Stop buying AI seats. Start buying decisions.

Microsoft passed 20 million paid Copilot seats this year. PwC asked 4,454 chief executives what AI returned and 56% said nothing measurable. Those are the same fact — and the seat is the reason. Here is the audit that reprices your licence bill before the subsidy does.

Andrew Ramsden 6 September 2026

Somewhere in your organisation there is a line on a bill for AI licences, and nobody has asked out loud what it bought. I have sat in rooms where that question hung in the air unasked, and I have not always been the one to ask it. Asking looks like penny-pinching, because the licences are cheap. The adoption chart is going up. And the person who signed the contract is usually two seats away.

Two numbers, measured at opposite ends of the same market, say the question is overdue.

Microsoft told its April earnings call it had reached 20 million paid Microsoft 365 Copilot seats, up from 15 million at the call three months before. Accenture alone is rolling Copilot out to around 743,000 of its people, which Microsoft says is the largest enterprise deployment so far.

Then PwC asked 4,454 chief executives across 95 countries what twelve months of AI had returned. Fifty-six per cent said they had seen no significant financial benefit. One in eight, 12%, had got both lower costs and higher revenue.

Twenty million seats. Fifty-six per cent nothing. Those are not two findings in tension, waiting for somebody to explain one of them away. They are one fact seen from the purchasing end and from the profit-and-loss end.

The seat is the bug

A seat is a licence to access a capability. It is not a change to how the work runs.

That sounds like a pedantic distinction until you follow the money. When you buy a seat you have bought one person the option to use a model, whenever they like, on whatever they choose to point it at. Nothing about the process changed. No step was removed, no handoff eliminated, no cycle time cut. Someone drafts an email faster, and the email still gets written, sent, read, actioned and filed by the same people in the same order.

That is why the dashboards look healthy while the profit-and-loss sits still. Utilisation is not a business outcome. A well-adopted seat produces exactly one thing you can measure, and that thing is a bill.

It is a strange bill too, because it is priced per person rather than per unit of work. Your cost scales with headcount and with enthusiasm, how many people hold a licence and how chatty they are, rather than with how many claims you process or how many jobs you quote. There is no denominator. You cannot say what a seat costs per transaction, because a seat was never attached to a transaction.

You already know this. You have looked at the licence count and the adoption chart and thought, quietly, what did we actually get? And yet seats have been cheap enough that the question never quite got asked.

They are about to stop being cheap enough.

The price you are paying is not the price

OpenAI's 2025 financial documents, obtained by the journalist Ed Zitron and reported in June, show revenue of $13.07 billion and an operating loss of $20.92 billion. Add those together and the company spent something close to $2.60 for every dollar it took in. Its headline net loss of nearly $39 billion is mostly a one-off charge from the restructure, and quoting that figure would overstate the case. The operating loss is the honest number, and the honest number is enough.

That gap is your discount.

Investors are funding the gap, at a scale that cannot run forever. Crunchbase counted a record $510 billion of global venture funding in the first half of this year alone, more than the $440 billion invested across the whole of 2025, with more than 70% of second-quarter capital going to AI companies. That money buys market share, and the way it buys market share is by selling inference for less than it costs to serve.

I cannot tell you when that ends, and nobody credible can. What I can tell you is what it means for a per-seat contract. Your renewal price is set by somebody else's willingness to keep losing money. You do not control that variable, you cannot forecast it, and right now it is the largest hidden assumption in your AI budget.

The standard advice is to negotiate harder, lock in a multi-year deal, push adoption up so the cost per seat looks better. All of that treats the price as the problem. The price is not the problem. The unit is the problem, and a cheaper seat is still a cost with no denominator.

Two questions, in order

So run the audit, in two passes, in this order. The second question means nothing without the first.

Pass one: does any named piece of work run differently?

Go line by line down the licence bill. For the people holding those seats, ask whether any named piece of work runs differently because they have this. Not "do they like it". Not "do they log in". Does something a customer or a colleague receives arrive faster, cheaper or better, in a way somebody would put a number against?

You will find three groups. A small one where the answer is obviously yes: they have both the work and the skill, and they will tell you exactly what changed. A middle group who use it pleasantly and produce nothing you can point at. And a group who were provisioned because provisioning everybody was easier than deciding.

Most licence bills are mostly the last two. That is what happens when a company buys access instead of change, and most of us have signed off on exactly that at some point.

Pass two: at full price, would you still buy it?

For every line that survived pass one, ask the harder question. If this cost three times what it costs today, would you renew?

Three is a stress test, and it is roughly the shape of no subsidy at all against a supplier spending $2.60 to earn a dollar. Some lines pass instantly, because the value of the work obviously exceeds the price. Some pass for a handful of people and fail for the rest, which tells you the seat count is wrong rather than the tool. And some fail flatly, which tells you the thing was only ever viable while somebody else paid part of it.

Every line lands in one of three places.

Theatre. No named work runs differently. The line fails pass one, so pass two never applies. Cancelling it changes nothing except the bill.
Countdown. Real work runs differently, and yet the sums only hold at today's subsidised price. There is a clock on this line that you did not start and cannot see. This is the dangerous group, because it looks exactly like success on every dashboard you own. You are renting an advantage on somebody else's balance sheet.
Load-bearing. Real work runs differently, and you would still buy it at three times the price. The business leans on it and it does not move.

Then count the seats in each group, and the whole bill takes the name of whichever holds most of them. Theatre if most of your seats buy access rather than outcomes. Load-bearing if half or more earn their keep at any plausible price. Countdown for everything in between: real work, real value, running on a clock somebody else set.

Where the money should go instead

Cancelling licences is not a strategy either.

The alternative to a seat is what I have come to call a judgement point: one bounded decision, made by a model, at a single named step inside a process. It classifies a case, matches a record, flags an exception, or summarises a file for the next step. Not a model sitting across the whole workflow reading everything, but one call, on one question, with an escalation to a person when its confidence is low.

That shape fixes the thing the seat broke, because it has a denominator. A judgement point costs a predictable fraction of a cent per transaction, because you have told it precisely what to decide. You can put that number in a spreadsheet beside the cost of the step it replaced. You can defend it at any token price, because if prices triple you know exactly what triples and exactly what it saves. A seat estate has no such row. It has a headcount and a hope.

There is a second benefit that turns up a year later. Because the model only ever sees one bounded question, you can change which model answers it. Judgement points move between suppliers in a way that a vendor's desktop assistant deliberately does not. When prices move, and they will move in both directions, you are choosing rather than renegotiating.

And yet the denominator is not the real prize.

What you are actually buying

Think about the processes your company has never automated. They were rarely blocked by the mechanical work: moving data, filling fields, routing, chasing, reconciling against a rule. That has been automatable for twenty years, and you have almost certainly done it wherever it was worth doing.

They were blocked in the middle, at one or two moments where somebody had to read something and decide. Does this invoice match this contract. Is this complaint about the product or the delivery. Does this application need a human. Is this photo showing the damage the claim describes. One unautomatable moment in the centre of an otherwise mechanical chain was enough to keep the whole process manual, so it stayed manual, year after year, until it stopped being discussed as a candidate at all.

That is the category that just opened. Not the vague version that sold twenty million seats, that AI can do knowledge work. The specific version: whole classes of process that were permanently off the automation list can now run end to end, because the one step that blocked them has a machine answer, a confidence signal, and a person behind it when confidence is low.

So when you fund a judgement point, you are not buying a faster step. You are buying the process that step was holding hostage.

Front-load the tokens into the build

Something else changed, and most budgets have not caught up with it. Building the automation itself got dramatically cheaper.

Agentic tooling now does the construction work that used to be a six-month project with a vendor and a statement of work: reading the existing system, writing the integration, generating the tests, wiring the exception paths. That work is genuinely token-hungry. It is also the right place to be token-hungry, because you spend it once.

Which gives you the shape of a sane AI budget, and it looks nothing like a seat licence.

The build bill. Large, deliberate, one-off. This is where you want the tokens going: agentic tooling constructing the automated process around your judgement points. Spend generously. It is capital, and this work has never been cheaper to do.
The run bill. Small, per-transaction, forever. Once built, the process runs the way processes have always run. Deterministic code does the mechanical steps at effectively no cost and calls a model only at the judgement points, only when it reaches one, only for as long as the decision takes.

That split is the whole reason rising token prices should not frighten a company that has done this work, and should frighten one that has not. Your run bill touches a model at three or four moments per transaction instead of continuously. If prices double, a per-seat estate doubles across every person holding a licence. A built process barely moves, because most of what it does was never a model call.

The uncomfortable part is the timing. The build is cheapest to do now, in the noisy part of the cycle, while agentic tooling is priced the way it is today. Convert seat money into build money this year and you own processes that run cheaply for a decade. Keep renewing per-person licences and you buy the same thin capability again every twelve months, at a price somebody else sets.

The best argument against everything above

Somebody in your room will make this one, and it is a good argument.

It runs like this. Tokens are about to get much cheaper anyway, so optimising now is premature. Open-weight models have flooded the market and commoditised what the frontier labs were charging a premium for. Underneath the models, every layer of the stack is on a downward cost curve: chips, power draw per unit of work, cooling, the useful life of the hardware, and eventually whatever exotic answers get built for the energy problem. Richard Sutton named the underlying dynamic in a 2019 essay he called the bitter lesson: general approaches that scale with computing power tend to beat cleverly hand-built ones, precisely because they ride the falling cost of computation. Betting against cheap compute has been a losing trade for seventy years.

I think that argument is right. And yet it does not get you out of this.

You do not get to spend the long horizon's prices today. Every month between now and then runs on this year's bill, and this year's bill is what your board is looking at. A company running an efficient process and a company running a wasteful one both benefit when prices fall. Only one of them is still in a position to press the advantage when it happens.

The horrific bills are also available at today's prices, and this is the part people underestimate. The failure is almost never the price per token. It is the shape of the usage: a model re-reading whole documents on every transaction, an agent looping because nothing told it when to stop, a workflow calling a frontier model for a decision a small one would have made correctly. Companies discover this monthly, at current, subsidised, historically cheap prices. If prices fall ninety per cent and the shape never changed, a horrific bill becomes a merely bad one.

Now the concession, and it is a real one. You should run an experimental phase, and you should run it early. You cannot reason your way from a whiteboard to which judgement points matter in your business. You find out by trying things, cheaply, in places where being wrong is survivable. Spend the money to learn, and shorten the phase wherever you can by learning from everyone who has already run the experiment, because the failures are widely published now and there is no prize for discovering them personally.

But a phase is a phase. It ends, and what it produces is decisions: which processes, which judgement points, which model for which decision, what gets built once and what gets paid for forever. That is the same discipline this audit is asking for. The experiment is how you earn the right to make those calls on evidence instead of vendor slides.

Cheap compute rewards whoever is positioned to use it. It does not refund whoever was not.

And the people

This is the part that decides whether the whole exercise was worth anything.

When a built process takes hours out of your week, those hours need a named destination or they evaporate into slack while the dashboard stays green. The destination is the work you currently cannot deliver enough of: the difficult conversation, the bespoke job you turn away, the follow-up nobody has time for, the quality pass that never survives month-end. That work is what customers choose you for, and it is the only part of any of this that a competitor cannot buy from the same vendor you did.

Seats spread a thin capability across everybody, forever, at a price you do not set. A built process puts a deep one exactly where the work was stuck, pays for itself once, and puts your people back where the process was never the point.

One boundary, honestly

It cannot tell you what tokens will cost next year. Nobody can, and no vendor is going to volunteer it. What it does is take the assumption out of your budget. It tells you which lines already earn at any price, which lines live on somebody else's losses, and which lines were never doing anything at all.

Run it on one licence bill. Give yourself thirty minutes and honest answers, and by the end you will have every line labelled, the seat count you actually need, and one word to take into the renewal conversation.

Start with your largest line by seat count and ask pass one: what runs differently? If the room cannot name a workflow in a sentence, you have just found the money for your first judgement point. I would like to hear how it lands.

Attribution

Written by Andrew Ramsden. AI tools assisted research and drafting; all outputs verified.

Accountable

Andrew Ramsden.

Limitations

OpenAI's figures come from internal documents reported by TechSpot, not published accounts. PwC's 56% is its own wording, "no significant financial benefit".

References

Seven sources, retrieved and hashed on 26 and 27 August 2026, each figure quoted with a locator in the SOURCED sidecar.