AI Tokenomics: Why Your AI Budget Has No Ceiling Anymore
Microsoft Copilot Cowork bills by usage since 16 June 2026. Why the Mittelstand budgets AI by the seat but pays by the token — and how to cap the open line.
A finance director at a machine builder with 1,200 staff opens the first Copilot Cowork invoice issued after 16 June 2026. The seat licence sits where it always has, line by line, name by name. Underneath it is a second line that was not there last month: usage, measured in credits, carrying a number nobody on the team had forecast. He signed the contract in spring. He learned what a heavy task actually costs in the third billing month. Between those two moments lies the gap this article is about.
The vendors stopped pricing AI by the seat a while ago. Most Mittelstand budgets did not notice. They still carry AI as a fixed cost per user, planned like any other software licence, while the invoice underneath has quietly switched to a meter that never stops turning.
The invoice now runs on two logics
A seat licence is a fixed price per user per month. You know it on the day you sign. Usage-based billing adds a second figure on top, and that figure is set by what your people actually do with the model.
Microsoft made the switch concrete on 16 June 2026, when it announced worldwide general availability of Copilot Cowork (MSDynamicsWorld, 16 June 2026). The model: roughly 30 dollars per seat, plus 0.01 dollars per credit consumed. A light task burns around 125 credits. A heavy one burns around 2,500. That is a factor of twenty for the same headcount, the same licence, the same month. The seat line holds steady. The usage line is the one that moves.
There is a spend limit. Administrators can cap consumption per user and per group in the admin centre. The catch sits in the default: at general availability that limit is switched off. A grace period runs until 1 July 2026, after which the meter bills in full. The control exists; nobody turns it on by accident.
The vendor knows it stings
The most honest description of the problem this month did not come from a critic. It came from a vendor strategist. Bill Patterson of Salesforce put it plainly on 23 June 2026: ask most finance directors whether their AI tools are paying off, and the answer is “no, but I’m paying for it” (diginomica, 23 June 2026). He went further and called a poorly governed AI deployment “a very expensive Google search.” Adoption is climbing. Yield is not following at the same speed.
That is the uncomfortable part. The spend is real and dated. The return is a forecast. When a vendor’s own go-to-market lead says the quiet part out loud, the budget conversation has already moved past marketing.
The numbers from the Mittelstand
Bitkom, the German digital industry association, published its 2026 study on artificial intelligence in February 2026, surveying 604 companies with 20 or more employees. The headline reads well: 41 percent use AI productively, and 77 percent report a stronger competitive position. The line that belongs in the budget meeting reads less well: 33 percent pay more than they expected (Bitkom 2026 study report, February 2026).
Gartner sharpens the same point from the project side. In its 2026 Hype Cycle for Generative AI, summarised at the end of May, the firm expects at least half of all GenAI projects to overshoot their cost budget, and at least half of proofs of concept to be abandoned, with “escalating costs” named among the reasons (The Register on Gartner, 28 May 2026). Gartner also frames per-token pricing across thousands of users as a TCO problem rather than a line item (Gartner press release, 18 June 2026).
The pattern is not “AI is too expensive.” It is narrower and more fixable: budgeted by the seat, billed by the token. That mismatch is the whole problem.
Why this lands now and not next quarter
Tokenomics went from a niche complaint to an analyst top story inside a single week. diginomica ran it as its weekly lead on 22 June 2026 (Enterprise hits and misses, 22 June 2026). Constellation Research published a warning on 21 June. ComputerWeekly built a CIO roundtable around it on 23 June (ComputerWeekly, 23 June 2026). Three independent analyst desks, one theme, one week.
The Microsoft date is what makes it operational rather than merely topical. The switch from seat to usage is no longer a trend on a slide. For Copilot Cowork it is a billing date with a grace window that closes on 1 July 2026. That is the difference between something to watch and something to put on the agenda before the next invoice. The same usage-cost dynamic is why some companies are bringing their ERP systems back once the open-ended cloud line shows up in the third month.
Key Takeaways
Before you sign the next AI contract, separate the bill into what is fixed and what is open. The seat budget hides the usage line until the meter has already run for a quarter. The fix is to split the figure before you commit, not after.
Fill in this table once, against your own last invoice:
| Column | What goes here | Why it matters |
|---|---|---|
| Fixed (seat / licence) | Price per user per month, planned for the year | The number your CFO already trusts |
| Usage (token / credit / task) | Per-activity charge, e.g. Copilot Cowork at 0.01 dollars per credit | The number that moves; the source of the overrun |
| Cap (contract clause) | The clause that limits the open line | The difference between a budget and a bet |
Four steps for Monday morning:
- Take the last three AI invoices and split each one into fixed versus usage.
- Name the usage drivers: model choice, tool calls, task runtime, context retrieval — the four cost levers behind a Cowork task.
- Set a spend limit per user and per group in the admin centre. It exists; at general availability it is switched off by default.
- Write a cap clause into the next contract: a consumption ceiling, an early warning at a defined threshold, and the right to downgrade to a cheaper model.
Before and after, in one line: a light Cowork task costs about 125 credits, a heavy one about 2,500 — a factor of twenty on identical seat counts. Knowing which of your tasks are heavy turns a surprise into a plan.
The limit: this does not apply if your AI use stays plain chat within the seat envelope. In that case the flat rate is still the cheaper model, and you can leave the meter alone.
What the market numbers don’t show
The strongest objection to all of this does not come from a marketer. It comes from the vendor strategists, and it deserves a straight answer. Pay-as-you-go, the argument runs, is fairer than a flat rate. You pay only for what you use; under a seat price the light users subsidise the heavy ones. Patterson makes that case, and Microsoft justifies the Cowork switch by pointing out that flat rates cannot survive a heavy user running hundreds of tasks a week.
The argument is not wrong. It simply answers a different question. Fairness per use does not solve budgetability. A fair variable bill is still a variable bill, and a finance director needs a number he can defend to the board a year in advance. “We will pay what is fair” is not a figure you can put in a spreadsheet in January.
There is a second objection worth holding onto, and it cuts the other way. Steve Pagram, speaking via Constellation Research on 21 June 2026, warned against letting every vendor solve the cost question for you, because that is the fastest route to losing your own AI strategy (Constellation Research, 21 June 2026). Budget control is not bookkeeping bolted on after the fact. It is part of architectural sovereignty, the same lesson that runs through the four mid-market reports nobody reads from the user’s side rather than the vendor’s. Hand the meter to the vendor and you have handed over the plan with it.
None of this argues against using AI. It argues for pricing it before you sign. The vendor-roadmap gap is older than tokenomics: SAP has positioned Joule as a native part of S/4HANA since TechEd 2024, yet the DSAG Investment Report 2026 (n=337, 26 February 2026) measures just 3 percent productive use. DSAG, the German-speaking SAP user group, keeps surfacing the same lesson the invoices now teach in cash: adoption is not yield, and AI without proper ERP integration stays an expensive toy no matter how the licence is priced.
Frequently asked questions about AI tokenomics
What is the difference between a seat licence and usage-based AI billing?
A seat licence is a fixed price per user per month, as predictable as any software licence. Usage-based means you also pay per actual AI activity — for Microsoft Copilot Cowork, in credits at 0.01 dollars each as of 16 June 2026. A light task runs around 125 credits, a heavy one around 2,500, a factor of twenty for the same number of users. The seat line stays flat; the usage line swings.
Can I cap the usage risk without banning AI outright?
Yes. Microsoft allows spend limits per user and per group in the admin centre, although at general availability the function is switched off by default and an administrator has to set it. Beyond that, a cap clause belongs in the contract: a consumption ceiling, an early warning at a threshold, and the right to move to a cheaper model.
Is AI still worth it for the Mittelstand under this billing model?
The numbers say yes, with a caveat. Bitkom’s February 2026 study reports 41 percent productive use and 77 percent claiming a stronger competitive position. But 33 percent pay more than they expected. The benefit is real; the cost side is what has to be calculated before you sign, not after.
Why are vendors moving to usage pricing if seats are more predictable?
Because flat rates lose money on heavy users. Someone running hundreds of AI tasks a week costs the provider a multiple of the licence fee. Salesforce and Microsoft argue pay-as-you-go is fairer, which is true for the light users. It does not solve the budgetability problem of a mid-market firm that needs one defensible annual figure.
Does the EU AI Act competence duty apply to me here too?
Indirectly, and it is budget-relevant. Article 4 of the AI Regulation (Regulation (EU) 2024/1689) has required adequate AI literacy in the workforce since 2 February 2025. In the Bitkom 2026 study, 53 percent name missing AI competence as their biggest hurdle and 43 percent still offer no AI training. Training is therefore not optional polish but a mandatory cost line — and it belongs in the same budget as the token spend.
Next step
Do you know which part of your AI bill is fixed and which part is open?
I am happy to compare your situation against the seat-versus-usage split — we take your last three AI invoices and separate the fixed line from the consumption line, then look at where a cap clause would actually bite. No deck, no pitch, just your numbers on the table.
→ Or read first: AI & automation with calculable costs · AI without ERP integration stays an expensive toy
Sources and links: diginomica — Patterson on tokenomics (23 June 2026) · diginomica — hits and misses (22 June 2026) · Constellation Research (21 June 2026) · MSDynamicsWorld — Copilot Cowork GA (16 June 2026) · The Register on Gartner GenAI (28 May 2026) · Gartner press release (18 June 2026) · ComputerWeekly roundtable (23 June 2026)
Read more on pfisterer.xyz: Cloud exit: why companies are bringing their ERP back · Mid-market ERP 2026: four reports, one pattern · AI without ERP integration is an expensive toy