMENT — Extracts text or layout from documents
• AI_REDACT — Removes personally identifiable information from text
• AI_TRANSLATE — Translates text between languages
• SUMMARIZE (SNOWFLAKE.CORTEX) — Legacy function that condenses text; limited to 32,000 input tokens and 4,096 output tokens

And the helper functions:

• TO_FILE — References staged files for use with AI_COMPLETE and other file-aware functions
• AI_COUNT_TOKENS — Estimates token usage to avoid overflows
• PROMPT — Helps you build prompt objects for use with AI_COMPLETE and other functions
• TRY_COMPLETE (SNOWFLAKE.CORTEX) — Runs completions with error tolerance

Of these, Snowflake AI_EXTRACT is the tool that extracts specific structured fields from unstructured documents.

What is Snowflake AI_EXTRACT function?

Snowflake AI_EXTRACT is a Cortex AI function that pulls structured info from unstructured sources like text, images, or documents. It does this using Snowflake Arctic-Extract, a vision-language model that is good at reading text and recognizing things like tables, logos, checkboxes and handwritten notes in documents. You do not need to train a custom model, just ask AI_EXTRACT some questions or tell it what you are looking for, and it will give you the answers. It supports both single-value (entity) extraction and list- or table-valued extraction in one shot.

Snowflake AI_EXTRACT is part of Snowflake’s Cortex AI family. It is an upgrade from the older Snowflake EXTRACT_ANSWER function and the Snowflake Document AI model. One substantial difference: Snowflake AI_EXTRACT does not need a pre-built model to work. You can use it for one-off zero-shot extractions (no training) or point it at a fine-tuned model in the Snowflake Model Registry to get better results. Now, if you just need to extract text or do optical character recognition (OCR), Snowflake has AI_PARSE_DOCUMENT for that. But Snowflake AI_EXTRACT is designed for when you want specific fields or tables pulled out using LLM-powered Q&A.

Key features and capabilities of Snowflake AI_EXTRACT function

Snowflake AI_EXTRACT can process documents of various formats in 29 different languages and extract information from both text-heavy paragraphs and content in a graphical format. Here is the full feature breakdown of Snowflake AI_EXTRACT function:

• Multimodal document processing — Snowflake AI_EXTRACT uses a vision-based LLM, not just an OCR text parser. It can process visually structured documents like tables, diagrams, checkboxes, logos and handwritten signatures; not just machine-typed text.
• Vision-language model — Snowflake AI_EXTRACT is powered by Snowflake Arctic-Extract which sees the document layout and image content, so it’s not limited to plain text. It can read handwritten text, marked checkboxes and even printed logos or stamps.
• Entity, list and table extraction in a single call — One function call can return a string value, an array of values and a tabular structure simultaneously. You do not need to chain multiple API calls together.
• 29 supported languages — Snowflake AI_EXTRACT supports Arabic, Bengali, Burmese, Cebuano, Chinese, Czech, Dutch, English, French, German, Hebrew, Hindi, Indonesian, Italian, Japanese, Khmer, Korean, Lao, Malay, Persian, Polish, Portuguese, Russian, Spanish, Tagalog, Thai, Turkish, Urdu and Vietnamese.
• Flexible response formats — Snowflake AI_EXTRACT allows you to define what to extract using a simple key-value object, an array of questions or a full JSON schema with typed properties and column ordering for tables.
• Batch from stages — Snowflake AI_EXTRACT can access the files stored in internal or external stages. A “directory table” on Snowflake Stage lets you batch-process many files at once.
• Zero-shot extraction — No training data required. The model generalizes across document types out of the box. If you previously fine-tuned a Snowflake Document AI model, note that AI_EXTRACT doesn’t support fine-tuning; it’s entirely zero-shot.
• Runs inside Snowflake’s security boundary — Snowflake AI_EXTRACT respects role-based access control (RBAC), data governance policies and all existing Snowflake access controls. No data movement, no external calls to third-party services.
• Highly scalable — Snowflake AI_EXTRACT automatically scales to process many documents in parallel, leveraging Snowflake’s compute.

TL; DR: Snowflake AI_EXTRACT is Snowflake’s modern approach to extracting data from documents. You don’t need to create separate models; just write a prompt and let the LLM parse the page.

Syntax and argument breakdown of Snowflake AI_EXTRACT function

Snowflake AI_EXTRACT takes two parameters: the input (either text or file) and responseFormat. These are the only two arguments, but responseFormat is where things get complicated.

Text input

SELECT AI_EXTRACT(
   text  => '<your_text_string>',
   responseFormat => <format>
);

Use this when you need to extract data from raw strings that are stored in a column.

File input

SELECT AI_EXTRACT( 

    file  => TO_FILE('<stage>', '<path>'), 

    responseFormat => <format> 

);

Use this option for documents that are stored on either internal or external Snowflake Stages.

Note: You cannot use both a text string and a file object in the same call. Choose one.

responseFormat

And the response format can be:

• An array of strings containing the information to be extracted
• An array of arrays containing two strings (label and the information to be extracted)
• A key-value map where the key is a name and the value is a question string
• A JSON schema object (for complex extractions). You define properties with descriptions and types (string, array or table object)

Format 1—Simple key-value object

responseFormat => {'name': 'What is the employee last name?', 'city': 'What city does the employee live in?'}

Format 2—Array of labeled pairs

responseFormat => [['name', 'What is the first name?'], ['city', 'Where does the employee live?']]

Format 3—Array of questions (no labels, labels auto-generated)

responseFormat => ['What is the invoice total?', 'What is the name of the vendor?']

Format 4—JSON schema (supports entity strings, arrays and tables)

responseFormat => { 

  'schema': { 

    'type': 'object', 

    'properties': { 

      'vendor_name': { 

        'description': 'What is the vendor name on this invoice?', 

        'type': 'string' 

      }, 

      'line_items': { 

        'description': 'What are the product descriptions on this invoice?', 

        'type': 'array' 

      }, 

      'charges_table': { 

        'description': 'The detailed list of charges', 

        'type': 'object', 

        'column_ordering': ['item', 'quantity', 'unit_price', 'total'], 

        'properties': { 

          'item': { 'type': 'array' }, 

          'quantity': { 'type': 'array' }, 

          'unit_price': { 'type': 'array' }, 

          'total': { 'type': 'array' } 

        } 

      } 

    } 

  } 

}

Note: You cannot mix JSON schema format with the other formats in the same call. If responseFormat contains the schema key, you must define all questions within the JSON schema. Also, string is currently the only supported scalar type in JSON schemas (numeric and boolean types aren’t supported yet).

Snowflake AI_EXTRACT vs Snowflake Document AI

So, what is the difference between Snowflake Document AI and Snowflake AI_EXTRACT?

Snowflake Document AI was Snowflake’s earlier approach to document extraction. Document AI and the <model_build_name>!PREDICT method are now deprecated. Snowflake recommends using the AI_EXTRACT function instead. If you are still using Snowflake Document AI pipelines, now is the time to migrate immediately.

Table 1: Difference between Snowflake AI_EXTRACT vs Snowflake Document AI

TL; DR: Snowflake Document AI used to be a more hands-on process with fixed models and limitations. Now, Snowflake AI_EXTRACT is a more modern approach that relies on code and uses large language models. It makes things easier by ditching the model building and fine-tuning steps – you can just ask questions as you need them. Plus, Snowflake AI_EXTRACT can handle a lot more languages and bigger outputs than the old solution. Switch to AI_EXTRACT if you are using Snowflake Document AI. Do it before February 28, 2026, because that is when the old service ends.

Snowflake AI_EXTRACT vs Snowflake AI_PARSE_DOCUMENT

Snowflake offers another new function, Snowflake AI_PARSE_DOCUMENT, for document processing. How does it differ from Snowflake AI_EXTRACT?

These two functions are easy to get mixed up with because they both deal with documents, but they are used for different things.

AI_PARSE_DOCUMENT is an OCR/layout extraction tool. It takes a document and turns it into text or Markdown, keeping its original layout and structure intact (headings, paragraphs and tables). What it does not do is analyze the document and answer specific questions about it. Its main goal is to make the document readable.

Snowflake AI_EXTRACT pulls out specific info from a document. You ask it a question and it gives you a simple, structured JSON answer. You will not get the full raw text of the document, just the details you are looking for.

Note: If you are porting an existing Snowflake Document AI pipeline, the rule of thumb is: use AI_EXTRACT when you need structured values and AI_PARSE_DOCUMENT when you just need OCR or text with layout. AI_PARSE_DOCUMENT is like exporting everything to JSON, whereas AI_EXTRACT picks out the specific parts you asked for.

Table 2: Difference between Snowflake AI_EXTRACT vs Snowflake AI_PARSE_DOCUMENT

TL; DR: AI_PARSE_DOCUMENT is useful for extracting the full text of a document into a readable format, perfect for RAG pipelines, search indexing, or training a summary model. Snowflake AI_EXTRACT is a better choice when you only need to extract specific fields and have them formatted into table-ready, structured JSON.

You can also try combining these two functions. First, use AI_PARSE_DOCUMENT to clean up the text and convert it to Markdown. Then, pass that text to Snowflake AI_EXTRACT with a text => … call. This two-step approach can sometimes improve accuracy on messy scanned documents.

What can you use Snowflake AI-EXTRACT for?

Snowflake AI_EXTRACT can be used anywhere you have documents or images with data you want to put in tables or fields. You can use Snowflake AI_EXTRACT in the following scenarios:

• Invoice and receipt processing — Pull vendor name, invoice number, line items, totals, due dates and tax amounts from PDF invoices; load directly into a staging table for accounts payable processing
• Contract and legal document analysis — Extract parties, effective dates, termination clauses, renewal terms and more from legal agreements; scale across thousands of contracts without manual review
• Table extraction for financial data ingestion — Pull structured tables from earnings reports, financial statements and data sheets directly into Snowflake tables
• Form processing — Extract filled-in values from application forms, tax forms, insurance claims and survey responses; Snowflake AI_EXTRACT handles checkboxes and handwritten fields through its vision model and is not limited to typed text
• Medical records/clinical notes — Extract patient info, diagnoses, medications and test results from clinical notes and discharge summaries; since Snowflake AI_EXTRACT runs entirely inside Snowflake with role-based access control (RBAC) enforcement, it keeps protected health information (PHI) within your governance perimeter
• Document search and RAG pipelines — Extract metadata fields (title, author, date, category, key topics) and store them alongside the document reference for downstream search and retrieval; even pair with Cortex Search for semantic lookup
• Financial reporting — Extract key performance indicators (KPIs), revenue figures, segment data and footnotes from quarterly reports across hundreds of filings at once using batch processing from a stage
• Email and ticket grouping — Pass the body of support tickets, complaint emails, or onboarding forms as text inputs to extract intent, product names, order numbers and customer details without storing them as files first
• Regulatory and compliance document review — Extract license numbers, compliance attestations, expiry dates and signatory names from regulatory filings at scale
• General data enrichment — Any time you want to add document-derived metadata to your database (like customer letters, support tickets, or user-written comments)

So, whenever you have unstructured documents with data you want to query or analyze in Snowflake, AI_EXTRACT is a big help. You define the fields once, and the LLM does the rest. This saves engineering time and reduces errors, making it better than manual processes.

How to calculate Snowflake AI_EXTRACT costs?

Snowflake AI_EXTRACT incurs two main kinds of costs: warehouse compute credits and Cortex (token-based) credits.

1) Warehouse compute costs

Every query you run still runs on a Snowflake virtual warehouse, so you pay for the warehouse time (just like any query). Because AI_EXTRACT is I/O-bound (the heavy LLM work is done by Snowflake’s managed service), you do not need a large warehouse. In fact, Snowflake explicitly recommends using a MEDIUM warehouse or smaller for AI_EXTRACT. Remember that bigger clusters do not speed up the LLM inference or reduce latency for AI functions. It just burns more credits on warehouse compute without any benefit. Your warehouse cost is therefore usually minimal, but if you queue many documents in parallel on a large cluster, it can add up fast.

2) Cortex AI (Token) billing

This is where the actual cost comes from. Snowflake AI_EXTRACT bills on both input tokens and output tokens using the Snowflake Arctic-Extract model, which runs at ~5 credits per million tokens (standard, non-fine-tuned).

What counts as input tokens

Input token billing has three components:

• Document content. Either page-based token counts for file inputs or actual text length for string inputs (more on this below)
• Your responseFormat payload. The questions, field names, and schema you pass in the argument all count as input tokens
• Snowflake’s managed system prompts. This is added automatically to every call, and you cannot see or control it

How do document pages convert to tokens?

For file-based inputs, Snowflake does not tokenize the actual text on the page. It uses a flat rate based on document format:

Table 3: Snowflake Cortex AI token cost breakdown

Note: The cost of a mostly blank PDF page is the same as a dense one with a lot of text, due to the flat 970 token rate.

Output token limits

Output tokens are billed based on the length of the JSON response AI_EXTRACT returns. The caps depend on what you are extracting:

• Entity and list extractions – 512 output tokens per question (maximum 100 questions per call)
• Table extractions – 4,096 output tokens per question (maximum 10 questions per call)

Table extraction can be expensive. If you are pulling full structured tables from documents, those output tokens can easily dominate the total cost of a call.

The token billing rate itself (5 credits per million tokens for standard Snowflake Arctic-Extract) is consistent across cloud providers. What changes by region is the dollar value of a Snowflake credit.

Check the current rates in Snowflake’s Service Consumption Table.

Sample cost estimate

Let us say we process 1k invoices. Each invoice has 3 pages, and we ask 5 entity questions per invoice.

• Input tokens per invoice ⇒ (3 pages × 970) + responseFormat tokens + system prompt ~3100 input tokens
• Output tokens per invoice ⇒ 5 questions × ~100 tokens average ~ 500 output tokens
• Total per invoice ⇒ ~3600 tokens
• Total for 1k invoices ⇒ ~3.6 million tokens

At 5 credits per million tokens, that is 18 credits. On AWS US East, at the Standard edition rate ($2/credit), it costs $36 for Cortex billing.

Add warehouse compute on top. An XS warehouse runs at 1 credit per hour. A 20-minute batch adds another ~0.33 credits, which is negligible.

These are rough estimates. Always run a sample before processing a large batch.

To monitor spend, use Snowflake’s account usage views.

Snowflake provides the CORTEX_FUNCTIONS_USAGE_HISTORY account usage view to track Cortex token consumption.

You can also use METERING_DAILY_HISTORY to track overall credit consumption at the account level and set budget alerts.

How to configure and use Snowflake AI_EXTRACT to extract data from documents?

Now, we have gotten the basics covered, so let us dive into the practical side of things. Next up, we’ll walk through setting up Snowflake AI_EXTRACT step by step, starting from an absolute scratch.

Prerequisites and setups:

Here is what you should have before you start:

• A Snowflake account on an edition/region that supports Snowflake Cortex AI functions and AI_EXTRACT.
• The ACCOUNTADMIN role, or a role with the ability to grant CORTEX_USER privileges
• A warehouse, database and schema to hold Snowflake Stages, tables and tasks. Use a dedicated warehouse for Cortex workloads to track credits separately.
• Cross-region inference enabled if your region doesn’t natively support AI_EXTRACT (covered in Step 5 below)
• Documents in a supported format (PDF, DOC, DOCX, PNG, JPEG, JPG, TIFF, TIF, HTML, HTM, TXT, TEXT, PPT, EML, BMP, GIF, WEBP, or MD). Files must be < 100 MB and < 125 pages. Snowflake AI_EXTRACT limits: up to 100 entity extraction questions or 10 table extraction questions per call; tables count as 10 entities each.

Step 1—Sign in to Snowsight or SnowSQL

Log in to your Snowflake account via the Snowsight web interface or the Snowflake CLI. All SQL in this guide runs in Snowsight or via SnowSQL.

Step 2—Create a dedicated warehouse, database and schema

Now, with your session open, the first real task is setting up the infrastructure. Keep Snowflake AI_EXTRACT workloads on a dedicated, right-sized warehouse- that way, it is easier to track costs later. (The MEDIUM size is recommended; X-Small or Small also often suffice). The warehouse will charge credits while running your AI queries.

CREATE WAREHOUSE IF NOT EXISTS doc_extract_wh 

    WITH WAREHOUSE_SIZE = 'MEDIUM' 

    AUTO_SUSPEND = 60 

    AUTO_RESUME = TRUE;

Figure 2: Creating the doc_extract_wh Snowflake virtual warehouse

Now create the database and schema that will hold your stages, tables and tasks.

CREATE DATABASE IF NOT EXISTS doc_processing_db; 

CREATE SCHEMA IF NOT EXISTS doc_processing_db.extraction;

Figure 3: Creating the doc_processing_db database and extraction schema—Snowflake AI_EXTRACT

Step 3—Create a custom role and assign least privileges

Do not run extraction workloads under ACCOUNTADMIN in production. The principle of least privilege applies here just as much as anywhere else—scope the permissions to exactly what the pipeline needs.

USE ROLE ACCOUNTADMIN; 

CREATE ROLE IF NOT EXISTS doc_extractor_role;

Figure 4: Creating the doc_extractor_role custom role—Snowflake AI_EXTRACT

Now, grant the role access to the warehouse, database and schema objects it’ll need.

GRANT USAGE ON WAREHOUSE doc_extract_wh TO ROLE doc_extractor_role; 

GRANT USAGE ON DATABASE doc_processing_db TO ROLE doc_extractor_role; 

GRANT USAGE ON SCHEMA doc_processing_db.extraction TO ROLE doc_extractor_role; 

GRANT CREATE TABLE ON SCHEMA doc_processing_db.extraction TO ROLE doc_extractor_role; 

GRANT CREATE STAGE ON SCHEMA doc_processing_db.extraction TO ROLE doc_extractor_role; 

GRANT CREATE STREAM ON SCHEMA doc_processing_db.extraction TO ROLE doc_extractor_role; 

GRANT CREATE TASK ON SCHEMA doc_processing_db.extraction TO ROLE doc_extractor_role;

Figure 5: Granting warehouse, database and schema privileges to doc_extractor_role—Snowflake AI_EXTRACT

Step 4—Grant Cortex privileges (control access)

The CORTEX_USER database role in Snowflake includes the necessary privileges to call Snowflake Cortex AI functions. It is automatically granted to the PUBLIC role, which means every user and role gets it by default.

For a production environment, it is a good idea to have more control over who can call Cortex functions. To do this, grant the CORTEX_USER role directly to a custom role you have created.

GRANT DATABASE ROLE SNOWFLAKE.CORTEX_USER TO ROLE doc_extractor_role; 

GRANT ROLE doc_extractor_role TO USER <your_username>;

Figure 6: Granting SNOWFLAKE.CORTEX_USER to doc_extractor_role and assigning it to a user—Snowflake AI_EXTRACT

You must run these as ACCOUNTADMIN or a similarly powerful role. Now, any user with doc_extractor_role can call Snowflake Cortex AI functions. Also make sure doc_extractor_role (or the granting role) has USAGE on the target warehouse, database and schema.

Step 5—(Optional) Enable cross-region inference

Snowflake AI_EXTRACT is natively available in select Snowflake regions. If your account is in a region that does not support it natively, you need to enable cross-region inference via the CORTEX_ENABLED_CROSS_REGION account parameter.

USE ROLE ACCOUNTADMIN;

To allow inference in any supported region:

ALTER ACCOUNT SET CORTEX_ENABLED_CROSS_REGION = 'ANY_REGION';

Or restrict processing to a specific cloud or region group:

ALTER ACCOUNT SET CORTEX_ENABLED_CROSS_REGION = 'AWS_US';

Note: Cross-region inference is not supported in U.S. SnowGov regions. User inputs, service-generated prompts and outputs are not stored or cached during cross-region inference. If both the source and destination regions are on AWS, the data stays within the AWS global network and is automatically encrypted at the physical layer.

Step 6—Switch role, set context, create prompt/metadata tables

Switch to your new role and set the working context. From here on, all objects are created under doc_extractor_role.

USE ROLE doc_extractor_role; 

USE WAREHOUSE doc_extract_wh; 

USE DATABASE doc_processing_db; 

USE SCHEMA extraction;

Figure 7: Setting the active role, warehouse, database and schema—Snowflake AI_EXTRACT

Now create a table to track documents queued for processing.

CREATE OR REPLACE TABLE document_queue ( 

    document_id     VARCHAR(100), 

    file_path       VARCHAR(500), 

    document_type   VARCHAR(50), 

    uploaded_at     TIMESTAMP_NTZ DEFAULT CURRENT_TIMESTAMP(), 

    processed_at    TIMESTAMP_NTZ, 

    status          VARCHAR(20) DEFAULT 'pending'  -- (pending, processed, error) 

);

Figure 8: Creating the document_queue tracking table—Snowflake AI_EXTRACT

And a table to store the extraction output.

CREATE OR REPLACE TABLE extraction_results ( 

    document_id     VARCHAR(100), 

    file_path       VARCHAR(500), 

    extracted_at    TIMESTAMP_NTZ DEFAULT CURRENT_TIMESTAMP(), 

    raw_response    VARIANT, 

    error_message   VARCHAR(500) 

);

Figure 9: Creating the extraction_results output table—Snowflake AI_EXTRACT

Step 7—Create Snowflake internal stage for document storage

Snowflake Cortex AI functions that process media files require the files to be stored on an internal or external stage. The Snowflake internal stage must use server-side encryption (client-side encrypted stages are not supported by Snowflake AI_EXTRACT). You also need a directory table to query stage contents or run batch processing.

Let’s create Snowflake internal stage with server-side encryption and directory table enabled.

CREATE OR REPLACE STAGE doc_processing_db.extraction.documents_stage 

  DIRECTORY = (ENABLE = TRUE) 

  ENCRYPTION = (TYPE = 'SNOWFLAKE_SSE') 

  COMMENT = 'Snowflake Internal stage for documents for Snowflake AI_EXTRACT';

Figure 10: Creating “documents_stage” Snowflake internal stage with SSE encryption and directory table enabled—Snowflake AI_EXTRACT

The DIRECTORY = (ENABLE = TRUE) setting is what lets you run FROM DIRECTORY(@documents_stage) in batch queries later. Do not skip this.

Note: SNOWFLAKE_SSE is server-side encryption managed by Snowflake. Client-side encrypted stages are not supported by Snowflake AI_EXTRACT.

Step 8—Uploading files to the Snowflake Internal Stage

Now, with the Snowflake Stage ready, you can load documents. There are a few ways to do it.

Via Snowsight:

Navigate to Catalog > Database Explorer > doc_processing_db > extraction > Stages > documents_stage >Stage Files, then drag and drop your files directly into the UI.

Figure 11: Uploading files via the Snowsight interface—Snowflake AI_EXTRACT

Via SnowSQL CLI:

snowsql -a <account> -u <user> 

PUT file:///local/path/to/<your_file>.pdf @doc_processing_db.extraction.documents_stage/invoices/ OVERWRITE = TRUE;

Via SQL (Python connector or Snowpark):

# via snowflake-connector-python 

import snowflake.connector 

conn = snowflake.connector.connect( 

    user='<user>', 

    password='<password>', 

    account='<account>' 

) 

cs = conn.cursor() 

cs.execute("PUT file:///local/path/*.pdf @doc_processing_db.extraction.documents_stage/invoices/ OVERWRITE=TRUE")

After uploading, refresh the directory table, so Snowflake indexes the new files.

ALTER STAGE doc_processing_db.extraction.documents_stage REFRESH;

Finally, verify the upload:

SELECT RELATIVE_PATH, SIZE, LAST_MODIFIED 

FROM DIRECTORY(@doc_processing_db.extraction.documents_stage) 

WHERE RELATIVE_PATH LIKE 'invoices/%';

Figure 12: Verifying uploaded files via the DIRECTORY() table function—Snowflake AI_EXTRACT

Or

LIST @doc_processing_db.extraction.documents_stage/invoices/;

Figure 13: Verifying uploaded files via LIST—Snowflake AI_EXTRACT

Note: for external stages you can configure automatic directory refresh via cloud notifications; for Snowflake internal stages ALTER STAGE … REFRESH is the manual refresh command.

Here is a screenshot of what our invoice looks like:

Figure 14: Sample invoice for testing extraction logic—Snowflake AI_EXTRACT

Step 9—Define document extraction schema and sample responseFormat

Here is where the real design work happens. A well-crafted responseFormat is the single biggest factor in extraction quality.

Let us look at some realistic examples.

Simple key-value document extraction

The simplest format is a flat object that maps field names to natural language questions.

SELECT AI_EXTRACT( 

  file => TO_FILE('@doc_processing_db.extraction.documents_stage', 'invoices/snowflake_ai_extract_invoice_001.pdf'), 

  responseFormat => { 

    'vendor_name': 'What is the vendor or supplier name on this invoice?', 

    'invoice_number': 'What is the invoice number or ID?', 

    'invoice_date': 'What is the invoice date?', 

    'due_date': 'What is the payment due date?', 

    'total_amount': 'What is the total amount due?' 

  } 

);

Figure 15: Running a key-value extraction with Snowflake AI_EXTRACT

Array document extraction (list of values)

When you need a list of values rather than a single answer, use the JSON schema format with “type”: “array.”

SELECT AI_EXTRACT( 

  file => TO_FILE('@doc_processing_db.extraction.documents_stage', 'invoices/snowflake_ai_extract_invoice_001.pdf'), 

  responseFormat => { 

    'schema': { 

      'type': 'object', 

      'properties': { 

        'line_item_descriptions': { 

          'description': 'What are all the product or service descriptions listed on this invoice?', 

          'type': 'array' 

        } 

      } 

    } 

  } 

);

Figure 16: Extracting a list of values using an array schema—Snowflake AI_EXTRACT

Table extraction with column ordering

For extracting tabular data (like a full line-items table) use an “type”: “object” with column_ordering and array-typed columns. You can mix table extraction with scalar fields in the same call.

SELECT AI_EXTRACT( 

  file => TO_FILE('@doc_processing_db.extraction.documents_stage', 'invoices/snowflake_ai_extract_invoice_001.pdf'), 

  responseFormat => { 

    'schema': { 

      'type': 'object', 

      'properties': { 

        'line_items': { 

          'description': 'The itemized line items table on this invoice', 

          'type': 'object', 

          'column_ordering': ['description', 'quantity', 'unit_price', 'line_total'], 

          'properties': { 

            'description': { 'type': 'array' }, 

            'quantity': { 'type': 'array' }, 

            'unit_price': { 'type': 'array' }, 

            'line_total': { 'type': 'array' } 

          } 

        }, 

        'vendor': { 

          'description': 'What is the vendor name?', 

          'type': 'string' 

        }, 

        'invoice_total': { 

          'description': 'What is the total amount due?', 

          'type': 'string' 

        } 

      } 

    } 

  } 

);

The description field is your main lever for improving extraction accuracy. Vague descriptions produce vague results. Be very very specific, especially on documents with multiple subtotals, nested tables, or repeated fields. That extra context helps the model locate the right value.

Figure 17: Extracting table data with column ordering defined—Snowflake AI_EXTRACT

Note: In a single Snowflake AI_EXTRACT call, you can ask a maximum of 100 entity extraction questions, or up to 10 table extraction questions (each table question counts as 10 entity questions toward the limit).

Step 10—(Optional) Create Snowflake AI_EXTRACT wrapper function

If you are running the same extraction schema repeatedly, wrapping it in a user-defined function (UDF) reduces repetition and makes schema updates a one-place change.

GRANT CREATE FUNCTION ON SCHEMA doc_processing_db.extraction TO ROLE doc_extractor_role; 

CREATE OR REPLACE FUNCTION doc_processing_db.extraction.extract_invoice_fields( 

  stage_path VARCHAR, 

  file_name  VARCHAR 

) 

RETURNS VARIANT 

LANGUAGE SQL 

AS 

$$ 

  CAST( 

    AI_EXTRACT( 

      file => TO_FILE(stage_path, file_name), 

      responseFormat => { 

        'vendor_name':    'What is the name of the vendor or supplier?', 

        'invoice_number': 'What is the invoice number?', 

        'invoice_date':   'What is the invoice date?', 

        'due_date':       'What is the payment duenames,?', 

        'total_amount':   'What is the total amount due including taxes?', 

        'currency':       'What currency is used?', 

        'payment_terms':  'What are the payment terms?' 

      } 

    ) AS VARIANT 

  ) 

$$;

Figure 18: Creating extract_invoice_fields as a reusable SQL UDF wrapping Snowflake AI_EXTRACT

Here’s how to call it:

SELECT doc_processing_db.extraction.extract_invoice_fields( 

  '@doc_processing_db.extraction.documents_stage', 

  'invoices/snowflake_ai_extract_invoice_001.pdf' 

) AS extraction_result;

Figure 19: Testing the wrapper function with the sample invoice—Snowflake AI_EXTRACT

Step 11—Test Snowflake AI_EXTRACT on sample documents

Before building automation, test your document extraction logic manually. Run through each of these patterns to confirm everything works in your environment.

Test 1–Extract from a text string

Snowflake AI_EXTRACT works on plain text too, not just files. It’s good for quick prototyping.

SELECT AI_EXTRACT( 

  text => 'Invoice #INV-2026-NVD-800 from Dragon Corp dated Feb 20, 2026. Total due: $7,432,250. Payment due by February 21, 2026.', 

  responseFormat => { 

    'vendor': 'What is the vendor name?', 

    'invoice_number': 'What is the invoice number?', 

    'total_due': 'What is the total amount due?', 

    'due_date': 'What is the due date?' 

  } 

) AS extraction_result;

Figure 20: Extracting fields from an inline text string using Snowflake AI_EXTRACT

Test 2–Extract from a single PDF

SELECT AI_EXTRACT( 

  file => TO_FILE('@doc_processing_db.extraction.documents_stage', 'invoices/snowflake_ai_extract_invoice_001.pdf'), 

  responseFormat => {'vendor': 'Vendor name', 'total': 'Total amount due'} 

) AS result;

Figure 21: Extracting vendor name and total from a single PDF using Snowflake AI_EXTRACT

Test 3–Extract table data with JSON schema

SELECT 

  relative_path, 

  AI_EXTRACT( 

    file => TO_FILE('@doc_processing_db.extraction.documents_stage', relative_path), 

    responseFormat => { 

      'schema': { 

        'type': 'object', 

        'properties': { 

          'charges': { 

            'description': 'Itemized charges table', 

            'type': 'object', 

            'column_ordering': ['item', 'amount'], 

            'properties': { 

              'item': { 'type': 'array' }, 

              'amount': { 'type': 'array' } 

            } 

          } 

        } 

      } 

    } 

  ) AS extraction_result 

FROM DIRECTORY(@doc_processing_db.extraction.documents_stage) 

WHERE relative_path LIKE 'invoices/%';

Figure 22: Batch table extraction across all invoice files in the Snowflake Stage using Snowflake AI_EXTRACT

Test 4–Parse the JSON output into columns

The raw output from Snowflake AI_EXTRACT is a VARIANT with response and error keys. Use Snowflake’s semi-structured data operators to flatten it into columns.

WITH extracted AS ( 

  SELECT 

    relative_path, 

    AI_EXTRACT( 

      file => TO_FILE('@doc_processing_db.extraction.documents_stage', relative_path), 

      responseFormat => { 

        'vendor': 'Vendor name', 

        'invoice_number': 'Invoice number', 

        'total': 'Total amount due' 

      } 

    ) AS result 

  FROM DIRECTORY(@doc_processing_db.extraction.documents_stage) 

  WHERE relative_path LIKE 'invoices/%' 

) 

SELECT 

  relative_path, 

  result:response:vendor::VARCHAR AS vendor, 

  result:response:invoice_number::VARCHAR AS invoice_number, 

  result:response:total::VARCHAR AS total, 

  result:error::VARCHAR AS error 

FROM extracted;

Figure 23: Parsing JSON extraction output into relational columns—Snowflake AI_EXTRACT

Step 12—Set up Snowflake Streams and Snowflake Tasks for automated processing

Once your document extraction logic is in place, automate it. A Stream on the stage directory table detects new files; a Snowflake Task processes them on a schedule.

First, create a Snowflake Stream on the directory table to detect new files.

CREATE OR REPLACE STREAM documents_stream 

    ON STAGE doc_processing_db.extraction.documents_stage;

Figure 24: Creating a stream to detect new file uploads—Snowflake AI_EXTRACT

Then, create a Snowflake Task that runs every 5 minutes and processes new documents

CREATE OR REPLACE TASK process_new_documents 

  WAREHOUSE = doc_extract_wh 

  SCHEDULE  = '5 MINUTE'  

  WHEN SYSTEM$STREAM_HAS_DATA('documents_stream') 

AS 

INSERT INTO extraction_results (document_id, file_path, extracted_at, raw_response, error_message) 

SELECT 

  MD5(RELATIVE_PATH) AS document_id, 

  RELATIVE_PATH AS file_path, 

  CURRENT_TIMESTAMP() AS extracted_at, 

  AI_EXTRACT( 

    file => TO_FILE('@doc_processing_db.extraction.documents_stage', RELATIVE_PATH), 

    responseFormat => { 

      'vendor_name':    'What is the vendor name?', 

      'invoice_number': 'What is the invoice number?', 

      'invoice_date':   'What is the invoice date?', 

      'total_amount':   'What is the total amount due?' 

    } 

  ) AS raw_response, 

  AI_EXTRACT( 

    file => TO_FILE('@doc_processing_db.extraction.documents_stage', RELATIVE_PATH), 

    responseFormat => {'error_check': 'Return "NO_ERROR" if extraction successful, otherwise return the error message.'} 

  ):response:error::VARCHAR AS error_message 

FROM documents_stream 

WHERE METADATA$ACTION = 'INSERT';

Figure 25: Creating the process_new_documents Snowflake Task to automate Snowflake AI_EXTRACT on new stage files

Finally, start the Snowflake Task. Tasks are created in a suspended state by default.

ALTER TASK process_new_documents RESUME; 

Figure 26: Resuming the task to enable automation—Snowflake AI_EXTRACT

Step 13—Upload new documents for automated processing

Once Snowflake Stream and Snowflake Task are running, any new file you upload will be picked up automatically on the next task run (within 5 minutes by default).

Let’s start by uploading a new invoice via Snowsight (the process is the same as above) or you can do so via SnowSQL

PUT file:///path/to/new_invoice.pdf @doc_processing_db.extraction.documents_stage/invoices/ OVERWRITE=TRUE;

Then refresh the directory table so the Stream can detect the new file.

ALTER STAGE doc_processing_db.extraction.documents_stage REFRESH;

Step 14—Create error handling and validation

Snowflake AI_EXTRACT does not raise SQL exceptions for extraction failures. It returns an error key in the JSON response instead. If you do not build validation into your pipeline, failed extractions will silently pile up in your results table.

To do so, first create a view to surface failed extractions.

GRANT CREATE VIEW ON SCHEMA doc_processing_db.extraction TO ROLE doc_extractor_role; 

CREATE OR REPLACE VIEW extraction_errors AS 

SELECT 

  file_path, 

  extracted_at, 

  raw_response:error::VARCHAR AS error_message 

FROM extraction_results 

WHERE raw_response:error IS NOT NULL 

  OR error_message IS NOT NULL;

Figure 27: Creating a view to surface failed extractions—Snowflake AI_EXTRACT

Then set up a Snowflake Task that logs errors hourly.

CREATE TABLE IF NOT EXISTS extraction_error_log ( 

  logged_at TIMESTAMP_NTZ, 

  error_count INTEGER, 

  sample_files ARRAY 

); 

 

CREATE OR REPLACE TASK alert_on_extraction_errors 

  WAREHOUSE = doc_extract_wh 

  SCHEDULE = '60 MINUTE' 

AS 

INSERT INTO extraction_error_log (logged_at, error_count, sample_files) 

SELECT 

  CURRENT_TIMESTAMP(), 

  COUNT(*), 

  ARRAY_AGG(file_path) WITHIN GROUP (ORDER BY extracted_at DESC) 

FROM extraction_errors 

WHERE extracted_at >= DATEADD(hour, -1, CURRENT_TIMESTAMP());

Figure 28: Setting up an hourly Snowflake Task for error logging—Snowflake AI_EXTRACT

For even more powerful error handling inside stored procedures, wrap Snowflake AI_EXTRACT in a TRY/CATCH block using Snowflake Scripting.

GRANT CREATE PROCEDURE ON SCHEMA doc_processing_db.extraction TO ROLE doc_extractor_role; 

CREATE OR REPLACE PROCEDURE doc_processing_db.extraction.safe_extract_document( 

  stage_path VARCHAR, 

  file_path  VARCHAR 

) 

RETURNS VARIANT 

LANGUAGE SQL 

AS 

$$ 

DECLARE 

  result VARIANT; 

  err_msg STRING; 

BEGIN 

  -- call AI_EXTRACT and capture result 

  result := AI_EXTRACT( 

    file => TO_FILE(stage_path, file_path), 

    responseFormat => { 

      'vendor': 'Vendor name', 

      'total':  'Total amount due' 

    } 

  ); 

  INSERT INTO doc_processing_db.extraction.extraction_results (document_id, file_path, extracted_at, raw_response, error_message) 

  VALUES (MD5(file_path), file_path, CURRENT_TIMESTAMP(), result, result:error::VARCHAR); 

  RETURN result; 

EXCEPTION 

  WHEN OTHER THEN 

    -- capture SQL error and store a minimal failure row for later investigation 

    err_msg := SQLERRM; 

    INSERT INTO doc_processing_db.extraction.extraction_error_log (logged_at, error_count, sample_files, sample_errors) 

    VALUES (CURRENT_TIMESTAMP(), 1, ARRAY_CONSTRUCT(file_path), ARRAY_CONSTRUCT(err_msg)); 

    -- return a JSON-like VARIANT with the error so callers get a predictable shape 

    RETURN PARSE_JSON('{"error": "' || REPLACE(err_msg, '"', '''') || '", "response": null}'); 

END; 

$$;

Figure 29: safe_extract_document stored procedure with TRY/CATCH error handling wrapping Snowflake AI_EXTRACT

Step 15—Monitoring and alerting

Finally, with extraction running in production, monitor both Cortex token consumption and warehouse credit spend. Do not rely on a fixed token-per-page magic number. Snowflake reports Cortex usage in account-usage views. Use the account usage views below to measure token and credit consumption.

Query Cortex AI SQL usage => hourly aggregates

SELECT 

  USAGE_TIME, 

  FUNCTION_NAME, 

  MODEL_NAME, 

  SUM(TOKEN_CREDITS) AS token_credits, 

  SUM(TOKENS)        AS tokens_used 

FROM SNOWFLAKE.ACCOUNT_USAGE.CORTEX_AISQL_USAGE_HISTORY 

WHERE USAGE_TIME >= DATEADD(day, -30, CURRENT_TIMESTAMP()) 

GROUP BY 1,2,3 

ORDER BY USAGE_TIME DESC, token_credits DESC;

Figure 30: Analyzing token usage history—Snowflake AI_EXTRACT

Use CORTEX_AISQL_USAGE_HISTORY for the most up-to-date AISQL usage view. CORTEX_FUNCTIONS_USAGE_HISTORY is deprecated for some use cases; prefer the AISQL view for SQL-based function calls.

Query daily metering (credits) for warehouses

SELECT 

  USAGE_DATE, 

  SERVICE_TYPE, 

  WAREHOUSE_NAME, 

  SUM(CREDITS_USED) AS credits_used 

FROM SNOWFLAKE.ACCOUNT_USAGE.METERING_DAILY_HISTORY 

WHERE USAGE_DATE >= DATEADD(day, -30, CURRENT_DATE()) 

GROUP BY 1,2,3 

ORDER BY USAGE_DATE DESC;

Figure 31: Monitoring warehouse credit consumption—Snowflake AI_EXTRACT

Last step: Set up a Snowflake Alert to notify your team when Cortex spending reaches a threshold within an hour. First, make sure you have a notification setup in place, as you’ll need it to send the alert. Also, verify that the email addresses of the recipients are valid. For more information, check out Notifications in Snowflake. Once that is done, execute the following code:

GRANT CREATE ALERT ON SCHEMA doc_processing_db.extraction TO ROLE doc_extractor_role; 

CREATE OR REPLACE ALERT doc_processing_db.extraction.high_cortex_spend_alert 

  WAREHOUSE = doc_extract_wh 

  SCHEDULE  = '1 HOUR' 

IF ( 

  EXISTS ( 

    SELECT 1 

    FROM SNOWFLAKE.ACCOUNT_USAGE.CORTEX_AISQL_USAGE_HISTORY 

    WHERE USAGE_TIME >= DATEADD(hour, -1, CURRENT_TIMESTAMP()) 

    GROUP BY FUNCTION_NAME 

    HAVING SUM(TOKEN_CREDITS) > 100   -- choose threshold 

  ) 

) 

THEN 

  CALL SYSTEM$SEND_EMAIL( 

    'cortex_alerts', 

    'pramitx47@gmail.com', 

    'High Cortex AI Spend Detected', 

    'Cortex AI spend exceeded threshold in the last hour. Check CORTEX_AISQL_USAGE_HISTORY' 

  );

Figure 32: Configuring alerts for high credit spend—Snowflake AI_EXTRACT

Pair CORTEX_FUNCTIONS_USAGE_HISTORY and METERING_DAILY_HISTORY to get a complete view of Snowflake AI_EXTRACT costs. Note that token charges and warehouse charges are billed separately.

And that’s it. You should now have a good understanding of how to configure, set up and use Snowflake AI_EXTRACT.

What are the limitations of Snowflake AI_EXTRACT ?

Snowflake AI_EXTRACT has certain limitations that you should be aware of when you’re planning.

1) Text and file parameters are mutually exclusive

You can’t pass text => and file => in the same call. If your workflow sometimes receives text and file references, route them through separate query paths.

2) Document size limits

Documents longer than 125 pages will fail. The document limit is 125 pages. Also, files cannot be larger than 100 MB. Another thing: client-side encrypted stages are not supported here.

3) Question limits

The maximum output length for entity extraction is 512 tokens per question. For table extraction, the model returns a maximum of 4,096 tokens. Per call, you can ask up to 100 entity extraction questions. Table extraction counts as 10 entity questions regardless of actual column count.

4) Output token truncation

Big tables can cause problems. If a table exceeds 4,096 tokens, the output will get cut off. The JSON will just stop without warning. To fix this, you can break the table into smaller groups by column or process the data a page at a time.

5) No confidence scores

Snowflake Document AI gave you confidence scores to help flag extractions that were not accurate for human review. Snowflake AI_EXTRACT does not do this. You will need to create your own way to check the results, such as looking for missing values, running test queries, or using AI_FILTER to make sure the extracted values are correct.

6) JSON schema type support

The model is quite particular about the JSON schema it accepts. For entity questions, it only supports string types. In the JSON schema, you cannot specify numeric or boolean types – all values will be returned as strings. Make sure to cast them correctly in your subsequent SQL queries.

7) Regional availability

Snowflake AI_EXTRACT is natively available in select regions. For other regions, cross-region inference is required and must be explicitly enabled by an ACCOUNTADMIN. Cross-region inference is not available in U.S. SnowGov regions.

8) Client-side encrypted stages not supported

Server-side encryption (SNOWFLAKE_SSE or cloud provider SSE) works fine. Client-side encrypted stages are not supported.

9) Custom network policies

Custom network policies are not currently supported for Snowflake Cortex AI functions.

10) No fine-tuning

Zero-shot extraction is good enough for most documents. But documents with very specific or unique formats might not work as well with this method. If that is the case, you won’t be able to train the model using your own examples. On a positive note, legacy Snowflake Document AI models that have been fine-tuned can still be used with the Snowflake AI_EXTRACT(model => …, file => …) legacy syntax after you migrate. One thing to note, though, is that you won’t be able to create any new custom models from scratch.

Even with limitations, Snowflake AI_EXTRACT can handle a lot of common documents without issues. The trick is to create prompts and schemas that avoid super long answers and huge tables all at once.

What are the best practices for using Snowflake AI_EXTRACT effectively?

To make the most of Snowflake AI_EXTRACT, here are some best practices to keep in mind:

1) Keep prompts specific and concise

Keep your questions short and sweet. Vague questions usually get vague answers. The clearer you ask, the more exact the info you get.

2) Use the description field to localize

In a document with multiple tables or sections, the description field in JSON schema is how you point the model to the right one. To avoid confusion, provide some context for the model by describing the table’s title, location, or other details. That way, it knows exactly what you are referring to.

3) Use AI_COUNT_TOKENS before large batch jobs

Run a sample before committing. This takes 30 seconds and can save you from an unexpected credit bill. Use AI_COUNT_TOKENS to estimate the cost of a sample before committing a full run.

4) Materialize results—do not reprocess

Every Snowflake AI_EXTRACT call costs tokens. If you have already extracted data from a document, store the result in a table and query the table. Do not call Snowflake AI_EXTRACT on the same document twice. Use the Snowflake Stream and Snowflake Task pattern to process each document exactly once.

5) Split large tables by column group

When a table has a lot of rows, it might get cut off due to the 4,096-token output limit. To avoid this, try extracting just 3-4 columns at a time, rather than the whole table, and then combine the results based on the row index.

6) Break large documents into chunks before extraction

If you are near the 125-page limit, splitting your documents into smaller chunks is often a clever idea. Break them down by chapter, section, or topic and process each one separately. This usually gives you more accurate results than processing a large document all at once.

7) Use MEDIUM or smaller warehouses

Larger warehouses do not speed up AI_EXTRACT performance. That is because Cortex functions run on Snowflake’s serverless infrastructure, not your warehouse’s compute nodes. Your warehouse is mainly used for query orchestration and data movement. A MEDIUM-sized warehouse is plenty big enough.

8) Use column_ordering in table extraction schemas

In your JSON schema, the column_ordering field determines the order of columns in the extracted output. Without this field, the column order can be all over the place when looking at tables across different documents. Always include column_ordering when loading table data into a Snowflake table with predefined column positions.

9) Process documents directly from stages

Do not load file content into VARCHAR columns and pass it as text when you have documents – use the file parameter instead. It is more efficient and handles binary formats like PDFs correctly.

10) Enable directory tables before batch queries

If you forget to enable DIRECTORY = (ENABLE = TRUE) on your Snowflake Stage at creation time, you can alter the stage later. Just be sure to manually refresh the stage before querying the directory table.

11) Use role-based access to scope Cortex function usage

Grant SNOWFLAKE.CORTEX_USER only to the roles that need it for running Snowflake Cortex AI functions. Out of the box, the CORTEX_USER role is available to the PUBLIC role, so all users in your account can use Snowflake Cortex AI functions. But in enterprise environments, chances are you will want to limit access.

12) Combine with AI_PARSE_DOCUMENT for complex layouts

Documents with unusual formatting, such as multiple columns, mixed text and images, or dense tables, can be tricky. But running Snowflake AI_PARSE_DOCUMENT in LAYOUT mode first can improve AI_EXTRACT accuracy because the model gets a cleaner, Markdown-structured version of the document to work from.

And that’s a wrap!

Conclusion

Snowflake AI_EXTRACT is a powerful function that makes analyzing documents in Snowflake much easier. You can extract key information, lists and tables from documents without needing external tools or manual parsing. It is a one-step query that replaces the older Snowflake Document AI workflow and scales with your existing data pipelines. To get started, follow the setup steps above. Be very cautious that there are some limitations, such as page count and token length, and always keep track of your costs. The benefit is that you can now access all the data hidden in your reports, contracts and more, all within Snowflake.

In this article, we have covered:

• What is Snowflake Cortex AI?
• What is the Snowflake AI_EXTRACT function?
• Syntax and argument breakdown of the Snowflake AI_EXTRACT function
• Snowflake AI_EXTRACT vs Snowflake Document AI
• Snowflake AI_EXTRACT vs Snowflake AI_PARSE_DOCUMENT
• What can you use Snowflake AI-EXTRACT for?
• How to calculate Snowflake AI_EXTRACT costs
• Step-by-step guide to configure and use Snowflake AI_EXTRACT to extract data from documents
• Limitations of Snowflake AI_EXTRACT
• Best practices for using Snowflake AI_EXTRACT effectively

… and so much more!

Want to learn more? Reach out for a chat

FAQs

SaaS has become the default way organizations consume software, but the challenges created by SaaS sprawl vary dramatically based on the size and complexity of the business.

Small businesses and mid-market organizations tend to see SaaS growth in manageable pockets. They want visibility, cost control and basic workflow automation. Enterprises operate in environments that are exponentially more complex, distributed and regulated. Their SaaS challenges aren’t simply bigger versions of SMB problems—they’re fundamentally different.

Let’s break down those differences and explain why enterprise SaaS management requires a very specific platform approach.

Visibility: from “what are we using?” to “what’s connected to what?”

SMB needs

• Identify which SaaS applications are in use
• Consolidate duplicate tools
• Reduce untracked, department-level purchases

Enterprise needs

Enterprises often run hundreds or thousands of SaaS applications across regions, identities, business units and cloud ecosystems. Their visibility needs include:

• Discovering sanctioned and unsanctioned SaaS
• Mapping cloud-to-cloud integrations
• Understanding how sensitive data moves between apps
• Tracking region-specific usage for data residency and compliance
• Identifying access across multiple identity providers and SSO systems

For enterprises, visibility isn’t a dashboard. It’s governance.

Spend and financial governance: from waste reduction to strategic cost control

SMB needs

• Stop paying for unused licenses
• Gain clarity into subscription renewals
• Avoid redundant tools performing the same job

Enterprise needs

Enterprises face financial governance challenges at a global scale, including:

• Multicurrency SaaS spend
• Cost allocation across departments, regions or projects
• Complex tiered licensing models
• High-value contract renewals that require strategic negotiation
• Thousands of provisioned but unused licenses
• Redundant tools embedded in critical business processes

For SMBs, SaaS waste is a cost-saving opportunity. For enterprises, it’s a financial risk.

Complexity of SaaS ecosystems: from subscriptions to multilayered architecture

SMB needs

• Basic license usage tracking
• Understanding seat utilization
• Standardized metrics and simple vendor models

Enterprise needs

SaaS in an enterprise environment is rarely plug-and-play. It often includes:

• Multiple licensing metrics
• Regional pricing differences
• Complex licensing bundles
• Integrations across CRMs, HRIS platforms, finance systems, security tools and cloud services
• High volumes of configuration and data interchange
• Deep dependencies between applications

Enterprises don’t just use SaaS — they operate SaaS ecosystems.

Lifecycle (JML) management: from convenience automation to risk mitigation

SMB needs

• Simple automated onboarding and offboarding
• Reduced IT admin effort
• Basic tooling updates

Enterprise needs

Enterprises need lifecycle orchestration tied to compliance, identity and risk, including:

• Provisioning dozens of apps on an employee’s first day
• Automated, policy-driven access based on roles and personas
• Federated identity across multiple directories
• Monitoring contractors, partners, subsidiaries and temporary workers
• Strict, auditable offboarding
• Integrations with HRIS, ITSM, security tools and cloud services

Where SMBs focus on efficiency, enterprises focus on control and compliance.

Security and compliance: from reducing risk to managing global exposure

SMB needs

• Ensure access is revoked correctly
• Reduce shadow IT
• Maintain a basic security posture

Enterprise needs

Enterprise SaaS security has broader implications, including:

• Regulatory compliance requirements like GDPR, HIPAA, PCI and SOX
• Sensitive data shared through cloud-to-cloud connections
• Orphaned accounts across region-specific systems
• Application sprawl that increases attack surfaces
• Visibility into where data is stored, moved, synced and shared
• Vendor-specific compliance validation and audit readiness

For SMBs, SaaS security is important. For enterprises, it’s existential.

Why Flexera stands apart: the only SaaS management platform built for enterprise scale

Most SaaS management tools were built for small or mid-market companies. Flexera takes a different approach.

Flexera is designed to handle:

• Global, distributed SaaS environments
• Complex licensing and financial structures
• Large-scale data flows and cloud integrations
• Enterprise-grade lifecycle orchestration
• Deep visibility across hybrid IT estates (SaaS, on-premises and cloud)
• Risk, compliance and governance requirements
• The performance, accuracy and discovery depth needed for thousands of apps

This is why enterprises trust Flexera: it’s not a scaled-up SMB tool. It’s an enterprise SaaS management platform by design.

Ready to talk about how Flexera can help close the SaaS gap?

Reach out for a chat

Innovation area: IT Visibility

Export your full private catalog via public API

Flexera One has made available a new full download capability for IT Visibility’s Private Catalog. The new public API endpoint now allows customers to export all catalog records in a single CSV file, making large‑scale data access faster and more flexible.

Previously, customers were able to download catalog records through the Private Catalog UI of up to the first 250 records per export. With this release, retrieve your entire catalog programmatically, eliminating manual, UI‑based exports and record limits.

Why it matters

For customers managing large or complex catalogs, access to complete data is critical. This enhancement enables better automation, more accurate reporting, and easier integration with downstream systems. By removing export limits and manual steps, teams can save time, reduce errors, and work with their full catalog data whenever and however they need it.

For more information see our community blog.

Clean up device data at scale with Bulk Ignore via API

Bulk Ignore of devices via API in Flexera One IT Visibility gives customers a supported, automated way to ignore devices they no longer want included in dashboards and reports.

Previously, devices that were no longer relevant created discrepancies and manual cleanup challenges. With the new Bulk Ignore API, you can now ignore devices in bulk, reduce mismatches between expected and reported device counts, and keep your environment clean and accurate in near real time.

Why it matters

Accurate device data is essential for trusted reporting and confident decision making. This release gives you greater control over how device data is managed, improves operational efficiency through automation, and helps ensure consistency across dashboards and reports so IT Visibility better reflects the reality of your environment.

Developer documentation is available here.

Innovation area: IT Asset Management

AI Contract Ingestion—early access

AI Driven Contract Ingestion is now in Early Access for select customers for Flexera One ITAM. Leverage AI to streamline and accelerate the ingestion of contract data, significantly reducing manual effort while improving consistency and time to value.

Why it matters

Manually inputting contract information is time-consuming and prone to human error. Worst still errors can lead to a cascading effect of inaccurate data and failed optimization of licenses costing you money in missed opportunities in software cost reduction or even opening you up to software audits.

As this is an Early Access feature we are continue to refine this feature from real-world usage, gathering customer feedback and tweaking and enhancing the feature before general release.

Today, all customers have access to our AI entitlement ingestion feature. For more information on AI entitlement ingestion read our blog and consult our documentation.

License consumption for stand-by Oracle databases

This enhancement resolves a persistent issue where SAM tools struggle to reliably detect and report Active-Passive (stand-by) Oracle database relationships due to the absence of any direct indicator. To solve this Flexera has introduced intelligent pairing and recognition of these nodes.

Why it matters

Flexera One ITAM customers now have greater accuracy and visibility into their Oracle environments by accurately accounting for these stand-by databases. In the Licensing for Options and Management Packs section we have automatically aligned these changes between primary and stand-by nodes to ensure compliance and full coverage for our customers.

Enhanced Oracle WebLogic recognition

Flexera One ITAM has introduced advanced detection capabilities for Oracle WebLogic. We’ve expanded detection rules to ensure WebLogic server configurations are identified even when files are located outside standard locations, we’ve improved differentiation between active/configured WebLogic domains, and enhanced evidence gathering from Oracle Fusion Middleware audit data.

Why it matters

These updates deliver more reliable and precise WebLogic Server detection across a wide range of environments, supporting greater accuracy in inventory and reporting.

Innovation area: FinOps

Cloud Cost Optimization (CCO) now brings Partners more ways to improve margins and more ability to scale.

Key highlights for Partners

Multi-level hierarchy management and control: Partners now get unprecedented levels of granular control to improve margins and improve customers’ billing and invoicing experience across multiple levels of partner hierarchies. This new capability comes with guard rails—users in the L1 or L2 will have elevated privileges to their respective child organizations but are prevented from deleting or causing any forms of data loss or concerns with integrity.

Why it matters

Partners often struggle to manage more levels than parent / child hierarchy levels, adding manual effort and introducing the risk of inaccuracy for complex FinOps billing, invoicing and governance needs. With this capability, partners now can manage parent / child / grandchild relationships in CCO across multiple lines of business (LOBs) with distinct customer bases and configurations within a single platform.

Next Generation Billing Management: We now enable partners with fine-grained control over how costs are allocated and transformed in CCO, helping them meet a key element of a mature FinOps practice, Billing.

Billing Plans enable partners and customers to begin creating groups of Billing Rules that will transform their vendor’s bill data to align with what they intend to present or invoice to their end customers. New Billing History capabilities enable partner customers the ability to lock, unlock or reprocess the bill data as they make configuration changes. With customers being empowered, partners spend less time managing their customers’ billing and have more time for higher-value activity.

Why it matters

MSPs and Distributors frequently struggle to maximize profits while controlling what is presented in their billing data to their stakeholders and customers in reports or invoices. With this CCO capability, partners can better perform cloud pricing strategies such as:

• Arbitrage to manage the presentation of commitment / program discounts.
• Cost allocation to better align the costs of consumed assets to their consumers, with more granularity than the vendor may provide at times.
• Custom pricing or margin enhancement to better align the costs of the cloud usage to the prices they’re passing to their customers.

Innovation area: Security Vulnerability Management

AdminStudio 2026 is here: modern app packaging, broader integrations, stronger security

AdminStudio 2026 is now available and ready for download from the Product and License Center. This release delivers a powerful set of new features, platform enhancements, and fixes designed to streamline application packaging, testing, and deployment across modern enterprise environments.

Key highlights include:

• Omnissa App Volumes integration, enabling a unified workflow for onboarding and managing App Volumes applications
• Tanium Distribution System support for direct publishing and deployment visibility
• MP4 video recording in Repackager to capture and document installation steps with greater accuracy.

AdminStudio 2026 also introduces enhanced password encryption, support for Windows 11 25H2 (64bit), compatibility with Microsoft ConfigMgr 2503, and includes InstallShield 2025 R2.

Why it matters

AdminStudio 2026 helps teams reduce complexity while keeping pace with evolving technologies. By expanding integration options, strengthening security, and supporting the latest operating systems and deployment platforms, this release enables faster application readiness, better visibility, and more confident decision making—so you can modernize your application estate without slowing down.

For full details, see the release notes and explore more insights in the community blog.

Software Vulnerability Research (SVR) and Manager (SVM) updates

SVR upgrades to Python 3.12

The SVR platform has been successfully upgraded from Python 3.10 to Python 3.12. This upgrade ensures continued security compliance, improved performance, and longterm support as Python 3.10 approaches end of life.

SVM upgrades to MariaDB 10.11.15

In addition, the SVM production environment has been upgraded to MariaDB version 10.11.15. This update also improves database stability, security, and performance while maintaining full vendor support and longterm maintainability.

Why it matters

Keeping the platform on supported, up-to-date technologies is critical to delivering a secure, reliable, and high-performing experience. These upgrades reduce longterm risk, ensure compliance with evolving security standards, and provide a strong foundation for future enhancements, all without disrupting your daytoday operations.

Want to learn more about these and future updates? Get in touch!

Let’s chat

The use of large language models (LLMs) has moved from research labs into production systems faster than anyone predicted. A whole lot of companies and businesses are racing to figure out how to use AI on their own data without building custom infrastructure or hiring an army of engineers. The problem? Most AI tools live outside your data platform. You’re shuffling information back and forth, worrying about security, dealing with inconsistent answers and watching costs spiral. That gap created demand for AI that reasons over enterprise data while enforcing governance. For users of the Snowflake platform, there’s good news. In mid-2025, during the Snowflake Summit event, Snowflake announced its state-of-the-art Snowflake Intelligence feature, which reached general availability on November 4, 2025.

Snowflake Intelligence is the answer to bringing AI directly to your data instead of the other way around. It’s an agentic platform that lets business users ask questions in plain English and get answers, charts and actions, all within the governed environment of the Snowflake platform and without having to write complex query or wait for analysts to build custom dashboards. No data leaves your account in the cloud, and there’s no need to manage separate vector databases or complex retrieval-augmented generation (RAG) pipelines.

In this article, we’ll provide an in-depth technical overview of what Snowflake Intelligence is, how it works under the hood, how to set it up step by step, what you can do with it, its limitations and so much more.

What is Snowflake Intelligence?

Snowflake Intelligence is an enterprise AI agent platform built into Snowflake. It’s designed to let any user talk to their data in plain English without writing complex SQL and get instant insights and visualizations. Under the covers, Snowflake Intelligence spins up Snowflake Cortex Agents – AI agents that orchestrate multiple services (LLMs, search, SQL) to answer questions or take actions. These agents connect to your structured tables (via the Snowflake Cortex Analyst tool and Snowflake semantic models), unstructured documents (via the Snowflake Cortex Search feature and embeddings) and any custom tools (via stored procedures or integrations) you configure. The agent can then translate a user question into the right mix of SQL queries, semantic search queries and actions and combine the results into a response or chart.

Figure 1: Snowflake Intelligence (Source: Snowflake)

Key elements and features of Snowflake Intelligence

Here are some key features of Snowflake Intelligence:

• Data analysis through LLM-to-SQL—Snowflake Cortex Analyst translates natural language questions into SQL queries against your structured data. It’s using Snowflake semantic models defined in YAML files to understand business terminology and generate accurate SQL that respects your schema relationships.
• Data insight generation—The Snowflake Intelligence uses retrieval-augmented generation (RAG) combined with summarization to extract insights from unstructured data. You can upload PDFs, Microsoft Word documents or spreadsheets to Snowflake stages, index them with the Snowflake Cortex Search feature, and the agent can answer questions about their contents alongside structured data.
• Custom actions—You’re not only limited to read-only queries. Agents can execute stored procedures, call external APIs via webhooks or trigger automated workflows.
• Managed embeddings and vector search—The Snowflake Cortex Search feature handles all the embedding generation and infrastructure for semantic search. You’ll point it at a table column containing text, and it builds a hybrid index (vector embeddings and keyword search) with automatic semantic reranking.
• In-Snowflake LLM and embedding functions—Snowflake Cortex AI Functions let you call large language models (LLMs) directly in SQL for tasks like sentiment analysis, text classification and summarization.
• Agent orchestration—Snowflake Cortex Agents handle the planning layer. Whenever you ask a complex question requiring both structured and unstructured data, the agent parses your request, splits it into subtasks, routes queries to the appropriate tools and synthesizes a coherent response.
• Governance, auditability and RBAC—Every query respects your existing role-based access control (RBAC). If a user can’t see certain data in Snowflake, the Snowflake Intelligence won’t expose it. All queries are logged, auditable and governed by the same security framework as the rest of your Snowflake environment.
• Multi-model support with automatic selection—The Snowflake Intelligence supports leading LLMs, including Anthropic Claude models and OpenAI GPT models. You can set the agent to “Auto” by default, and Snowflake will choose the best available model (currently that’s Claude 4.5/4.0/3.7/3.5 and GPT 5/4.1). But if your region doesn’t have a model right there, you can switch on cross-region inference for access.
• On-the-fly insights and visualizations—Beyond just raw answers, the agent can create charts or tables from query results. The Data-to-Chart tool in an agent can make Vega-Lite charts on the spot, so you can see data trends visually without needing to set up a dashboard manually.
• Conversation threading—Interactions are stateful. The Snowflake Intelligence platform supports threaded conversations so an agent remembers context. You can follow up or refine your queries naturally, and the system uses the thread history to maintain continuity.

Snowflake Intelligence puts all these pieces together so end users can explore data by asking questions, without the need to code or wait for business intelligence (BI) teams.

Cost breakdown of the Snowflake Intelligence

Snowflake Intelligence doesn’t have its own pricing tier. Instead, you pay for the underlying services it uses. There are three main cost drivers:

• Snowflake Cortex Analyst requests (for structured data queries)
• Snowflake Cortex Search feature usage (for unstructured data searches)
• Snowflake Virtual Warehouse execution (for running the generated SQL)

Let’s break down the cost of each one.

1) Snowflake Cortex Analyst cost

Snowflake Cortex Analyst tool uses message-based pricing when accessed through the Snowflake Intelligence interface.

Snowflake Cortex Analyst costs ~67 credits per 1000 messages when used through the Snowflake Cortex Analyst API directly.

But when you’re using it via the Snowflake Intelligence platform, the pricing shifts to a token-based model per million tokens:

Table 1: Credit Consumption Rates for Snowflake Cortex / Snowflake AI Models (Per 1M Tokens) 

When using Snowflake Cortex Analyst through Snowflake Cortex Agents (which powers Intelligence), credit rates are slightly different:

Table 2: Credit Consumption Rates for Snowflake AI Models with Caching (Per 1M Tokens) 

Notice the markup? Using Snowflake Cortex Analyst through Intelligence adds roughly 25% to the base token costs. That’s the price of convenience.

Here’s what those credits actually cost based on your Snowflake edition and region (using US East Northern Virginia on AWS as the baseline):

• Standard edition: $2.00 per credit
• Enterprise edition: $3.00 per credit
• Business Critical edition: $4.00 per credit
• VPS edition: $6.00 per credit

If you’re running claude-sonnet-4-5 on Enterprise Edition through Snowflake Intelligence, you’re paying:

• Input: 3.45 credits × $3.00 = $10.35 per million input tokens
• Output: 17.26 credits × $3.00 = $51.78 per million output tokens

2) Snowflake Cortex Search

Snowflake Cortex Search handles queries against unstructured data. The pricing model is more complex because you’re paying for both indexing and searching.

Snowflake Cortex Search costs 6.3 credits per GB/month of indexed data.

But that’s just the serving compute. You also pay for:

• Embedding tokens during indexing
• Snowflake virtual warehouse compute for refreshes
• Storage for the indexes ($23/terabyte (TB) per month base rate)
• Cloud services costs

A 100GB search index costs 630 credits/month just for serving compute. On Enterprise Edition, that’s $1890/month before you’ve run a single search query.

3) Snowflake virtual warehouse execution costs

Here’s what many teams miss: the Snowflake Cortex Analyst tool generates SQL, but you still pay for executing that SQL. The generated queries run on your Snowflake virtual warehouses, consuming credits based on:

• Snowflake virtual warehouse size
• Execution time
• Query complexity

Standard Snowflake virtual warehouse sizes:

Table 3: Credit Consumption Rates for Snowflake Virtual Warehouse Sizes 

If your Intelligence-generated queries run on a Medium Snowflake Virtual warehouse and take 5 minutes to execute, that’s (4 credits/60 minutes) × 5 minutes = 0.33 credits per query. Add this to the Analyst costs.

Here’s the breakdown of what it actually costs to run Snowflake Intelligence:

Let’s say your team runs 10,000 Intelligence queries per month on the Enterprise Edition using claude-sonnet-4-5 model.

Here’s the cost breakdown:

Snowflake Cortex Analyst costs (assuming avg ~250 input + 750 output tokens per query):

• Input: (250 tokens * 10000 queries) / 1000000 = 2.5M tokens * 3.45 credits = 8.625 credits * $3 = $25.88
• Output: (750 tokens × 10000 queries) / 1000000 = 7.5M tokens * 17.26 credits = 129.45 credits * $3 = $388.35

Snowflake Virtual Warehouse execution (assuming Medium warehouse, 30 seconds avg per query):

• (4 credits/hour) * (30 seconds/3600 seconds) * 10000 queries = 33.33 credits * $3 = $100

Snowflake Cortex Search (if using a 50GB index):

• Serving compute: 6.3 credits * 50 GB = 315 credits * $3 = $945/month

Monthly total: $1,459.23

That’s for pretty average use. Now scale it up to 100k queries a month and you’re looking at ~$13,092.30 (not counting the extra costs for embedding, storing all that data, or actually running the searches).

Why use Snowflake Intelligence?

Simple. Secure. Fast.

Data stays inside Snowflake. Inference and embedding generation run within the Snowflake platform’s execution environment or through vetted model integrations that honor Snowflake access controls.

Snowflake handles embedding generation, vector storage, model inference, and orchestration. Your team can focus only on Snowflake semantic models and agent logic, not infrastructure operations.

Responses return in seconds.

The only main downside is vendor lock-in. If data lives outside Snowflake you must move it or build connectors. Swapping in a custom fine-tuned model requires integrating external model endpoints or using an approved execution path, which adds complexity.

Architecture overview of Snowflake Intelligence

Snowflake Intelligence is built on a modular agent-based architecture within the Snowflake Cortex AI suite.

Figure 2: Snowflake Intelligence Architecture

User interface layer

The User Interface layer is the Snowsight web interface (under AI & ML > Agents > Snowflake Intelligence or the separate Snowflake Intelligence UI). Users type questions into a chat UI and see answers, charts, and the agent’s reasoning. Admins can configure agents here, too. The interface is simple (search or chat), but it connects to a rich backend for handling complex queries.

Orchestration layer (Snowflake Cortex Agent)

Snowflake Cortex Agents sit at the core of the system. When a user asks a question, the agent doesn’t immediately query data or search documents. Instead, it plans. Cortex Agent dynamically plans a multi-step workflow: for example, it might first fetch some context with a search query, then fire a SQL query via Snowflake Cortex Analyst, then maybe transform results with a function, and finally compose a natural-language answer. The planning loop constantly updates as new data comes in.

Tool integration layer

The agent invokes a variety of built-in and custom tools to execute tasks.

Snowflake Cortex Analyst

Snowflake Cortex Analyst is the LLM-powered text-to-SQL engine. It converts natural language questions into optimized SQL queries against structured tables. It relies on Snowflake semantic models to map natural language to table/column names, metrics, and business concepts. Snowflake builds an internal schema-aware model from your database for better understanding. Queries execute on live data using your specified Snowflake Virtual warehouses.

Snowflake Cortex Search

Snowflake Cortex Search is a fully managed hybrid search service. It combines vector embeddings (for semantic similarity) with keyword search (for exact matches) and applies a neural reranker to surface the most relevant results. You create a Cortex Search service by pointing it at a table column containing text. Snowflake automatically generates embeddings, builds an index, and keeps it synchronized as data changes. You can filter results by metadata (like document type or date range) and configure which embedding model to use.

Data visualization (Data-to-Chart)

An agent can use a built-in tool to render query results as charts. If a user’s question would benefit from a graph, the agent auto-generates a Vega-Lite chart. This tool is enabled by adding data_to_chart in the agent’s configuration.

Custom Tools and Stored Procedures

You can integrate custom tools via Snowflake stored procedures or user-defined functions (UDFs) for custom logic, such as refreshing an external table or calling an HTTP webhook to an external service. These are added as steps in the agent’s reasoning flow, with configurations for name, description, parameters and warehouse. Grant USAGE privileges to enable access.

Snowflake Cortex AI Functions (LLM and embeddings in SQL)

Snowflake Cortex AI Functions are the built-in SQL functions that let the agent (or any user) run LLM operations directly. For example, agents might call AI_COMPLETE for a chat completion, or AI_EMBED for embedding text to compare similarity. These functions are optimized for batch workloads.

Data layer

Data layer contains your actual data:

• Structured data in Snowflake tables and views
• Unstructured data (PDFs, documents, images) stored in Snowflake stages
• Snowflake semantic models that map business concepts to physical schemas for Cortex Analyst
• Search indices managed by Cortex Search, including Snowflake Cortex Knowledge Extensions for external or shared content

Infrastructure and runtime layer

At the bottom sits Snowflake’s data platform infrastructure. Snowflake handles compute, storage, and model serving. Snowflake Virtual warehouses execute SQL queries. Serverless compute runs AI functions. The Cortex runtime manages LLM inference, embedding generation, and agent orchestration without you needing to provision anything.

TL;DR: Snowflake Intelligence is essentially Cortex Agents + Cortex Analyst + Cortex Search + Cortex AI Functions all working together in a secure Snowflake environment. Workflow:

1) User asks a question via the UI (ai.snowflake.com or Snowsight)

2) Snowflake Cortex Agent receives the query and creates an execution plan using LLMs

3) Agent routes tasks to appropriate tools:

• Snowflake Cortex Analyst for SQL queries on structured data
• Snowflake Cortex Search for semantic retrieval from unstructured data
• Custom tools for specialized logic or external integrations
• Data-to-Chart for visualization

4) Each step executes as a regular Snowflake operation (SQL query or AI function call)

5) Agent synthesizes results and presents answer with charts/visualizations 6) All operations benefit from Snowflake’s scalability, security, and governance

Now that we’ve covered what Snowflake Intelligence is and how it works, let’s dive into how to configure and run it from scratch.

Step-by-step guide to configuring and running the Snowflake Intelligence

Let’s walk through the full setup from scratch. Make sure you have a Snowflake account with administrative access.

Prerequisites and setup

Before starting, verify these requirements:

• Regional availability. Snowflake Intelligence is available across AWS, Azure, and GCP regions. If a model isn’t available in your region, enable Cortex Cross-region inference to route requests to regions where the model is hosted.
• Supported Snowflake AI models. Snowflake Intelligence supports the following models => Claude 4.5, Claude 4.0, Claude 3.7, Claude 3.5, GPT 5, GPT 4.1.
• Auto model selection is recommended to let the agent choose the best model for each query.
• Administrative privileges. The ACCOUNTADMIN role is required to create the Snowflake Intelligence object, as it holds the exclusive CREATE SNOWFLAKE INTELLIGENCE ON ACCOUNT privilege.
• Defined Snowflake roles following least-privilege principles. Plan to create custom Snowflake roles like SI_ADMIN, CORTEX_USER, and CORTEX_EMBED_USER for different access levels.
• Provisioned virtual warehouses. You’ll need active warehouses to execute SQL queries generated by Snowflake Cortex Analyst.
• Data infrastructure. Structured data modeled in tables, unstructured data stored in Snowflake stages, and prepared Snowflake semantic model YAML files or Snowflake semantic views.
• Basic knowledge of Snowflake SQL for roles, databases, schemas, tables, Snowflake stages, procedures, and YAML for semantic models.
• Private Connectivity (optional). For private network access, complete AWS PrivateLink or Azure Private Link setup before proceeding.

Step 1—Log in to the Snowflake platform

Sign in to Snowflake’s web UI (Snowsight) using an admin role. You’ll do most of the initial setup here.

Step 2—Create Snowflake roles, databases and Snowflake Virtual warehouses

Open a SQL worksheet in Snowflake Snowsight and execute the following setup commands:

USE ROLE ACCOUNTADMIN;  
CREATE ROLE IF NOT EXISTS SI_ADMIN;  
GRANT ROLE SI_ADMIN TO USER <-----your_username------->;

Figure 3: Creating and granting privilege to SI_ADMIN role – Snowflake Intelligence

CREATE DATABASE IF NOT EXISTS intelligence_db; 

CREATE SCHEMA IF NOT EXISTS intelligence_db.intelligence_schema;

Figure 4: Creating Snowflake database and schema – Snowflake Intelligence

CREATE WAREHOUSE IF NOT EXISTS intelligence_wh   
WITH WAREHOUSE_SIZE = 'SMALL'    
AUTO_SUSPEND = 60    
AUTO_RESUME = TRUE    
INITIALLY_SUSPENDED = TRUE;

Figure 5: Creating Snowflake Virtual Warehouse – Snowflake Intelligence

Grant appropriate privileges on the database, schema and warehouse to your Snowflake roles.

Step 3—Initialize the Snowflake Intelligence object

You can now create the Snowflake Intelligence object via the Snowflake Snowsight interface or SQL.

Via the Snowflake Snowsight interface:

Navigate to AI & ML > Agents > Snowflake Intelligence tab.

Figure 6: Navigating to Snowflake Intelligence

Figure 7: Navigating to Snowflake Intelligence

Open settings. The Snowflake Intelligence object is created automatically when you access settings for the first time.

Via SQL:

CREATE SNOWFLAKE INTELLIGENCE SNOWFLAKE_INTELLIGENCE_OBJECT_DEFAULT; 

Only one Snowflake Intelligence object can exist per account. This object acts as a container holding a curated list of agents available to users.

Step 4—Configure cross-region inference (optional)

If you need access to models not available in your region, enable cross-region inference.

First, identify your region category (aws_us, azure_us):

SELECT CURRENT_REGION(); 

Figure 8: Identifying Snowflake Region – Snowflake Intelligence

Then configure cross-region inference:

ALTER ACCOUNT SET CORTEX_ENABLED_CROSS_REGION = 'ANY_REGION';

(Or specify regions like ‘AWS_US, AWS_EU’).

ALTER ACCOUNT SET CORTEX_ENABLED_CROSS_REGION = 'AWS_US';

This allows the Snowflake Intelligence platform to transparently call models deployed elsewhere. Without this, you’re stuck with whatever the local region offers.

Figure 9: Configuring cross-region inference – Snowflake Intelligence

Note that cross-region inference incurs additional data transfer costs.

Step 5—Grant Snowflake Intelligence privileges and access control

Management rights:

GRANT MODIFY ON SNOWFLAKE INTELLIGENCE SNOWFLAKE_INTELLIGENCE_OBJECT_DEFAULT  TO ROLE SI_ADMIN;

Figure 10: Granting MODIFY privileges on the default Snowflake Intelligence object to role SI_ADMIN

Usage rights for end users:

CREATE ROLE IF NOT EXISTS SI_USER;

Figure 11: Creating the SI_USER role for Snowflake Intelligence

GRANT USAGE ON SNOWFLAKE INTELLIGENCE SNOWFLAKE_INTELLIGENCE_OBJECT_DEFAULT  TO ROLE SI_USER;

Figure 12: Granting USAGE privileges on the default Snowflake Intelligence object to role SI_USER

Database roles for Snowflake Cortex services:

GRANT DATABASE ROLE SNOWFLAKE.CORTEX_USER TO ROLE SI_USER;

Figure 13: Granting the SNOWFLAKE.CORTEX_USER database role to role SI_USER—Snowflake Intelligence

GRANT DATABASE ROLE SNOWFLAKE.CORTEX_EMBED_USER TO ROLE SI_USER;

Figure 14: Granting the SNOWFLAKE.CORTEX_EMBED_USER database role to role SI_USER—Snowflake Intelligence

Step 6—Create a Snowflake Cortex Agent

Agents must be created before they can be added to the Snowflake Intelligence object.

Navigate to AI & ML > Agents > Create agent in Snowflake Snowsight.

Figure 15: Navigating to Agents in Snowflake Snowsight to create a new agent—Snowflake Intelligence

Select the Database and Schema, then enter the Agent object name. Next, add the name in the Display Name field, and finally, click on Create Agent.

Figure 16: Configuring agent name and details to create a new agent—Snowflake Intelligence

Provide an agent name, description, questions and configure tools (Snowflake Cortex Analyst, Snowflake Cortex Search services, custom tools). Then click Save.

Figure 17: Configuring agent details, questions, and tools in Snowflake Snowsigh—Snowflake Intelligence

Figure 18: Configuring agent details, questions, and tools in Snowflake Snowsight—Snowflake Intelligence

Agents are NOT automatically added to the Snowflake Intelligence object; you must add them explicitly in the next step.

Step 7—Configure agent visibility

Control which agents appear in the Snowflake Intelligence interface:

To add an agent navigate to AI & ML > Agents > Snowflake Intelligence tab > Open settings > Review agent.

Select the agent.

Figure 19: Selecting an agent for review in the Snowflake Intelligence settings

To remove an agent, click on that kebab menu and select “Remove agent.”

Figure 20: Removing an agent using the more actions menu— Snowflake Intelligence

Alternatively, you can also easily configure it via SQL directly:

ALTER SNOWFLAKE INTELLIGENCE SNOWFLAKE_INTELLIGENCE_OBJECT_DEFAULT ADD AGENT my_db.my_schema.SalesAgent;

To remove it, you can use DROP AGENT. Grant USAGE on each agent object to any role that should query it. Note that all agents you create in an account are visible to any user with access unless you curate the list. Creating a curated list (by populating the intelligence object) gives better control.

Users see agents based on this hierarchy: if the Snowflake Intelligence object exists and they have USAGE on it, they see only agents in the curated list. If no curated list exists, they see all agents they have individual access to.

Step 8—Configure private connectivity (advanced)

If your Snowflake platform is set up in private mode, you might need a PrivateLink connection to allow AI services. On the AWS cloud platform, ensure an AWS PrivateLink endpoint exists for Snowflake. On the Microsoft Azure cloud service, ensure Azure Private Link is configured for your account. These steps are outside the Snowflake Intelligence feature itself; they’re the same steps used for Snowflake’s Snowpark Container Services or native apps.

See the Snowflake documentation below on the PrivateLink feature for a more in-depth guide.

• AWS PrivateLink and Snowflake
• Azure Private Link and Snowflake

Once the PrivateLink feature is set up, the Snowflake Cortex AI inference APIs can be reached from your network.

Step 9—Customize the Snowflake Intelligence interface

From AI & ML > Agents > Snowflake Intelligence > Open settings, you can:

• Configure general settings like display name and description

Figure 21: Configuring general Snowflake Intelligence settings in Snowflake Snowsight

• Customize appearance and branding

Figure 22: Customizing appearance and branding—Snowflake Intelligence

• Add sample questions to guide users

Save your configuration and verify that agents appear correctly for end users.

Step 10—Final fix

With the above done, test an agent: switch to a role with access, open a “New Chat” in the Snowflake Snowsight interface, pick your agent and ask a question. Verify it returns sensible answers or charts. Something like this:

Figure 23: Testing an agent’s response in the Snowflake Snowsight chat interface—Snowflake Intelligence

If something is wrong, check the following:

• User has the correct role and Snowflake virtual warehouse (Snowflake Intelligence uses the user’s default role and warehouse by default)
• Agent has the correct tool configurations (semantic view, search service)
• Permission errors in the agent’s log (use SHOW AGENTS and DESCRIBE AGENT to inspect)
• Account usage or logs for any Cortex AI errors

If you encounter any issues, Snowflake documentation and community forums are good resources. But in many cases, it’s just a matter of adjusting permissions or tool settings. Once it’s all OK, you can easily start chatting with the agent and getting answers from your data in the cloud.

Example: Identifying top-performing products using the Snowflake Intelligence

Now, let’s walk through a complete example of building an agent that answers questions about product performance. We’ll outline thorough steps and objects you’ll set up.

Phase A—Prerequisites and global setup

Assume you’ve completed Steps 1 to 5 from the configuration guide above. If your region doesn’t natively support your preferred model, enable cross-region inference as shown in Step 4.

Phase B—Objects, data and staging

Step 1—Create role, users, database, schema and warehouse

First, we need to build the sandbox. Let’s start by creating the administrative backbone: Snowflake roles, users and the containers for our data. Using a dedicated role like product_analyst helps keep your permissions clean and follows the principle of least privilege.

USE ROLE ACCOUNTADMIN; 

CREATE ROLE IF NOT EXISTS product_analyst; 

CREATE USER IF NOT EXISTS analyst_user PASSWORD='Password@123!' DEFAULT_ROLE=product_analyst; 

GRANT ROLE product_analyst TO USER analyst_user;

Figure 24: Provisioning the product_analyst role and user—Snowflake Intelligence

Next, set up the database and schema. Always keep the analytics layer separate so things don’t get messy later.

CREATE DATABASE IF NOT EXISTS sales_db; 

CREATE SCHEMA IF NOT EXISTS sales_db.analytics; 

USE SCHEMA sales_db.analytics;

Figure 25: Defining the sales database and schema—Snowflake Intelligence

Finally, we need some computing power. A MEDIUM Snowflake virtual warehouse handles these operations easily, and the auto-suspend feature keeps your credits from draining.

CREATE WAREHOUSE IF NOT EXISTS analytics_wh 

 WAREHOUSE_SIZE = 'MEDIUM' 

 AUTO_SUSPEND = 300 

 AUTO_RESUME = TRUE;

Figure 26: Creating the analytics Snowflake virtual warehouse—Snowflake Intelligence

Step 2—Create tables and load sample data

A database is just an empty shell without data. Let’s create a products table to house our catalog. Notice that we are just using standard data types to help keep it simple, which makes the LLM’s job easier.

CREATE OR REPLACE TABLE products ( 

  "product_id"   INT, 

  "product_name" STRING, 

  "category"     STRING, 

  "price"   DECIMAL(10,2) 

);

INSERT INTO products ("product_id", "product_name", "category", "price") VALUES 

  (1,  'Apple iPhone 16',              'Phones',     799.99), 

  (2,  'Apple iPhone 16 Pro Max',      'Phones',    1199.99), 

  (3,  'Samsung Galaxy S25 Ultra',     'Phones',    1299.99), 

  (4,  'Samsung Galaxy A16 5G',        'Phones',     199.99), 

  (5,  'Google Pixel 10 Pro',          'Phones',     999.99), 

  (6,  'Apple iPhone 17',              'Phones',     829.99), 

  (7,  'Samsung Galaxy S25',           'Phones',     799.99), 

  (8,  'Apple AirPods Pro 3',          'Earphones',  249.99), 

  (9,  'Sony WF-1000XM5',              'Earphones',  299.99), 

  (10, 'Bose Comfort Ultra Earbuds',   'Earphones',  249.99), 

  (11, 'Samsung Galaxy Buds 3 Pro',    'Earphones',  229.99), 

  (12, 'Google Pixel Buds Pro 2',      'Earphones',  229.99), 

  (13, 'Apple AirPods 4',              'Earphones',  129.99), 

  (14, 'Anker Soundcore Liberty 4 NC','Earphones',   99.99), 

  (15, 'Nothing Ear (2025)',           'Earphones',  149.99);

Figure 27: Populating the products table—Snowflake Intelligence

Now, let’s add some transaction history. This sales table tracks what was sold, when, and for how much.

CREATE OR REPLACE TABLE sales ( 

  "sale_id"      INT, 

  "product_id"   INT, 

  "sale_date"    DATE, 

  "quantity"     INT, 

  "total_amount" DECIMAL(10,2) 

);

INSERT INTO sales ("sale_id", "product_id", "sale_date", "quantity", "total_amount") VALUES 

  (1,  1,  '2026-01-10',  1,   799.99), 

  (2,  2,  '2026-01-12',  1,  1199.99), 

  (3,  5,  '2026-01-15',  2,  1999.98), 

  (4,  8,  '2026-01-18',  3,   749.97), 

  (5,  4,  '2026-01-20',  1,   199.99), 

  (6, 11,  '2026-01-22',  4,   919.96), 

  (7,  2,  '2026-01-25',  1,  1199.99), 

  (8, 13,  '2026-01-28',  5,   649.95), 

  (9,  9,  '2026-02-02',  2,   599.98), 

  (10,14, '2026-02-05', 10,   999.90);

Figure 28: Adding transaction records to the sales table—Snowflake Intelligence

Step 3—Create Snowflake internal stage and upload Snowflake semantic model YAML file

Snowflake Intelligence needs a map to understand your SQL schema. This map is a YAML file called a semantic model. First, let’s create a Snowflake stage, which is essentially a folder inside Snowflake, to hold that file.

CREATE OR REPLACE STAGE semantic_models;

Figure 29: Creating the semantic_models Snowflake internal stage—Snowflake Intelligence

Once you have created your Snowflake Stage, upload the following YAML file to that Snowflake stage. You can either use the Snowflake Snowsight UI or the PUT command. In this guide, we will upload it via Snowsight. To do so, navigate to Catalog > Database Explorer in the sidebar, select your database and schema, then click on Stages and select your Snowflake stage. Click the + Files button to upload files.

Figure 30: Uploading the YAML file via Snowflake Snowsight UI—Snowflake Intelligence

Here’s what your Snowflake semantic model YAML should look like:

name: SALES_ANALYSIS 

description: Semantic model for product sales analysis 

tables: 

  - name: PRODUCTS 

    description: Product catalog with pricing 

    base_table: 

      database: SALES_DB 

      schema: ANALYTICS 

      table: PRODUCTS 

    primary_key: 

      columns: 

        - PRODUCT_ID 

    dimensions: 

      - name: PRODUCT_NAME 

        synonyms: 

          - product 

          - item 

        description: Name of the product 

        expr: PRODUCT_NAME 

        data_type: STRING 

      - name: CATEGORY 

        synonyms: 

          - product category 

          - type 

        description: Product category 

        expr: CATEGORY 

        data_type: STRING 

  - name: SALES 

    description: Sales transactions 

    base_table: 

      database: SALES_DB 

      schema: ANALYTICS 

      table: SALES 

    primary_key: 

      columns: 

        - SALE_ID 

    dimensions: 

      - name: SALE_DATE 

        description: Date of sale 

        expr: SALE_DATE 

        data_type: DATE 

    metrics: 

      - name: TOTAL_REVENUE 

        description: Total sales revenue 

        expr: SUM(TOTAL_AMOUNT) 

      - name: UNITS_SOLD 

        description: Total units sold 

        expr: SUM(QUANTITY) 

relationships: 

  - name: SALES_TO_PRODUCTS 

    left_table: SALES 

    left_column: PRODUCT_ID 

    right_table: PRODUCTS 

    right_column: PRODUCT_ID

Save the above YAML file as sales_analysis.yaml on your PC. Once that’s done, navigate to the upload interface, select the Database, schema, and stage that you created earlier, upload that YAML file to that Snowflake stage, and click upload.

Figure 31: Uploading the YAML file via Snowflake Snowsight—Snowflake Intelligence

OR

Use the PUT command from SnowSQL:

PUT file://sales_analysis.yaml @SEMANTIC_MODELS AUTO_COMPRESS=TRUE;

Phase C—Snowflake Cortex setup (semantic layer, search and agent)

Step 1—Configure the Snowflake Cortex Search feature (optional for unstructured data)

Now we turn that YAML file into a living “Semantic View.” This is a schema-level object that keeps your business logic inside the database where it belongs.

You have two options to create the semantic view:

Option 1—Create from YAML file

Create the semantic view from the YAML:

CALL SYSTEM$CREATE_SEMANTIC_VIEW_FROM_YAML( 

  'sales_db.analytics', 

  GET_PRESIGNED_URL(@semantic_models, 'sales_analysis.yaml'), 

  FALSE 

);

OR

Option 2—Create directly in SQL

If you prefer keeping everything in a single SQL script, you can define the view directly. It does the same thing: it maps metrics and dimensions so the LLM doesn’t have to guess what columns mean.

CREATE OR REPLACE SEMANTIC VIEW sales_analysis 

TABLES ( 

  products AS sales_db.analytics.products PRIMARY KEY ("product_id"), 

  sales AS sales_db.analytics.sales PRIMARY KEY ("sale_id") 

) 

DIMENSIONS ( 

  products.product_name AS "product_name", 

  products.category AS "category", 

  sales.sale_date AS "sale_date" 

) 

METRICS ( 

  sales.total_revenue AS SUM("total_amount"), 

  sales.units_sold AS SUM("quantity") 

) 

COMMENT = 'Semantic view for analyzing phone and earphone sales data'

Figure 32: Generating the sales_analysis semantic view—Snowflake Intelligence

Both approaches create a semantic view, which is a schema-level object that stores business concepts directly in the database. Snowflake semantic views address the gap between how business users describe data and how it’s stored in database schemas. They’re metadata objects that improve LLM accuracy by combining reasoning with rule-based definitions.

Step 2—Configure Snowflake Cortex Search (Optional for Unstructured Data)

If you have product documentation or reviews in unstructured format, you can set up Snowflake Cortex Search to enable semantic search over text data. This is useful when you want to search through product reviews, descriptions, or support documentation.

First, create a table for unstructured content:

CREATE OR REPLACE TABLE product_reviews ( 

  review_id INT, 

  product_id INT, 

  review_text STRING, 

  review_date DATE 

);

Figure 33: Creating the product_reviews table—Snowflake Intelligence

ALTER TABLE product_reviews SET CHANGE_TRACKING = TRUE;

Figure 34: Enabling change tracking for incremental search updates—Snowflake Intelligence

Note: Change tracking must be enabled with non-zero time travel retention on all underlying objects to support incremental refreshes.

Let’s throw in some sample data. Here, we’ve added a mix of praise and complaints to test the agent’s reasoning later.

INSERT INTO product_reviews (review_id, product_id, review_text, review_date) VALUES 

  (1, 1, 'Amazing phone with great camera quality and battery life. Highly recommended!', '2026-01-15'), 

  (2, 2, 'Premium device but expensive. Display is stunning and performance is top-notch.', '2026-01-18'), 

  (3, 8, 'Best noise cancellation I have experienced. Comfortable for long listening sessions.', '2026-01-20'), 

  (4, 9, 'Sound quality is exceptional. Battery life could be better but overall excellent.', '2026-01-22'), 

  (5, 14, 'Great value for money. Sound quality rivals premium brands at half the price.', '2026-02-08');

Figure 35: Inserting sample review records—Snowflake Intelligence

Now let’s provision the actual Snowflake Cortex Search service.

CREATE OR REPLACE CORTEX SEARCH SERVICE product_review_search 

  ON review_text 

  ATTRIBUTES review_id, product_id, review_date 

  WAREHOUSE = analytics_wh 

  TARGET_LAG = '1 hour' 

  AS ( 

    SELECT  

      review_id, 

      product_id, 

      review_text, 

      review_date 

    FROM product_reviews 

  );

• ON review_text: Specifies the column to index for full-text search.
• ATTRIBUTES: Lists additional columns you want to return in results or filter on when querying.
• WAREHOUSE: The Snowflake virtual warehouse used for running the source query, building the search index, and keeping it refreshed.
• TARGET_LAG: Defines how fresh the search results need to be. A value of 1 hour means the service will check for updates approximately once per hour.

Snowflake Cortex Search service uses a combination of vector embeddings and keyword matching. It uses the snowflake-arctic-embed-m-v1.5 embedding model by default, though you can specify a different model using the EMBEDDING_MODEL parameter.

Here are available Snowflake AI models:

• snowflake-arctic-embed-m-v1.5 (default, 768 dimensions)
• snowflake-arctic-embed-l-v2.0 (1024 dimensions, higher quality)
• snowflake-arctic-embed-s (384 dimensions, faster)
• Other supported models, depending on your region

Figure 36: Provisioning the Snowflake Cortex Search service—Snowflake Intelligence

Step 3—Create the Snowflake Cortex agent

The Agent is the “brain” that orchestrates everything. It decides whether a user’s question requires a SQL query (via Analyst) or a text search (via Search).

CREATE AGENT sales_agent 
  COMMENT = 'Agent for analyzing product sales data'

Figure 37: Defining the sales_agent object—Snowflake Intelligence

The rest of the configuration happens in the Snowflake Snowsight UI. Navigate to AI & ML > Agents > sales_agent to start plugging in your tools.

Figure 38: Opening the agent configuration in Snowflake Snowsight—Snowflake Intelligence

Click Edit in the top-right corner. Then navigate to the Tools tab and add Cortex Analyst and Cortex Search, which you configured earlier.

Figure 39: Accessing the agent Tools menu—Snowflake Intelligence

Adding Snowflake Cortex Analyst

For Cortex Analyst, click the +Add button and select the database and schema you configured previously. Once selected, you will see the sales_analysis semantic view that you created earlier, and choose it from the dropdown.

Next, provide a name for the tool, such as “Sales_Analyst.” For the description, you can either click “Generate with Cortex” to auto-generate one or write a description manually.

After that, go to the Warehouse section below. Click Custom and select your preferred Snowflake virtual warehouse. For this guide, we will use “analytics_wh.” Finally, set the Query Timeout to around 60 seconds, and then click Add.

TL;DR:

• Name: “Sales_Analyst”
• Semantic View: sales_analysis
• Warehouse: analytics_wh
• Query Timeout: 60 seconds

Figure 40: Configuring the Sales_Analyst tool settings— Snowflake Intelligence

Adding Snowflake Cortex Search

Now For Cortex Search, if you have configured Cortex Search do the same for that as well. Earlier, we configured Cortex Search for Product Reviews, so let’s set that up.

Click the +Add button and select the database and schema you configured previously. Once selected, you will see product_review_search, which you created earlier, choose it from the dropdown.

Next, provide a name for the tool, such as “Product_Reviews_Search.” Then, in the ID column, add Review_id, and in the Title column, add Review_text (or adjust according to your preference). Once completed, click Add.

TL;DR:

• Name: “Product_Reviews_Search”
• Search: product_review_search
• ID: Review_id
• Title: Review_text
• Query Timeout: 60 seconds

Figure 41: Adding the Product_Reviews_Search tool—Snowflake Intelligence

Once everything is configured, your Tools section should look like this:

Figure 42: Reviewing the completed tool configuration— Snowflake Intelligence

Finally, click Save to complete the configuration. You’re done.

Phase D—Verify, run and sample queries

Step 1—Navigate to the Snowflake Intelligence interface

Everything is wired up. Time to see if it works. Head over to AI & ML > Snowflake Intelligence (or just hit ai.snowflake.com).

Figure 43: Accessing the Snowflake Intelligence interface

Step 2—Enter natural language query

Try out these questions:

Question 1: What are the top 5 products by revenue?

Figure 44: Querying top products by revenue— Snowflake Intelligence

Question 2: Show me sales by category

Figure 45: Querying category-level sales— Snowflake Intelligence

Question 3: How many units of Apple iPhone 16 were sold?

Figure 46: Asking for specific product sales volume— Snowflake Intelligence

Question 4: Top performing products by revenue

Figure 47: Ranking revenue performance— Snowflake Intelligence

Question 5: What are customers saying about the Sony WF-1000XM5 in their reviews?

Figure 48: Analyzing customer sentiment through search—Snowflake Intelligence

And the best part? You can even ask follow-up questions. The agent maintains conversation context, so it understands these refer to the previous query.

Limitations of Snowflake Intelligence

Snowflake Intelligence is seriously powerful, but it has some key limitations you should be aware of.

1) Token and context window limitations

Each LLM has a fixed context window. Claude 3.5 Sonnet supports 200k tokens, but if you use 195k tokens for input, you can only generate 5k tokens of output (the total can’t exceed 200k).

For semantic models, a good rule of thumb is to keep it under 50-100 columns across all tables. Go beyond that, and you’ll probably run into latency issues or see a decline in quality. That’s because the whole semantic model has to be loaded into the LLM for every query.

2) Unpredictable costs

Consumption-based pricing sounds flexible until you get a surprise bill. Run one bad query using AI_COMPLETE on millions of rows, and you’ll be shocked at how fast your credits disappear. You can’t really set a hard limit to prevent costs from getting out of hand, so you just have to monitor them and hope for the best.

On top of that, cross-region inference also comes with data transfer charges that may not be clear at first. If your account is in Europe but your preferred model is in AWS US, every query will incur egress fees.

3) Snowflake-only data access

Snowflake Intelligence is confined to data within the Snowflake environment. It can’t directly query external databases or data sources outside Snowflake without first ingesting that data.

4) Limitations in query handling

Snowflake Cortex Analyst works best with well-defined semantic models. If your data model is messy, with issues like tables missing primary keys, inconsistent naming, or complex, unnormalized schemas, the quality of the generated SQL will suffer. The agent can’t be expected to fix poor data architecture on its own.

Multi-step reasoning has its limits. If you ask a question that needs more than 2-3 steps to answer, you will probably get an incomplete or incorrect response. Agents can break down tasks and use the right tools, but they are not good at complex planning that involves backtracking or testing ideas.

5) Language and document processing limitations

Snowflake Cortex Search’s embedding models are optimized for English. Non-English text gets lower-quality semantic search results. You can still search it, but relevance scores won’t be as accurate.

Document processing has size limits. Images must be smaller than 8000 * 8000 pixels. Text chunks are recommended to stay under 512 tokens for best retrieval quality. Larger documents need to be split before indexing.

6) Region-specific and integration limits

Not all features are available in all regions. Some models require cross-region inference, which adds latency and cost. If you’re in a region with limited Cortex support, you might not have the best experience.

Integration with external systems relies on stored procedures and webhooks. There’s no native connector for tools like Salesforce or HubSpot, so you’ll need to write custom code to connect them.

Snowflake Intelligence can’t directly access data behind authentication walls like private Google Docs or password-protected file shares. You have to copy that data into Snowflake stages first.

7) No separate logging for agent-generated queries

There’s no built-in way to distinguish queries generated by agents from manual SQL queries in query history. You must filter by Snowflake virtual warehouse or Snowflake role and plan your environment accordingly for tracking.

8) Accuracy depends on Snowflake semantic model quality

If semantic models are poorly defined or incomplete, the agent will generate incorrect SQL. The system is only as good as the business logic you encode.

9) Rate limiting and quotas

The Cortex REST API has rate limits: tokens per minute (TPM) and requests per minute (RPM). If you hit these limits, you get a 429 error and have to retry. High-throughput applications need to implement backoff logic.

Query timeouts can kill long-running SQL generation. If Cortex Analyst generates a query that takes more than your configured timeout (default 60 seconds), it fails without returning partial results.

Conclusion

Snowflake Intelligence is a really powerful feature that Snowflake released this year. It brings state-of-the-art AI agents to your data warehouse, with LLM-powered reasoning, visualization, and action tools all in one place. The best part’s that it stays within the secure Snowflake environment, so your data never leaves and still follows your governance rules.

Now, engineers and developers can easily build AI features on top of their data by setting up Snowflake semantic models and agent tools and letting the Snowflake Intelligence feature handle the rest. It’s not magic, of course. You still need to set up your semantic models, create data indexes and monitor performance and costs. Once everything’s in place, the Snowflake Intelligence feature works like a super smart data assistant. You can chat with it, and it comes up with SQL on the fly, digs up hidden details in documents and even kicks off workflows with just your voice or text. It’s basically a layer above your data cloud, turning your questions into answers in an instant.

In this article, we’ve covered:

• What is Snowflake Intelligence?
• Key features of Snowflake Intelligence
• Cost breakdown of the Snowflake Intelligence feature
• Architecture overview of the Snowflake Intelligence
• Step-by-step guide to configuring and running the Snowflake Intelligence feature
• Example: Identifying top-performing products using the Snowflake Intelligence feature
• Limitations of the Snowflake Intelligence platform

… and so much more!

Want to learn more? Reach out for a chat

FAQ

In today’s rapidly growing, ever-evolving tech landscape, organizations face unprecedented complexity in managing their IT environments. The proliferation of SaaS, cloud, and AI-driven solutions, coupled with decentralized procurement and rising security threats, has made IT Asset Management (ITAM) more critical, and more challenging than ever before.

Yet, despite its strategic importance, ITAM programs often struggle to secure executive sponsorship, cross-functional buy-in, and investment needed to deliver maximum value. Selling ITAM internally is not just about justifying a tool or process, it’s about positioning ITAM as a business enabler that drives measurable outcomes across cost, risk, and operational efficiency in SaaS management.

Why ITAM, why now?

The urgency for robust ITAM and a solid ITAM strategy has never been greater. Key technology trends are reshaping the IT landscape:

• Distributed procurement: IT typically purchases enterprise-level productivity tools, but the bulk of technology purchasing is distributed across line of business budgets. Departments acquire technology through direct, partner, pay-as-you-go, and marketplace models, increasing the risk of shadow IT and redundant spend.
• SaaS and AI sprawl: The ease of acquiring SaaS and AI tools has led to uncontrolled growth, making it difficult to track usage, manage costs, and ensure compliance.
• Cloud complexity: New licensing models (PAYG, BYOL) and hybrid environments require sophisticated tracking and optimization.
• Security pressures: With the rise of cyber threats, IT security is a board-level concern. Unmanaged assets and unknown applications increase vulnerability. AI applications and agents make this challenge even more difficult.
• Cost and ROI scrutiny: The cost of AI and cloud infrastructure is significant, and CIOs are under pressure to demonstrate ROI and control budgets to enhance cloud optimization and cost optimization.

According to Gartner’s 2025 Magic Quadrant for SaaS Management Platforms, “Per-employee SaaS spend currently averages $1,562, a 51% increase since 2022. As many as 25% of provisioned SaaS licenses are not regularly used by employees. Organizations are generally only aware of 40% of applications in use.” These statistics underscore the urgent need for visibility, control, and optimization which are core tenets of ITAM.

How do you connect ITAM to business goals?

To sell ITAM internally, it’s essential to align the program with your organization’s strategic objectives. Common drivers include:

• Mergers and acquisitions: Staying compliant, rationalizing technologies, and removing technical debt and risk.
• IPO readiness: Achieving SOC2 Type 2 compliance and meeting regulatory standards.
• Cost optimization: Freeing up budget for innovation (e.g., AI investments) or improving margins.
• Security enhancement: Addressing a recent breach or strengthening cyber defenses.
• Regulatory compliance: Meeting evolving standards and audit requirements.

Ask yourself: What are your company’s top priorities? How can ITAM help achieve them? Tailoring your message to these goals makes ITAM relevant to executive stakeholders in SaaS management.

What are the questions that ITAM must answer?

A compelling ITAM program provides clear answers to critical questions:

• What assets do we have, and where are they?
• Who is using them, and how much are they used?
• What are the costs, and risks associated with our technology portfolio?
• Are we getting value from these assets? Should we keep this technology, expand its use or retire its use?

These questions form the foundation for building trust and credibility with business leaders.

Defining success: Which KPIs and metrics are most important in ITAM?

To gain and sustain support, ITAM and the ITAM strategy must demonstrate measurable impact. Key performance indicators include:

• Percentage of environment inventoried
• Percentage of licenses unallocated and unused
• Reduction in unused/redundant applications and contracts
• Inventory of contracts, invoices, and product use rights
• Savings during renewals/keeping renewals flat
• Reduction in EOL/EOS/vulnerable assets
• Elimination of license compliance risk
• Removal of deny-listed and shadow applications

Tracking and reporting on these KPIs shows progress and justifies ongoing investment.

Scoping ITAM: Where do you focus?

ITAM’s scope can be defined by technology domains (on-premises, SaaS, cloud, AI, containers) or organizational units (IT, engineering, business units). It’s important to clarify where ITAM operates and where it intersects with other functions, such as FinOps (focused on cloud infrastructure spend) or security teams. Highlighting these boundaries helps identify collaboration opportunities and gaps to address.

Gap Analysis: the path to maturity

A structured gap analysis compares the current state to the desired future state across key metrics. By technology area (hardware, on premises end user software, data center software, cloud infrastructure, cloud software, SaaS software), answer these questions:

• What are the key products we must track (top spenders/most risk)? E.g., Windows vX.X. Do we know of all the assets we should track (shadow IT, etc.)
• How is this software purchased and budgeted? (via partner, marketplace, direct, on a Pcard)
• Who manages the software/asset (license allocations/reclamation, renewal, software upgrades if applicable, retiring the asset, etc.)
• What are the tools and processes used to manage the assets? Are there tooling gaps? Many times a tool can inventory (what is it, and where is it) but not detect usage or analyze the ELP, etc.
  • Tooling should answer:
    • What is it, where is it
    • Is it obsolete, end of support or does it have security flaws? Is it deny listed, vulnerable?
    • Is it being used and by whom, or how many resources is it consuming?
    • How much did it cost?
    • What are the use rights? Can we use it in prod and dev, or can it be use across on-premises and cloud environments, etc.
  • Processes should answer:
    • Who and what tooling is responsible for inventorying the data and understanding security, cost and usage aspects? How often is this done? Is it always on (e.g. via APIs, etc.)
    • Who validates data accuracy and how often?
    • What are the analysis to be run on the data? (KPIs)
    • What is done with the output? Is there an automated process that is run with the data? How do you know you are making progress?

Sample framework for mapping IT asset data and process gaps.

Stakeholder engagement: Is ITAM a team sport?

ITAM’s success depends on strong partnerships across the organization. Key stakeholders include:

• ITSM/CMDB: Ensures accurate configuration data and lifecycle management.
• Security/GRC: Reduces risk by identifying unauthorized or vulnerable software.
• Procurement/vendor management: Optimizes contracts and renewals.
• Finance: Supports budgeting and cost optimization.
• Internal audit: Ensures compliance and transparency.
• Cloud/FinOps: Manages cloud resource allocation and spend for cloud optimization.
• Enterprise architecture: Guides technology standards and rationalization.
• App owners/business units: Provides critical input on usage and requirements.

Mapping these relationships and clarifying roles (e.g., through a RACI matrix) ensures alignment and accountability, and helps ensure there are no gaps in your processes.

The above is a sample RACI you can work through with your organization.

Making ITAM a business enabler

To close gaps and drive outcomes, align ITAM strategy and priorities with business objectives. For example, focus on initiatives that reduce risk, cut costs, or improve compliance, all areas that resonate with executive sponsors. Use real-world examples, such as audit risk reduction, cost savings, or improved renewal outcomes, to illustrate value.

Asking for support

Finally, selling ITAM internally means inviting feedback and support. Ask stakeholders: Do our priorities reflect the organization’s needs? Where can we better integrate or support your objectives? This collaborative approach builds trust and ensures ITAM remains relevant and impactful.

Create your ITAM sales document

Start creating your internal selling document by answering the questions outlined above with a clear ITAM strategy in mind. Here’s a template you can use to start building your pitch. By aligning ITAM with strategic priorities, measuring and communicating impact, and building strong cross-functional partnerships, you can secure the buy-in and investment needed to make ITAM a true business enabler.

Download our template: ITAM mission and roadmap.

Need help? Connect with our team of experts

Want to learn more? Reach out for a chat

Cloud costs are rising fast. Adding to the pain are other complexities—bursty AI workloads, expanding SaaS usage and Scope 2 and 3 sustainability reporting. The explosion of cloud spend combined with complexity means there’s a massive opportunity for services business delivered by partners for cloud cost management and optimization—but only if they have the right solution to reduce their customers’ pain. Flexera One FinOps platform has not only been recognized as a cloud cost management leader by major industry analysts, but it’s also been adopted by global partners like SHI for their FinOps practices.

With this release, Flexera One Cloud Cost Optimization (CCO) adds to these strengths, bringing powerful new capabilities to help partners operate more profitably and efficiently. Whether you’re an MSP managing your customers’ cloud spend, a distributor supporting downstream partners or a system integrator building FinOps services at scale, these enhancements help you drive more value, expand differentiating offerings, and accelerate growth through better visibility, stronger controls and next-generation billing automation.

L1-L2-L3+ multi-level hierarchy control helps partners scale

Partners rarely operate in isolation. You likely are participating in multi-tier ecosystems that include resellers, end customers, and sometimes customers of those customers. The new multi-level hierarchy management capability in Flexera One supports that reality, giving you the granular control needed to manage FinOps across any partner ecosystem.

These capabilities come with guard rails. Users in L1 or L2 will have elevated privileges to their respective child organizations but are prevented from deleting or causing any forms of data loss, relieving concerns about data integrity.

Our next-generation Billing Engine builds trust while improving margins

The new billing engine in CCO makes it far easier to manage cloud resale and provide the accuracy partner customers need. Partners can now seamlessly build and support complex billing plans, apply rules, manage discounts, reprocess bills and control margins with precision.

Billing Rules will enable partners to perform common cloud pricing strategies such as:

• Arbitrage to manage the presentation of commitment and program discounts. Partners can easily manage their program discounts and control their allocation according to their agreements with their customers.
• Cost allocation to better align the costs of consumed assets to their consumers, with more granularity than the cloud vendor may provide.
• Custom pricing or margin enhancement to better align the costs of the cloud usage to the prices partners pass on to their customers. This could include percentage-based margin adjustment, custom credit memos, line items, and other enhancements to the cost data.

Further, Billing History capabilities will give partner customers the ability to lock, unlock or reprocess the bill data as they make configuration changes. This makes partner customers more self-sufficient and helps partners more efficiently manage their customers’ billing.

Sharpen your competitive edge with Flexera FinOps

Out-compete with world-class insights and optimization for your customers across their IT estate. It starts with visibility—get a 360-degree view across all cloud spending by and for your customers—and continues with leading optimization capabilities that help partners:

• Grow revenue with more service offerings
• Improve margins on cloud resale
• Scale operations without scaling headcount
• Strengthen your trusted-advisor role

Flexera One’s fast, seamless enablement of high-value capabilities helps expand your FinOps practice. Partners can now enable Flexera One modules, including CCO, Sustainability and other FinOps-aligned capabilities, for customers in seconds, right within Flexera One, speeding your time to services revenue.

With Flexera One FinOps, partners get visibility across all FinOps Scopes that can support cross-sell and upsell services growth. Flexera One unifies visibility across SaaS, cloud, on-prem, AI services, Kubernetes, and data clouds, giving partners more routes to revenue.

Partners can increase customer satisfaction, retention and margins with the automated savings and high-value optimization capabilities in Flexera One FinOps and Flexera ProsperOps. By delivering savings without additional overhead via automated commitment management, spot optimization, Kubernetes optimization and data cloud optimization, you’ve got the agility and efficiency needed to grow faster.

Learn how to serve more FinOps use cases and grow now, and in the future

Current partners can reach out to their Partner Business manager for a custom demo of what’s new in Flexera One CCO and Flexera FinOps. If you’re interested in becoming a Flexera Partner, let us know!

Reach out for a Partner demo

New Partner interest