The model contract: what not to sign with an AI vendor

A mid-sized distributor has spent a year teaching a hosted model to read supplier price lists. Hundreds of prompts tuned by hand, a few thousand graded examples, a small fine-tune that finally handles the odd formats from the two biggest suppliers. The purchasing team has stopped retyping anything. It is the best thing the company built that year.

Then an e-mail arrives. The model version they built on will be retired in ninety days. The replacement is better on the vendor’s benchmarks and worse on the price lists. The fine-tune cannot be downloaded, the graded examples live in the vendor’s console, and the monthly bill has already grown past what anyone approved, because usage went up as the tool got useful.

Nobody did anything wrong. The owner signed the standard terms, the same ones everybody signs, and the terms did exactly what they were written to do. The cause is not the vendor’s behaviour. It is that an AI contract is treated as a software subscription when it is in fact a lease on a piece of the company’s brain, and nobody read it as one.

Where the control leaks out

Your data trains their next product. Many standard terms allow the vendor to use inputs, outputs or both to improve their service. Every graded example your team produces, every correction, every document you send to be classified makes the model better for your competitor too, and you will be charged for the improvement next year.

Deprecation is theirs alone. The vendor can retire a model version on notice they set, with no obligation that the replacement behaves the same way. Your process was built on a specific behaviour. The contract only promises that some model will answer.

The things you built cannot leave. Fine-tuned weights, prompt libraries, evaluation sets and logs often sit in the vendor’s system with no export clause, or with an export that hands back a format nothing else can read. A year of tuning becomes a year of rent.

Usage pricing has no ceiling. Pay per token, per call, per seat, with the rate changeable on notice. The better the tool works, the more the company uses it, the higher the bill. No cap, no fixed rate period, no obligation to warn before a step change.

The audit runs one way. The vendor reserves the right to inspect how you use the service, to suspend you for breaches they define, and to change the acceptable use policy later. You get no equivalent right to see what happens to your data or where it is processed.

Jurisdiction and processing are a footnote. Where the data sits and whose law applies is in a schedule nobody reads, and it can be changed by updating a web page. For a company in the EU this is not a footnote; it is the clause that decides whether the whole arrangement is legal next year.

What to insist on instead

No training on your data, in writing, by default. Not an opt-out buried in a settings page but a clause in the agreement. If the vendor will not give it, that tells you what the real product is.

Deprecation with a floor. A minimum notice period you can live with, continued access to the version you built on for that period, and a right to export your assets before it goes. Ask for it in the contract, and run your own evaluation set against every replacement the day it is announced. We covered how to build that set in evals: how you know your AI works.

Export of everything you made. Fine-tunes as weights where the technology allows it, prompts and evals as plain files, logs in a documented format. If the vendor cannot hand back the fine-tune, treat it as rented and keep the training data on your side so it can be rebuilt elsewhere.

Pricing with a cap and a term. A rate that is fixed for a period, a monthly ceiling you set, and advance notice of changes long enough to switch. Volume discounts are fine. An open-ended meter is not.

Rights that go both ways. A description of where data is processed and under which law, notice before it changes, and the right to leave without penalty if it does. If the acceptable use policy can change by web page, so should your right to terminate.

A second model in the plan. The cheapest clause is the one you do not need, because the process can run on another endpoint or on open weights you keep. Why that option matters is the whole argument of open weights are the means of production.

None of this is legal advice, and the wording belongs with your lawyer. The point is to walk in knowing which clauses decide who controls the work, so the lawyer is negotiating the right things.

A test for this afternoon

Open the agreement behind the one AI tool your company would miss most. Find five things: the sentence about training on your data, the notice period for retiring a model, the clause on exporting what you built, the mechanism for changing the price, and the place where data is processed. Write each down in one line, in plain words, as it reads today.

Then read the five lines as if they described a key supplier of parts rather than software. Would you accept that the supplier can change the part, keep your tooling, raise the price without a ceiling and move production to another country by updating a web page?

Every line you would not accept from a parts supplier is a line you have already accepted from a model vendor. That is the negotiation list for the next renewal, and the tool belongs on the register from the dependency audit with a note about which of the five it fails.

Owners who negotiated these terms, or walked away and built on their own weights, are the people this podcast is being made with. What it is and who it is for is on the home page.