After spending considerable time in the trenches of AI engineering — building systems, watching them fail, fixing them, and occasionally watching them genuinely transform how organisations work — I have arrived at a perspective that is neither the breathless optimism of the AI booster nor the cynicism of the AI sceptic. It sits in a more uncomfortable, more honest place.

AI is genuinely, materially useful — but only when treated as an engineering discipline. The organisations that have captured real value from AI are not the ones that moved fastest, spent the most, or deployed the most impressive demos. They are the ones that invested in the unglamorous foundations: clean data, rigorous evaluation, cost observability, and operational discipline. The ROI from AI is not fundamentally different from the ROI of any other software investment — it comes from solving real problems well, not from adopting technology for its own sake.

The cost conversation is long overdue. For too long, AI budgets have been justified with vague appeals to competitive necessity and transformative potential. Both are real — but they are not substitutes for a proper cost model. I am encouraged that the conversation is maturing. More engineering leaders are asking about token efficiency, cost-per-user, and quality-adjusted ROI. More business leaders are asking for concrete baselines and measurable milestones rather than capability demonstrations. This is progress.

The build vs. buy pendulum has swung too far in both directions. Two years ago, every serious AI team wanted to self-host. Today, many teams have overcorrected toward hosted APIs for everything, including genuinely sensitive data that should never leave their infrastructure. The right answer has always been contextual: use hosted APIs for commodity tasks, invest in self-hosting only when the economics and security requirements genuinely demand it, and treat fine-tuning as a precision tool rather than a default approach.

The hidden costs are not a reason to avoid AI — they are a reason to plan honestly. Every technology investment has hidden costs. The ones specific to AI — evaluation infrastructure, prompt maintenance, hallucination management, vendor risk — are learnable and manageable. The teams that have been burned by them failed not because the costs were unknowable, but because they chose not to look. Forewarned is forearmed.

The most important skill in AI engineering right now is knowing when not to use AI. This sounds counterintuitive, but I mean it earnestly. Deterministic logic is cheaper, faster, more reliable, and more explainable than LLMs for most structured tasks. Rule-based systems do not hallucinate. Classical ML models are often superior for prediction tasks with well-defined features. Engineers who reach for LLMs by default, for every problem, are optimising for novelty rather than outcomes. Engineers who use AI precisely — for the tasks where natural language understanding, generation, or reasoning genuinely matter — consistently deliver better ROI.

Looking forward: The cost curve for capable AI is bending downward faster than most people expected. Models that required frontier pricing eighteen months ago are now available at commodity prices. This changes the ROI calculus substantially — tasks that were previously uneconomical to automate are becoming viable. But lower prices also lower the bar for careless deployment. As AI becomes cheaper, the discipline required to deploy it well does not decrease — it becomes more important, because the volume and surface area of AI systems will expand rapidly.

The organisations that will win the next five years of AI are not the ones with the largest AI budgets. They are the ones that have built the internal capability to evaluate, deploy, and improve AI systems as a repeatable competency — not a one-time project. That capability is built through engineering discipline, honest cost accounting, and a refusal to let excitement substitute for evidence.

Written By:
Debarpan Chakraborty | AI Engineer, ThirdEye Data