AI Vendor Lock-In Turns Usage Into a License Metric
Anthropic now bills Claude Enterprise as a seat plus usage at API rates. OpenAI priced GPT-5.2 at $1.75 per million input tokens, up from $1.25 on GPT-5.1. GitHub moved Copilot from request units to GitHub AI Credits on 1 June 2026. Zapier surveyed 542 US executives with paid AI contracts: 89 percent still thought they could switch vendors within four weeks. Of those who had actually tried, 58 percent said the migration failed or took significantly more effort than expected. That is the AI vendor lock-in assumption.
That’s the problem.
This is not a cloud-bill story. It is a closed buffet meeting a stack you cannot leave in a week. Tokens, inference, and restricted compute are the new license metric. Do not let the vendor frame the next conversation as “you still have a seat” or “this is just usage.” The question is what you are entitled to, what you actually consume, and whether you can leave the meter you are already on.
What the primary sources actually say
Zapier says it polled 500 U.S. enterprise executives. Nearly nine in ten (89 percent) said they could switch within four weeks. Forty-one percent said two to five business days. Two-thirds had already attempted a migration. Among that group, only 42 percent said it went smoothly. Seventy-four percent said losing their primary AI vendor would disrupt day-to-day operations or leave them reliant. Only 6 percent said they could stop using it without interruption.
Zapier’s own write-up is the SAM question. When AI is woven into processes and tuned with “temporary” adaptations nobody documented, you do not have a four-week project. You have a dependency you never inventoried. Data migration and single-vendor overdependence were the top lock-in concerns, each at 46 percent. Sudden price hikes sat at 41 percent.
The commercial side moved in the same window. Anthropic’s Help Center now states that the Enterprise seat “doesn’t include any usage” and that every token bills at standard API rates. GitHub’s Copilot seat prices did not move; the unit inside the seat did. OpenAI’s own GPT-5.2 post put the input rate at $1.75 per million tokens against $1.25 for GPT-5.1: a published model-generation step-up, not a rumor.
Do not invent a 2–3× and call it your forecast. That range is not a published Anthropic multiplier. Your number is your history against the meter you are now on.
The meter is a license problem, not just a cloud bill
FinOps will see a token invoice and treat it like a cloud bill. That is half the file.
SAM already has language for the other half. Entitlement is what the paper still gives you: seats, a named plan, an included pool, a fixed tier. Consumption is what you burn once the buffet closes: tokens, inference, compute that can be restricted, a model that can be withdrawn from the plan you thought you bought.
If that feels familiar, the problem is not your team’s maturity. You have been managing AI like a SaaS seat. The vendor is now selling a seat plus a meter, on a stack you cannot swap in a week. A reporting layer that only prints named users will miss the bill. A FinOps view that only prints tokens will miss the entitlement you still pay and the lock-in that makes those tokens non-portable. More AI output and slower exec decisions is the five signs pattern; the model is the commodity is the Microsoft-side proof that switching models is not switching vendors.
Traditional SAM tools collect comprehensively and answer slowly. They were built to prove who has a seat, not to recast a closed buffet. The gap is interpretation work.
Picture your procurement director, sixty days from an AI renewal. The vendor walks in with a usage-based bill and a story that models are interchangeable. From Copilot to coworker is the workforce overlay if those meters are digital employees; Microsoft integration is the Graph file if Copilot sits next to the token invoice. Your CIO wants three answers this week: what we are entitled to on the current paper, what we actually consume, and whether we can move in four weeks if the meter jumps. If that pack takes three weeks of specialist reading, you have already lost the framing of the meeting.
An analysis that arrives after the salesperson has named your plan is a document. One that arrives during the meeting is a decision input.
AI vendor lock-in is why you cannot shop the meter
A cheaper token rate is not a license position if you cannot move. The C-suite bet was that frontier models are a runtime. For a developer on a personal plan, that is still roughly true. For an estate with vendor APIs, proprietary training data, and undocumented workflow adaptations, it is a hallucination with a survey behind it.
FinOps can watch the burn. SAM has to say whether the burn is portable. If 58 percent of attempted migrations already fail or drag, the meter is a commercial trap, not a shopping list.
What SAM should treat as entitlement versus consumption
Do this in the next thirty days, as a decision pack rather than a project.
| Treat as entitlement | Treat as consumption | Why the distinction changes the meeting |
|---|---|---|
| Seats, named plans, and whatever the current paper still includes | Tokens, inference, restricted compute, model access that can be pulled from a tier | A clean seat file can still blow the budget the week the buffet closes |
| Any remaining fixed or “sweetener” tier | What you burn once that tier includes less compute | Renewal math is the delta, not last year’s flat fee |
| The vendor relationship you cannot unwind in four weeks | The workflows, context, and undocumented adaptations that sit on that vendor | Lock-in makes the meter non-portable. Shopping a cheaper rate is not a position if you cannot move. |
| On-prem or “open” weights you booked as an escape | Whether that stack is still supported, or has been abandoned for a proprietary successor | Entitlement you cannot operationalize is not a position |
Then say, in one page, which AI contracts are still a buffet, which are already a meter, and which estates you could not leave if the price moved this quarter. Can a non-specialist brief the CIO in fifteen minutes from the pack you have today? If the answer is no, you do not have an AI position. You have a seat list and a cloud invoice.
The decision layer, not another inventory
You already have seat files, usage exports, and a FinOps view of tokens. The gap is not another inventory. The gap is turning that estate into a decision: what you are entitled to, what you consume, and which vendor you are already locked to before the next conversation. LICENSEWARE sits on the inventory and ITSM tools you already run. It is not a rip-and-replace SAM suite. It is a decision layer: what matters, why it matters now, what should happen next. That is renewal optimization on entitlement versus consumption, not another token export.
If you are heading into an AI renewal and your current tools still need three weeks to recast seats, meters, and exit risk as one position, book a Software Intelligence Review. You can also start on the free plan and run the analysis on your own data.
The question to walk in with
Do not let the vendor frame this as “you still have a seat” or “this is just usage, like cloud.” The right question is: what are we entitled to on the current paper, what are we actually consuming, and can we leave if the meter jumps. That is a data question, not a sales question.
The vendor will walk in with a meter. The only question is whether you have the entitlement, the consumption, and the lock-in map first: current, defensible, and tied to the contract in front of you.
