The easy thing to do with a gap that wide is explain it away. Maybe the machines aren’t quite good enough yet, and the numbers will close as the output improves. Maybe customers are just being sentimental, the way people grumbled about self-checkouts before they started using them, and they’ll come round. I don’t buy either.

What unsettles people is something simpler and harder to fix: the sense that nobody on the other end actually cared — that the thing in front of them rolled off a line rather than came out of a head. The technology is rarely the real problem; the missing human almost always is.

That distinction matters, because it points at a completely different response. If the problem were the technology, the sensible thing would be to wait for it to improve. If the problem is the disappearance of visible human judgement, waiting makes it worse. The answer isn’t to use less AI — that ship has sailed for 97% of the market. It’s to make sure your fingerprints are still on the work.

The gap isn’t about quality

Look closer at what those consumers actually said — Canva ran the survey with The Harris Poll, polling more than 3,500 of them — and the picture sharpens. Eighty-seven per cent believe the best advertising still requires a human touch. Seven in ten say AI-made ads feel like they’re “missing something”. And yet 68% don’t mind AI in the mix at all, as long as the result is genuinely helpful or relevant to them.

What they’re responding to is judgement. Work that had a person’s discernment in it reads differently from work that didn’t, and people feel that difference long before they can name it.

This is the same instinct I wrote about a few weeks ago, when I argued that typos are becoming a trust signal — that in a world where anyone can generate flawless copy in seconds, the visible evidence of a real human becomes the scarce, valuable thing. The trust gap is that idea at scale. Polish used to signal effort. Now polish is free, and effort has to show up somewhere else.

Why hiding the machine backfires

What owners reach for constantly is to use AI everywhere and simply say nothing, letting the customer assume a person made it. The trouble is that customers have stopped assuming. Canva found mentions of “AI slop” have risen ninefold, which tells you people are now actively scanning for the seams. And when more than half of them (52%) say that openly disclosing AI use is one of the things that would build their trust, staying silent about it amounts to a bet against your own audience’s intelligence.

Disclosure is the part owners agonise over, and I don’t think there’s one right answer — but here’s the rule I’d use. Be upfront where trust is part of what you’re selling: professional advice, anything to do with health or money, anything where the customer is taking you at your word. A generated holiday-sale banner needs no note attached. A generated reassurance to a worried client does. Owners who get that call right find that owning up to the AI builds trust instead of denting it.

Get caught passing off generated work as hand-made and you don’t just lose that one piece of credibility. You teach the customer to discount everything else you’ve told them. That’s an expensive lesson to hand someone for the sake of a slightly slicker campaign.

Where the human has to be visible

So the practical question for an owner has moved past whether to use AI, and on to where the human judgement needs to be obvious. Three places earn the effort.

The first is the promise rather than the production. Nobody needs a human to have set the kerning on your email or resized the images, but they do need to believe a person stands behind the claim it makes. So put a name and a face at the point of accountability — the founder’s signature on the sales email, a reply address that reaches an actual human, a real opinion in the newsletter — and let AI handle the production behind it.

Then there’s the judgement itself. Deciding what to say, who to say it to, and what to leave out is the part that rings hollow the moment a model makes the call unsupervised. The AI can hand you a hundred options in a heartbeat, but choosing the right one for this particular customer is exactly the work AI doesn’t replace. It’s the bit you’re actually paid for.

Last comes the response, not just the broadcast. A generated newsletter is forgivable in a way that a generated reply to a real complaint never will be. The closer the contact gets to a specific person with a specific problem, the more unmistakably a human needs to be in the room.

Making the hand obvious

None of this requires a manifesto about your “human-first values” — customers can smell that coming a mile off. It’s quieter than that. It’s the founder recording a thirty-second video instead of commissioning a glossy explainer. It’s the line in the footer that says, plainly, where you use AI and where you don’t. It’s leaving the small human texture in: the aside, the off-script opinion a model would have sanded smooth, because that texture is now the proof that someone was here.

The owners who win here are the ones who are clear about what the machine made and intentional about what stays human. When 97% of your competitors are automating the same surfaces, the visible human stops being a nostalgic indulgence and becomes the one thing that’s genuinely hard to copy.

Your customers aren’t asking you to put the machines down. They’re asking to know there’s still someone there when they reach out — and that, not the AI, is what your marketing has to prove now.

The post The 97/78 trust gap: why AI-made marketing unsettles your customers appeared first on Matt Hopkins.

That deal is ending, and you can see it in the invoices. Which means the way most small businesses buy AI, one model for everything and usually the dearest one, has just turned from a quirk into a budgeting problem, and the businesses that sort it out first will be the ones still smiling at month-end.

On 1 June, GitHub raised the multipliers on its Copilot plans. The frontier models like GPT-5.4 and Gemini 3.1 Pro went from a 1× premium request to 6×. Claude Opus jumped from 7.5× to 27×. Developers running agentic sessions started reporting bills ten to fifty times higher than the month before, for the same work. Around the same time Anthropic brought in weekly rate limits on its Pro and Max plans, with a separate, tighter cap on its most capable model — not because demand fell, but because the compute to serve everyone at the old rate simply wasn’t there. How far the repricing runs from here, I don’t think anyone can honestly say; the direction, though, has stopped being in doubt.

I’ve been making half of this argument for nine months. Back then the point was that most of what your team does with AI runs perfectly well on the cheapest model available, and that defaulting everything to the flagship was a management failure dressed up as a technical default. That was true when the flagship was cheap. What’s changed is the price. The old argument was that you don’t need to. The new one is that you can’t afford to.

The maths isn’t complicated, which is exactly why it’s so easy to ignore until the bill arrives. Think about what you actually run through AI on a normal day: summarising email threads, pulling figures out of invoices, drafting first-pass replies, formatting reports, tidying meeting notes. None of that needs the cleverest model on the market; the cheapest one handles it without breaking a sweat. The work that genuinely earns a frontier model, the nuanced client deliverable or the call that touches money or legal exposure, is a thin slice of the week. Call it eighty-twenty, give or take; you won’t know your own split until you’ve looked.

When the top model cost barely more than the cheap one, mixing the two up cost you a rounding error. Now that it costs six, ten, twenty times more per call, every task you send to the wrong tier is a small overpayment that compounds quietly, until month-end makes it loud.

Your 90-day routing playbook

So what do you actually do about it? You’ve got a quarter, and the work breaks into four moves.

1. Audit what you’re spending (week one). Pull last month’s AI invoice and find the three workloads that ate the most tokens. Most owners have never looked, and the first look is usually a shock: the expensive model turns out to be doing work the cheap one could do in its sleep. This is the same discipline that let a Mark Cuban-backed cheese company, Rebel Cheese, point an AI agent at its shipping invoices and save four hundred thousand dollars in a year — a quarter of a million of it in overcharges the firm only caught after one frantic holiday season. You can’t route what you haven’t measured. If what you find makes you want to rebuild the budget from the ground up rather than just trim it, I’ve set out a framework for exactly that.
2. Set a cheap-first default (weeks two to four). Every task starts on the cheapest model that might handle it, and only moves up a tier when it genuinely needs to. Frontier access becomes the exception you reach for on purpose, not the default you fall into by accident. I run my own work this way, cheap model first and escalate only when the task earns it, and the thing that still surprises me is how rarely it does. If you want the mechanics, how to wire it up and who owns it once it’s running, I’ve written about that separately.
3. Run bake-offs on real work (month two). Take your actual tasks, not benchmarks, and run them side by side, cheap model against expensive, then look at the output honestly. If the cheap model gets it eighty per cent right, you’ve found your routing rule. The other twenty per cent tells you precisely where the frontier model still earns its multiplier — the spend you should stop apologising for. It’s how Cost Plus Drugs ended up having Claude produce a weekly price-comparison report across its twenty-five most expensive drugs: a recurring, well-scoped job that used to need software or a dedicated person, now done in minutes in a browser.
4. Give someone the job of keeping it fresh (ongoing). Models change every few weeks, so today’s bargain is next quarter’s overpriced default. Appoint a model sommelier, someone whose job is to know which model is the right pour for which task and to revisit the routing as the menu changes. It needn’t be a full role; it just needs to be someone’s actual job, because routing without an owner drifts back to one-model-for-everything inside a week.

The wider shift is that the cost of AI has stopped being someone else’s problem. For eighteen months the labs absorbed it; now they’re handing it back, politely, through multipliers, rate limits and repricing nobody bothered to announce. The businesses that come out of the next eighteen months in good shape will be the ones who built routing discipline before the bill forced their hand, not the ones still explaining, come December, why a line item they never watched managed to triple.

The post The end of cheap AI: how the subsidy era ending changes the SMB playbook appeared first on Matt Hopkins.

Then the thing arrives, and it’s a very capable assistant: it drafts, it suggests, it summarises, and then sits there politely until a person tells it what to do next. Useful, often genuinely so, but not quite what you thought you’d bought. That quiet disappointment is sitting underneath a lot of enterprise AI right now.

The word doing the damage is “agentic”, and it’s worth being precise about it. An agent acts: it initiates, it decides, it carries a task through without a human in the loop for every step, and (the part the vendors skate over) it owns the outcome. An assistant responds: you ask, it answers; you point, it moves, and there’s always a hand on the button. The difference isn’t how clever the model is, because the same frontier model can power both. It’s whether the vendor will let the software act without you pressing go, and whether they’ll stand behind what happens when it does. Right now, almost none of them will.

Agentic is a tense, not a capability

The trick, and it’s rarely even a cynical one, is that when a vendor says their product is agentic, they’re usually describing a roadmap in the present tense. The autonomy is real on a slide, in a closed beta, in the demo environment where the data is clean and the awkward edge cases have been removed. What’s sold as “does” is very often “will do”. The verb is carrying the whole sentence, and it’s pointing at next year.

You can see it in the naming if you look. When SAP unveiled its “Autonomous Enterprise” at Sapphire this spring, the headline suite was described as more than fifty Joule Assistants, orchestrating a couple of hundred specialised agents underneath. Read that back. Even the company building the most ambitious version of all this calls the things you’ll actually touch assistants. Salesforce will tell you Agentforce resolves the large majority of customer queries with nobody human involved (85%, by its own count), and for a narrow, well-bounded job like first-line support, that’s plausibly true. Press past the headline and the same shape appears everywhere else: real autonomy in one well-lit lane, and a supervised assistant doing the rest. The reporting that’s followed (early adopters wrestling with data quality, inconsistent behaviour, and pricing nobody could explain) is what that gap feels like once the contract’s signed.

The technology is real. The marketing has simply run a year or two ahead of the engineering, which is what happens in every adoption cycle. How long the gap lasts, I’m not sure; maybe the engineering closes on the marketing faster than these cycles usually allow. Either way, it only becomes your problem if you buy on the marketing’s timeline and then plan your business on it too.

Why the difference lands on you

If all this amounted to was a slightly grander word than warranted, it wouldn’t be worth writing about. But the agent-versus-assistant line is also the accountability line, and that’s where it gets expensive.

Deploy a genuine agent and you’ve got to answer some uncomfortable questions before you switch it on: which decisions it’s allowed to make, what it does when it’s unsure, who owns the result when it’s wrong. Deploy an assistant dressed as an agent and those questions get skipped, because everyone assumes a human is still in the chair, right up until the day the thing does something consequential and you find that nobody ever decided who was responsible. You’ve taken on all the risk of letting software act on its own, while still doing all the supervising yourself. The worst of both, paid for at the price of the better one.

This is the supply-side mirror of something I’ve written about before from the buyer’s end: most organisations want agents but can’t yet support them. The two halves feed each other. Buyers who aren’t ready can’t easily tell roadmap from running product, so they reward the grander slide; vendors, sensing that, sell the grander slide. And both bets come due at the same place: the moment something goes wrong and nobody can say who owned it. For a smaller business that sting lands harder, because you don’t have the spare process, the legal cover, or the slack in the budget that lets a large enterprise shrug it off.

How to tell, in the room

There’s no need to become a sceptic about AI to protect yourself here. You just have to drag the verb back into the present tense. A handful of questions does most of the work, and you can ask all of them in a single meeting:

1. Show me the autonomous part running. Not a video, not a sandbox, but the actual product, doing the actual thing, on data that looks like mine. If the demo always has a human clicking “approve” between steps, you’re buying a faster co-pilot, not an agent.
2. What does it do when it isn’t sure? A real agent has an answer to this: it escalates, or it pauses and flags the gap for a person. A relabelled assistant just hands the whole problem back to you and calls it a feature.
3. What’s generally available today, and what’s roadmap? Make them sort the slide into things you can use this quarter and things you’re being shown to close the deal. The honest vendors will do it without flinching.
4. Who’s accountable when it gets something wrong, and how do I undo it? Watch what happens to the word “autonomous” the moment liability and rollback enter the conversation. It tends to shrink.
5. What am I actually committing to? Pricing that assumes autonomy you won’t have for a year, integration work that quietly lands on your team, lock-in that’s hard to reverse. The cost of the roadmap is usually paid on today’s invoice.

If you want a sharper sense of what you’re being sold versus what you’re being shown, it’s worth knowing where the agentic maturity scale actually tops out before you walk in. Most products selling level five are shipping level one or two, and that’s fine, as long as you know which one is landing in your account.

Buy the assistant, just buy it on purpose

An assistant isn’t a con, and that’s the part worth holding onto. For most small and mid-sized businesses, a good assistant (one that drafts the proposal, reconciles the invoices, answers the first-line query) is exactly the right tool for where you actually are. The value is real and it’s available today, not next year.

The mistake isn’t buying one. It’s paying agent prices for it, and building agent-shaped plans around it — the headcount you were going to redeploy, the process you were going to leave unsupervised — for something that still needs you in the chair. That’s how you end up disappointed by software that was, on its own terms, working perfectly well.

So when the next slide says agentic, treat the word as a question rather than a claim, and make the vendor earn it before you pay for it.

The post Sold the agent, shipped the assistant appeared first on Matt Hopkins.

Most of us are treating this as weather. Bad weather, but weather: something that blew in and will eventually blow out, probably after the next Budget, probably with a bit of relief tucked into it for small employers. I understand the instinct. It’s also the most expensive mistake you can make this year, and the smarter move is to plan around it now.

The floor is legislated to 2031

The hope inside “let’s just get through this” is that the number is temporary: confidence recovers, the Chancellor finds some room, and the cost of a job drifts back toward where it sat in 2024. Nothing in the data supports that hope. The secondary threshold for employer National Insurance, the part that did most of the damage, is frozen until 2031. That isn’t a squeeze you wait out. It’s a new floor, legislated to stay put for the rest of the decade.

The wider picture offers no rescue either. The Bank of England’s agents report subdued growth through 2026 and pay settlements easing toward 3.6 per cent as the labour market loosens; ICAEW’s confidence monitor has sat in negative territory all year, with tax named as a worry at record levels. For an owner, that’s the worst of both worlds. You can’t lean on a buoyant market to pass the cost to customers, and you can’t bank on a generous Budget to take it off you. Which leaves the one lever that’s genuinely yours: the shape of your own operation.

So the realistic planning assumption is plain: this is the price now, not a number waiting to fall back. And once you accept that, the question stops being how do I survive it and turns into something far more useful: how do I rebuild the operation around it.

What £25,850 actually buys

Per head, the rise is worth roughly £2,872 a year. It helps to think about what that sum is, rather than letting it sit as an abstract line on the P&L.

For a lot of small firms, £2,872 a head is the entire training budget. It’s the margin on a decent-sized job. It’s the difference between a year where you take something home and a year where you don’t. In plenty of cases it’s the gap between hiring the tenth person and deciding you can’t. Underneath the headline figure, the cost has re-priced every individual decision that sits on top of a salary: every hire, every promotion, every pound of training now starts from a higher base.

That’s the bit owners miss when they file the number as a single tax event. It permanently changes the economics of every future decision that involves a person, and it does so for as long as the team exists.

Treat it as fixed, then design around it

The founders who come out of this in good shape will be the ones who do something slightly counter-intuitive: they stop arguing with the number. They book it as fixed, like rent, or the lease on the premises, and then they redesign the business so the higher cost of a head is matched by more value per head.

That can sound like a slogan, but it’s really a design problem, and it breaks into three concrete moves.

1. Price the cost back in. If a job costs more to deliver, some of that has to move into the price. Owners are oddly squeamish about this — far readier to swallow a £25,850 rise than to put four per cent on an invoice. I’ve made the case before in pricing, value and the incentive to buy: the firms holding their margin through this are the ones who’ve stopped apologising for what they charge.
2. Lift output before you add heads. If you can’t easily take on people at the higher cost, get more from the ones you’ve got — the question I worked through in scaling without hiring. The tools for raising output per person are cheaper and better than they’ve ever been, at the exact moment that adding a person got dearer. That timing isn’t a coincidence to waste.
3. Make every role earn its new cost. At £2,872 more a head, some roles that were comfortable at the old price are underwater at the new one, and not because anyone’s underperforming — the maths simply moved. Working that out is uncomfortable. Skip it, and the £25,850 becomes the thing that actually breaks you.

The counter-case: a Budget reprieve

There’s a counter-case worth taking seriously. A version of the next Budget could throw small employers a bone — a bigger Employment Allowance, a touch of relief on the threshold — and if it lands, the squeeze eases. I honestly don’t know how likely that’ll be. The allowance already rose to £10,500 for this year, which genuinely helps the smallest firms. But “it might get marginally better” is a thin thing to build a year on, and even the hopeful version doesn’t take you back to 2024.

Planning for the floor and being pleasantly surprised is a far safer way to run a business than planning for the rescue and getting caught out. So treat the £25,850 as what it now is — the standing cost of having a team — and the owners who price for it and build an operation plainly worth the money will still be standing when the ones holding out for a Budget miracle have quietly run out of room.

The post The £25,850 line: the new cost of a job, and how to plan for it appeared first on Matt Hopkins.

It’s the most convincing version of that pitch I’ve seen, and Anthropic has stripped out every obvious reason not to try it. You toggle it on inside Claude Cowork, connect the tools you already run on (QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, Microsoft 365), and pick a job: planning payroll, closing the month, chasing invoices, prepping for tax season. There are fifteen ready-made workflows and fifteen skills, and Claude does the work while you approve before anything sends, posts, or pays. Daniela Amodei, Anthropic’s president, pitched it modestly enough — “People run the business,” she said, “and Claude helps take the late-night work off their plates.” It doesn’t even cost extra; it’s a free toggle on whatever Claude plan you already pay for. No IT department, no code, no integration work, no new line on the budget.

Which leaves exactly one thing Anthropic can’t hand you, and it happens to be the thing that was always the real problem. The reason most owners can’t delegate has never been the absence of capable help, or the price of it. It’s that they’ve never defined the work clearly enough to hand it over.

The work nobody wrote down

Michael Gerber named this forty years ago in The E-Myth. Most small businesses are run by people who were good at the work (the baker, the plumber, the designer) and who assumed that being good at the work meant being good at running a business that does the work. It doesn’t. The owner who can’t get out from under is rarely short of effort, and rarely short of people. They’re short of definition. The job lives in their head, gets done a little differently every time, and so can’t be passed to anyone without them hovering over it. Gerber’s famous line, work on the business and not in it, was really an instruction to do that defining work: write the thing down, turn the craft into a procedure, so it can be handed over. I’ve written before about what that instruction means now that the systems you delegate to can think for themselves, and the short version is that it gets harder, not easier.

Take “chase overdue invoices,” one of the jobs in the box. It sounds like a single task. It’s really a dozen decisions you’ve never written down. Which customers do you chase, and which do you leave alone because they always pay late but always pay? After how many days? In what tone for a brand-new client versus the one you’ve had for nine years? At what point does a reminder become a phone call from you personally? An owner carries all of that around in their head and applies it differently every Tuesday. Hand the job over with none of it written down, and Claude will still chase your invoices — quickly, confidently, and by whatever rule it inferred rather than the one you’d have used.

An employee you can’t brief

That’s the bit the “fifteen silent employees” framing skips over. However capable, a new hire you can’t brief is a liability. You give them a fuzzy instruction, they do something plausible, and you spend more time correcting it than you’d have spent doing it yourself. Whether this tool helps you comes down to one thing: can you say precisely what you want? Most owners have never had to, because the work has always sat unspoken — done by hand, by habit, by instinct.

So the useful question runs the other way. Before you ask what Claude can do for the business, ask how clearly you could tell it what the business needs. That second question is the one that predicts the result, and no review of the software can answer it for you.

Seventeen of twenty

There’s a neat illustration of where the line sits. The New Stack built a fake profit-and-loss statement for a small consultancy, buried twenty deliberate problems in it, and set Claude for Small Business loose. It caught seventeen in under six minutes — every obvious and middling issue, and most of the hard ones. The three it missed were the most forensic, the kind you only notice if you already know what normal looks like for that particular business: a suspiciously flat interest figure, a small drop in bank fees, depreciation maths that was a shade too clean.

That result is more flattering to the human than it first looks. Claude caught everything that could be found by reading the numbers. It missed the things that could only be caught by someone who knew the business behind them. The tool brings the analysis; you bring the knowledge of your own operation. The software doesn’t move that line. It just shows you, rather precisely, where it runs.

Who should switch it on

If you’ve already done the defining work — if you could write your invoice-chasing rules, or your month-end checklist, on a single page without much thinking — then this is the most leverage you’ll have bought in years, and you should turn it on this week. If you couldn’t, the tool will still run, and that’s the trap. Plenty of owners will switch it on, run the month-end workflow, get a tidy-looking P&L, and never notice they’ve outsourced not just the arithmetic but the judgement that was the only reason they understood their own business in the first place. We’ve seen the smaller version of this play out already: a team switches on an AI feature, uses it for a fortnight, then quietly drifts back to the old way, because nobody ever decided what it was for.

There’s a more hopeful version, and I’m genuinely not sure which one wins. It’s possible the tool forces the discipline owners have dodged for forty years — that you can’t ask Claude to chase your invoices without finally deciding your rules for chasing invoices, so the clarity gets dragged out of you whether you fancied the work or not. Used that way, it earns its keep. Used as a way to avoid the thinking, it just adds a fast, plausible layer of output that nobody really owns. Anthropic doesn’t get to decide which you end up with. You do.

So the question Anthropic wants you to ask — what you’d do with fifteen tireless employees — is the right one. It’s just not the first one. The first is quieter and a good deal harder: could you write down a single job in your business clearly enough that someone could do it well without you in the room? If you can, switch it on. If you can’t, that isn’t a software problem, and no toggle is going to fix it.

The post Claude for Small Business removes every barrier but the one that matters appeared first on Matt Hopkins.

So what follows is a way of seeing. There’s a glossary at the bottom if you want one, but the value is in the picture itself. Once you’ve got it, you can’t unsee it, and plenty of decisions that felt like guesswork start to feel like choices.

The setting you haven’t touched

Start with the one you already have in front of you. Open Claude Cowork, or ChatGPT, or almost any AI product, and somewhere there’s a small selector offering you a choice of model — Opus or Sonnet, say. Most people leave it on whatever it arrived set to, because the names mean nothing and the whole thing looks like something for the technical folk to worry about. It isn’t. Opus is the powerful, slower, more expensive option; Sonnet is the lighter, faster, cheaper one. Same product, same buttons, same screen, same way of working; different model doing the actual thinking. For a hard piece of reasoning, reach for the powerful one. For drafting an email, summarising a document, the everyday back-and-forth, drop to the lighter one and your subscription stretches further. Nothing else changes. You’ve simply stopped paying for power you weren’t using.

That small toggle is the whole idea in miniature. The model — the part that does the reasoning — is a separate thing from the product wrapped around it, and you can change one without changing the other. I made a version of this point eighteen months ago from the other side, arguing that most of what you’re paying for lives in the product, not the model. This is the version you can act on without writing a line of code.

What you’re actually paying for

Two things are doing the work whenever you open up an AI tool. There’s the product — Claude Cowork, the ChatGPT app, the Codex tool, Cursor — and there’s the model running inside it: Opus 4.8, GPT-5.5, Gemini 3.1 Pro. The bill bundles them. The marketing blurs them. But they’re separate decisions made by separate parts of the same machine, and when the product is the kind that can run tasks and edit files rather than just answer questions, the trade calls that an agentic harness — a useful name because the harness is the layer that decides whether you can swap the model underneath.

None of the big sellers name these layers consistently. The cleanest example is Google: Gemini is the company’s AI brand, the product you open, and the model line, so when someone tells you they “use Gemini” you’ve learned almost nothing about what they’re actually running. OpenAI is barely tidier — “ChatGPT” is the product most people mean, and Codex is now both a product and a family of models. Anthropic happens to keep them separate, but that’s a stylistic choice on their part, not a market norm. The cost of letting the marketing pick your vocabulary is that you can’t tell which layer you’re being sold.

When the bundle comes apart

Where the distinction earns its keep is the moment your tidy bundle starts to come apart, and how far it comes apart depends on what you’ve bought. Claude Cowork lets you change the model within Anthropic’s own range, Opus for Sonnet. A tool like Cursor lets you pick across companies. And something like OpenClaw, or Claude Code itself, will run almost anything you point it at, including open-weight models such as Moonshot’s Kimi K2.6 that you can host on your own machine. That last part is the proof the separation is real: people genuinely run Claude Code on a non-Anthropic model, and it still works. If the two were one welded unit, you couldn’t. The harness is the thing you work in: the loop, the tools, the memory, the permissions. The model is the reasoning you plug into it. The bundle just decides how far you’re allowed to reach for a different one.

There’s even a whole class of tool whose entire reason for existing is to hand you that reach — OpenClaw, community-built and now moving to its own independent foundation, is the obvious one. It’s a harness that will run whatever model you give it: bring your own. With Anthropic, the company building the product also builds the model and would, understandably, prefer you used both. With a bring-your-own-model tool, the product stays the product and you pick the reasoning yourself, or you let it be no one’s, in the case of open models that belong to nobody. Whether that flexibility is worth the extra setup will depend on you; for most small teams, the truthful answer is they don’t switch often enough to feel locked in either way. For a business thinking past this quarter, though, that’s the un-locked-in path: you’re not betting your whole workflow on a single vendor’s roadmap and a single vendor’s pricing.

Everything I’ve described is shifting under your feet. Codex used to be a model and is now also a product. OpenClaw was Clawdbot, then Moltbot, then OpenClaw, all in the space of a few months. The vocabulary keeps sliding because the industry hasn’t settled on which layer it’s actually selling, so the names chase wherever the money is rather than describing what’s underneath. You’re being asked to make purchasing decisions in a fog the sellers are busy generating.

The integration question

There’s one more layer worth knowing about, and it’s a buying question rather than a what’s-inside-the-machine question. Can your AI actually touch your calendar, pull a document, post into Slack, edit a row in your accounting software? Or can it only describe those things from outside? The current standard for letting it do the touching is called MCP. You can cheerfully forget what the letters stand for; the only question that matters is whether your AI can connect to the other tools you already pay for. Without it, an assistant is clever but sealed off, able to talk about your work and never to touch it. With it, it can act on the real thing. So “do you support MCP?” is worth asking out loud when you’re weighing up a new bit of kit, because the answer tells you whether your AI will be able to work with it or whether you’ve just bought an island. Whether MCP itself wins as the standard or another contender takes its place isn’t yet settled; what’s settled is that the connect-to-the-rest-of-your-stack question is the right one to be asking. I’ve written before about how whole categories of software are now being rebuilt around AI agents as their primary customer rather than humans; whatever the protocol ends up being called, it’s the connector that decides whether the agent you’ve paid for can join in.

None of this asks you to write a line of code or read a single specification. It asks for one shift in how you look at the thing you’re paying for: stop treating “an AI” as a single, indivisible object, and start seeing the model, the product around it, and its reach into your other tools as three separate choices that are yours to make. Do that, and the intimidating little toggle becomes a lever for your own budget, and the glossy product page becomes a short list of questions you now know how to ask. Learning to name the parts is how you stay the one making the choices.

A short glossary

Company: the organisation that builds and sells the technology. Anthropic, OpenAI, Google, Moonshot.

Brand: the consumer-facing name the company markets under. Claude, ChatGPT, Gemini. Often, confusingly, doubling as something else on this list.

Product: the actual thing you open and use. Claude Cowork, the ChatGPT app, Cursor, the Codex tool.

Agentic harness: a product that lets the AI take real actions (run tasks, use tools, edit files) rather than only answering questions.

Model: the part that does the reasoning. Opus 4.8, GPT-5.5, Gemini 3.1 Pro, Kimi K2.6. Separate from the product wrapped around it, and often swappable.

MCP: the standard that lets an AI connect to your other tools and data. The thing that decides whether your assistant can act on your world or only describe it.

The post The part you’re not choosing appeared first on Matt Hopkins.

This is the founder bottleneck. It isn’t laziness or arrogance or a missing skill. It’s the slow drift from being someone who can do anything to being someone who has to do everything. The two are very different positions. The first builds a business; the second prevents it from growing.

The pattern that works at five and breaks at twenty

At five people, being in every decision is fine; it’s why the work is good. The team is too small for delegation to make sense, the cost of a bad call is high, and you’re closer to the customer than anyone else. Founder-as-decision-maker is the right model for that stage.

At twenty, the same pattern is the constraint. The discount approval queue stretches. The supplier swap waits for your call. A new hire’s first-week question goes to you because nobody else has been allowed to answer one. Growth at this stage is capped by managerial capacity, not market demand. The thing that grew the business now slows it down, and the team — the same team you hired to take work off you — learns to wait for you instead of move without you. Ask the Banker put a name to this in their 2026 report — the small-business execution gap — and it’s the same pattern EINEdge has been writing up under the heading of founder discipline: owners carrying the day-to-day judgement load past the point at which their week can hold it.

The founder model fades rather than fails. There’s no day it breaks, and so for a long time you tell yourself you’re being thorough.

The week audit

The fix starts with looking at what’s actually on your week. Most owners have never done this — we plan in lists and we measure in feelings. A week’s worth of calendar, sorted into four categories, tells the truth:

1. Decisions only you can make. The contract a director signs. The hire at executive level. The bet-the-business call. Maybe five per cent of the week, often less.
2. Decisions you happen to be making. Pricing on a job. A supplier swap. A discount approval. These could be made by someone with the right context, but you’re the one who has it.
3. Work you find yourself doing. Drafting the proposal. Approving the social post. Fixing the spreadsheet. There’s no judgement here that requires you; you do it because you’ve always done it.
4. Work you’re hiding behind. The inbox. The status calls. The “quick look” at the design before it goes out. We tell ourselves these are oversight. Most of the time they’re anxiety dressed as activity.

Categories three and four are the bottleneck. They feel productive, which is why they survive. They’re also the reason the team can’t move without you.

The discipline of letting go

Dan Amos has run Aflac for 36 years and gave HBR an interview last month that contained a line worth keeping. Asked what he had learned about leadership across that time, he said: listen for the silence. When a meeting goes quiet, it isn’t because everyone agrees. It’s because people don’t want to disagree with the person in charge. The longer you’re in the room, the more silence you create.

This is the founder bottleneck phrased in a different language. Being needed narrows the conversations around you. People stop bringing the hard call to your desk because they’ve learned the answer is “leave it with me.” And the business slowly reshapes itself around your inbox.

Letting go doesn’t mean stepping back. It means choosing which decisions still need you, deliberately, and then trusting the team to live with the others. That trust has a cost — they’ll make calls you would have made differently. Some will be worse. A few will be better. The first time a team member makes a decision that costs you a customer, the temptation to take the work back is almost physical. What you’re aiming for is movement; the business needs decisions made, not decisions held back for the one person who isn’t free to make them. Where the line falls between trust and abdication is rarely obvious — I’m still not sure I always read it right, and the founders I know best say the same.

This is the place where “working on the business” becomes a discipline rather than a slogan — a weekly choice about what work you accept as yours, made fresh every Monday.

What this actually looks like

You audit the week. You name the bottleneck out loud, first to yourself and then to the team. You move category-three work out, with a proper handover and an agreement on what “done” looks like. You move category-four work out too, and you accept the discomfort of not being copied in. You set a rule for category two: if someone has the context, the decision is theirs, with a written threshold so that “I wasn’t sure if I should ask” stops being a recurring conversation. And you protect category one with both hands, because that’s the work only you can do.

There’s no project plan in there. You sit down on a Monday with the week ahead and ask, of every meeting and approval, which of these genuinely need me — and which of these are me, being needed.

The founders who broke through

The owners I’ve watched break past the twenty-person ceiling didn’t get better at managing. They got worse at being indispensable. They made themselves a little harder to reach, a little less central to the day-to-day, and the business — given the space — grew into it. Being needed felt like the job for a while; somewhere around twenty people, it became the cap on the work they actually wanted to do, and they had the discipline to recognise the swap.

The post The founder bottleneck: why being needed is killing your growth appeared first on Matt Hopkins.

Six weeks ago, the strategy turned. The Technology Secretary, Liz Kendall, announced a £500 million Sovereign AI Unit — equity of up to £20 million per startup, a million GPU-hours each on the AI Research Resource supercomputer network, fast-tracked visas in one working day, and government help with procurement, data access and regulatory navigation. A £282 million Strategic Assets Grants Programme sits alongside it for shared infrastructure like high-value datasets and automated labs.

The first equity cheque went to Callosum, an AI infrastructure company. Six more got compute access: Prima Mente (biological foundation models), Cosine (world simulation), Cursive (agentic AI), Doubleword (sovereign inference), Twig Bio (engineering biology) and Odyssey (AI for national security).

This is the most concrete UK industrial-policy AI move in years, and on balance the right kind of move. It still deserves a careful grading, though, because the real risk here is that the fund succeeds at the wrong thing.

What the fund gets right

It prices speed over paperwork. Visas in a working day. Procurement fast-tracks. Help cutting through the regulatory pinball machine. This is the bit that matters most, and it’s the bit Whitehall has historically been worst at. The biggest reason the UK has lost AI companies to the US, in the end, has surprisingly little to do with tax or capital. It’s that decisions in Whitehall take six months and decisions in San Francisco take a week. If you’ve ever waited for a visa to clear while a competitor in Palo Alto hired the same person, you understand viscerally how much this matters. Speed is a competitive moat, and the fund finally treats it as one.

Compute, too, lands in the right place. A million GPU-hours per company on AIRR is the difference between training a credible foundation model and renting fragments of one. Britain has spent two decades excellent at AI research and second-rate at productising it. The gap has almost always been compute. Throwing a public supercomputer at the problem is precisely the kind of state investment that markets won’t make and that makes everything downstream possible.

And it backs serious technical bets, not press releases. Callosum, Prima Mente, Cosine, Cursive, Doubleword, Twig Bio, Odyssey — the list is heavy on hard problems. Biology, world models, inference infrastructure, defence. These aren’t chatbot wrappers or consultancy plays. Someone in the Sovereign AI Unit has been reading carefully and picking on technical merit rather than relationships. That alone is unusual enough to be worth noting.

What it still gets wrong

It doesn’t fix why British AI is worth more in California. That, in the end, is the heart of the problem, and the announcement doesn’t address it. Every one of those seven companies is, today, more valuable the day after it flips to a Delaware C-corp than the day before. The buyer pool is wider, the multiples are higher, the path to listing is shorter, and the option pool everyone competes for talent with is denominated in dollars. The British government can write a £20 million cheque and still be outbid two years later by a US acquirer paying £600 million for the same business. The pattern is well-rehearsed: DeepMind sold to Google. ARM relisted in New York rather than London. Mistral, on smaller GDP behind it, has done something Britain hasn’t quite managed yet. The same UK business, listed or owned across the Atlantic, simply commands a higher price.

France’s playbook of equity, compute and visas is the one the Sovereign AI Unit is copying. But France pairs it with €109 billion of AI infrastructure announced at the AI Action Summit, a deeper sovereign-cloud strategy through Mistral Compute, and at least the beginnings of a regulatory environment designed to make domiciling in Paris pay off. Britain hasn’t done the harder thing: making the UK a more valuable home for a scaling AI company than the US is. Until that changes, this fund is a down-payment on companies the UK won’t keep.

Scale is the other gap. £500 million is genuine money for seven startups. It’s also less than 1% of what France committed in February 2025, and a rounding error next to the $100 billion MGX fund the UAE is now deploying. If this is meant as a focused bet on a handful of frontier companies, fine. If it’s supposed to underwrite Britain’s place in the global AI economy, then this is the opening bid, not the strategy.

And the fund funds builders, not buyers. I made a similar point about the Budget six months ago, and it applies here too. A country with strong frontier firms and a long tail of poorly-managed companies that won’t or can’t adopt frontier technology doesn’t get richer from AI just because it has Callosum and Prima Mente. It gets richer when the average British firm uses what those companies build. I’m not sure how cleanly anyone can separate the two halves of the productivity problem, but the half the fund doesn’t touch — whether the rest of the British economy actually adopts what gets built — is a big chunk of it. The instruments for that half look completely different: SME procurement reform, sector-specific implementation grants, tax relief for AI adoption rather than only for AI creation, and a serious push on management capability uplift. None of those are on the table.

How it scores

It’s the right direction with the right instruments, at roughly a tenth of the right scale, and with the central trapdoor still wide open. The pivot from “we need a registration regime” in February to “here are seven equity cheques” in April is exactly the shift the case for action required, and anyone who argued the government should stop legislating and start unblocking — and I was one of them — has very little ground to be churlish about it now. But two months ago Britain was writing compliance bills and this month it’s writing cheques. Neither, alone, is the answer to why so many of the businesses the UK wants to keep are worth more the moment they stop being British. That answer sits in the tax, listing and capital-gains environment Whitehall keeps avoiding.

The post The UK’s £500M bet: what the Sovereign AI Fund gets right — and still gets wrong appeared first on Matt Hopkins.

The number itself is engineered for a podcast clip. Ninety-three is precise enough to sound rigorous and vague enough that nobody can cleanly disprove it. The substance underneath is genuine, though. Most enterprise marketing budgets are pointed in the wrong direction, most CMOs know it, and AI has now made the mid-funnel work they ought to be doing instead much cheaper to produce. The question worth asking is why the people running these departments, most of whom are perfectly capable, can’t move the money even when they see the gap. The answer the original argument doesn’t really address is that the wasted spend has a second job that the dashboard requires.

The dashboard’s real job

Impressions, reach, MQLs, top-of-funnel pipeline contribution, share of voice. These are activity numbers. They tell you what the function did each quarter — how much content went out, how many ad slots were filled, how many leads were qualified. They give the CMO a number to put on a slide each quarter, the CFO a column to compare with last year, and the board a way to evaluate a function that otherwise lacks an obvious evaluation criterion. Most of these numbers could be deleted tomorrow without anyone’s revenue moving by a pound. But if you delete them, the function loses its budget, because the people setting the budget don’t know what else they’d judge it on.

This is the part that doesn’t show up in marketing critiques very often. The reason the wasted spend exists is structural. Buying impressions produces the report. Cut the impression-buying and the report empties out. Cut the report and the budget gets a haircut at the next planning round. Most CMOs understand this perfectly well; it’s their job to.

Mid-funnel work, and why it’s hard to fund

Mid-funnel work is hard for exactly this reason. It doesn’t return a number in thirty days. Brand-building, organic social, podcasts, content that earns trust over six months — these activities produce results you can only see as compounding effects on win rates, deal size, and inbound velocity. By the time the effect shows up in a quarterly report, you can’t cleanly attribute it to any single piece of work. There’s no dashboard tile that says “this episode of our podcast contributed £40,000 of pipeline.” There can’t be. So when budget time comes around, the CMO who reallocated 20% to organic mid-funnel content has nothing to point at, while the colleague who kept buying programmatic impressions has a chart that goes up and to the right.

The 20% reallocation Vee is recommending is, technically, sound. AI tools have collapsed the cost of producing organic social. A small in-house team can now ship in a week what a creative agency used to bill for over a quarter, and the unit economics have flipped exactly as he says they’ve flipped. Producing it has never been cheaper. None of that matters in the boardroom, because the boardroom isn’t reading unit economics. It’s reading the same dashboard it read in 2019, and that dashboard has no row for “trust accumulated.”

There’s an argument I’ve made before that AI didn’t break marketing so much as expose which content was ever worth finding — the rules of discovery changed, and the content built to game the old ones became invisible. Vee’s 93% is downstream of the same shift. The economic case for the mid-funnel is overwhelming. What hasn’t shifted is the internal political case, and the political case is what actually funds it.

The vinyl signal

There’s a counter-signal worth holding next to all of this. Vinyl records hit $1 billion in US sales last year — the first time since 1983, growing 9.3% year on year. Independent bookshops are growing again. In-person conferences are oversubscribed and prices are rising. The marketing channels that have worked most reliably for premium brands over the past few years share a feature: they don’t appear on any funnel diagram, and they produce no impressions data at all.

What’s emerging is barbell-shaped. At one end sits the slow, expensive, hard-to-measure analog work: events, physical media, direct mail to a hundred named people, the things that look like 2002. At the other end sits the cheap AI-produced organic content that Vee is talking about: TikToks and shorts and podcast clips, the things that look like 2026. The middle of the bar — impression-buying, programmatic, MQL factories — is quietly hollowing out. That’s the shape of where attention has gone.

What the two ends share is a problem: neither produces a number you can put on a quarterly slide. A handwritten note to ten key prospects doesn’t make a dashboard. Neither does a TikTok that earns three million organic views over six months and changes how a category talks about itself. A marketing director who funds either is doing real work. A marketing director who keeps buying impressions is doing dashboard work. From the boardroom seat, only the dashboard work is visible. So that’s what gets resourced, every quarter, in every public company you’ve ever worked with.

The measurement question that comes first

Vee’s prescription is a budget reallocation, but the harder, prior question is about measurement. The CFO has to accept results that take six months to show up and can’t be cleanly attributed to a single campaign. Founder-led companies do this routinely, because the founder has the standing to say “we’re going to stop measuring the bit that’s easy and start measuring the bit that matters.” Public-company CMOs reporting to a CFO and a board generally can’t make that move alone. Their budget conversations are constrained by what their counterparts can see on a slide, and the slide hasn’t changed in twenty years.

Marketing is full of perfectly capable people producing reports nobody believes, because the alternative is a report nobody can read. I’ve written before about how the industry’s self-indulgent culture lets a lot of this go unchallenged — the awards, the jargon, the trend-chasing — but the part that survives every one of those critiques is the dashboard. The dashboard is the mechanism by which the function keeps its budget. Critics can attack the awards or the jargon and nothing happens. Attack the dashboard and the function’s whole budget conversation collapses, which is exactly why nobody does.

If you’re going into a planning round wanting to do what Vee suggests, there are three tests that decide whether the reallocation will actually hold:

1. What evidence will you accept? Before you move any money toward mid-funnel work, get clear on what trust-building evidence — pipeline velocity over a year, deal-size shifts, inbound-source mix — you’re willing to track and defend. If you can’t name it now, you won’t be able to defend it in six months when the impressions chart looks sparse.
2. Who in the boardroom will hold the line? The CFO and the board are reading a specific dashboard. You need someone senior — ideally the CEO, often the founder — who will keep saying “we’re running on the longer measure now” when the easy numbers start looking thinner.
3. What’s coming off the dashboard, and when do you tell them? You can’t add a measurement without retiring one. Decide which existing metric you’re letting go of and brief the board before the next quarter’s report shows up looking different.

Skip any of those three and Vee’s 20% reverses itself within two planning cycles, and the budget returns to the same line items it left.

Vee’s right that the marketing budget is misallocated. But the 93% is the consequence of an earlier change that hasn’t happened yet — about what the board is willing to call evidence — and I’m not sure whether that change has to come from the CFO or further up. Either way, until somebody at the top accepts evidence that’s slower, fuzzier, and harder to attribute, the dashboard wins. And the dashboard knows what it likes.

The post The 93% isn’t waste — it’s funding the dashboard appeared first on Matt Hopkins.

In the last fortnight, four vendors have stopped pretending.

Salesforce launched Headless 360, with co-founder Parker Harris asking the question out loud: why should you ever log into Salesforce again? Microsoft rolled out Agent Mode across Office. OpenAI shipped Workspace Agents. Google used Cloud Next 2026 to fold Vertex and Agentspace into Gemini Enterprise — pitched as the layer that runs your workflows so the humans don’t have to. Aaron Levie at Box has been saying for months that headless is inevitable; this is the fortnight the incumbents agreed with him.

That’s a lot of product news for a fortnight, and the substance underneath each launch is the same concession. The companies whose entire revenue model depends on humans paying to log in have started building the product that means humans don’t log in. That changes what you’re buying. And the contract sitting in your shared drive — the one signed at the last renewal — is the wrong document for the conversation that’s coming.

Why per-seat stops working

Per-seat pricing only makes sense if seats describe usage. The moment one human can direct twenty agents that each hit your CRM, your ERP and your support tool on her behalf, the seat stops describing anything. You can keep the unit if you like, and the vendor will too, for a while, because it sells the same number of dashboards as last year. You’ll be paying for nothing.

Per-seat also created a perverse incentive that has held SaaS in place for years. Vendors got rewarded for the number of people who logged in, not the work being done — which is why most products optimised for daily active users, why dashboards proliferated, why every workflow got its own button to click. The interface was the product because the licence was for the interface.

A headless world inverts that. The interface becomes decoration; the value sits in the API surface, the data the system holds, the actions it can execute on a user’s behalf. I covered the broader shift earlier this year in when the application layer is eaten from below, where agents route around software rather than replace it. Intercom made the same bet from the vendor side a few weeks ago, with its CLI signing up for an account without a human ever touching the form (more on that in agent-ready SaaS). What’s new this fortnight is that the incumbents have joined them, in public, with launches their own sales teams now have to defend.

Why the procurement playbook breaks

Most enterprise SaaS contracts are written around concepts that have just become inoperative. They count seats. They specify maximum concurrent users. They cap API calls per user. They penalise the customer for sharing logins. They assume the boundary of usage is a human looking at a screen, and they price the relationship accordingly. Your procurement team didn’t write these contracts; they inherited them. That’s worth saying out loud, because the next renewal is going to need to fight on ground nobody chose.

Picture a renewal meeting that’s coming for a lot of buyers this autumn. Finance and IT are both in the room, and the vendor’s account exec is on the call. The vendor’s audit shows seat utilisation up 40% year on year. Finance is delighted. IT looks suspicious — the utilisation rise is mostly automated API calls from an agent that ten people on the team have started using. Is that one seat or ten? Is it covered by the existing API entitlement, or is it overage? Who’s liable if the agent does something it shouldn’t with the data? The standard contract has nothing useful to say on any of those questions, and the vendor is about to suggest a new schedule. That’s the position most buyers will be in.

The vendor’s first instinct, when you next sit down to renegotiate, will be to charge you per agent, per token, per API call, per outcome, or whatever unit the spot-rate market favours that quarter. The vendor will write the new schedule before you do. That’s the leverage problem.

There’s a quieter problem underneath that one. Most procurement teams haven’t really renegotiated a SaaS contract for the agent era yet, because the renewals so far have been near-identical to the originals with the prices adjusted. The vendor sent the same redlines. The buyer ticked the same boxes. The clauses that defined what a “user” is, what a “session” is, what “concurrent” means — those were inherited from the original contract and never updated. They describe a world that has now ended.

What to actually do

If you’re a buyer with a renewal due in the next twelve months, here’s the list.

1. Read your current contract for “user” and “seat” definitions — and mark every adjacent clause. It isn’t just the user/seat clause that breaks. Audit rights, overage treatment, API-call entitlement, liability for actions taken by automated callers, data-access scope, service levels on API uptime, and any “no automated access” provision were all written for a world where seats and humans corresponded. Mark each one. Each is contested ground at your next renewal.
2. Ask for the agent pricing schedule now, before you need it. Every major vendor either has one or is drafting one. You’ll want to read theirs at your leisure, not at renewal under deadline pressure. The first generation of these schedules will be vendor-favourable and full of language about “consumption units” that nobody has yet defined consistently — better to scrutinise that on your own clock than under a redline timer.
3. Add agent-readiness criteria to your next RFP. Not “does the vendor have an MCP server” — every vendor will tick that box. The questions worth asking are about authentication for non-human callers, error message structure, audit trails when an agent acts on a user’s behalf, how the vendor handles a single user running twenty parallel sessions, who carries liability when an agent acts on bad input, what API uptime guarantees are on offer, and whether the data-access scope is clear enough to put into a DPIA. None of these are nice-to-haves; they’re the new procurement basics.
4. Decide which workflows you actually want behind an agent. Not every workflow benefits from being headless. The conversation with a customer-success manager isn’t better with a layer of intermediation. A finance approval that needs a human on the hook needs an interface. Some of your premium-tier value will turn out to be decoration the agent routes around; some will turn out to be load-bearing. Knowing which is which, before the vendor tells you, saves the negotiation.
5. Pre-empt the licence true-up. If your current contract counts seats and you’ve handed those seats to agents over the next twelve months, the vendor’s audit team will find it. Better to renegotiate the unit while the contract is still being honoured than to discover the cost when year-end reconciliation lands.

What “agent-readiness” should mean in your RFP

“Agent-ready” is about to do for SaaS RFPs what “cloud-ready” did for IT in 2012 — it’ll mean everything and therefore nothing, unless you sharpen it on the way in. Here’s a narrow definition worth using: a product is agent-ready when an off-the-shelf agent, given the public API and docs, can complete one of your common end-to-end workflows without a human helping it parse an error or guess an endpoint. Anything else is a marketing claim.

A couple of these aren’t fully settled yet. Authentication for non-human callers, in particular, is unclear — there isn’t a settled industry standard, and I’d be wary of any vendor claiming there is. The questions worth asking are still the same. Are errors machine-readable? Can authentication work without a browser redirect? Does the product expose telemetry that lets you see what the agent did and why? Is the pricing schedule legible when a single user runs ten parallel sessions? Who’s liable if the agent does something with your data that it shouldn’t? What’s the SLA on API availability when your workflows depend on it? These aren’t exotic concerns; they’re the basics, and most vendors will fail at least one of them today. The point of asking now is to discover which ones, before you sign.

The window that closes

The seat licence was a useful fiction, held in place for two decades because nobody had a reason to challenge it. The vendors have just torn up their side. Whether your finance team and your procurement function tear up theirs in time to negotiate from a position of strength — or whether they wait to be told what the new unit is — will decide who sets the price of the next decade of enterprise software.

The post The headless turn: what to renegotiate before your next SaaS contract appeared first on Matt Hopkins.