← All articles
AI Strategy4 min read

What to check in an AI vendor contract before you sign

Before you sign an AI vendor contract, check four things: whether your data trains their model, who owns the outputs, what the renewal and exit cost, and who pays when the model is wrong.

What to check in an AI vendor contract before you sign

An AI vendor's sales deck will show you the demo. The contract shows you what you actually bought. Most expensive surprises in AI projects do not come from the model — they come from clauses nobody read: who owns the outputs, what happens to your data, and what you are locked into when the vendor changes pricing or retires the model your product depends on.

Read the contract before the pilot. Four areas matter.

Data: where it goes and what it trains

The first question is not "is our data secure?" but "does our data train their model?" A vendor that trains on customer inputs is building its product with your data, and in regulated industries that alone can breach your own client obligations. Look for an explicit no-training clause and a data-processing addendum that names sub-processors, retention periods, and deletion on termination.

Ask where data is processed and stored, and whether support staff in other regions can access it. For a hospital, bank, or government client, the answer decides whether you can use the vendor at all.

Outputs, IP, and the model underneath

Who owns what the system produces? For most buyers the answer should be: you own the outputs, the vendor owns the model. If the contract is silent, assume the worst and get it written in. The same applies to fine-tuning: if you pay for a model tuned on your data, does the tuned model stay yours when you leave?

Then check the model dependency. If the product is a wrapper around a third-party model, what happens when that model is deprecated or repriced? Ask for notice periods and a migration path.

Pricing, lock-in, and the exit

AI pricing moves fast, and contracts written around per-token or per-seat rates can quietly double your bill as usage grows. Look for the renewal clause: auto-renewal with a price increase you have to opt out of is the most common trap. Negotiate caps on annual increases and a right to export your data — prompts, configurations, conversation history — in a usable format on exit.

Run the exit test before signing: if you cancelled in six months, what would you lose? If the answer is "everything, including our configuration and history," the switching cost is the real price of the contract.

Liability: who pays when the model is wrong

Models hallucinate, and sooner or later one will hallucinate on your customer's data. Check the liability cap and the indemnity language. A cap set at three months of fees sounds standard until the error costs a client relationship. You will rarely get unlimited liability, but you can usually get carve-outs for data breaches and IP infringement, plus a service-level agreement with real remedies — credits that actually bite.

None of this is legal advice, and for a large deal you should have a lawyer read it. But the engineering team should read it first, because they are the ones who will live with the data flows and the lock-in. We help companies evaluate AI vendors on the terms that actually matter — data, ownership, and exit — before the signature, not after the surprise.

Field note: In one engagement, a client was days from signing a support-automation vendor whose contract allowed training on all customer inputs and auto-renewed at "then-current pricing" with 30 days' notice. The demo was excellent. The contract would have put their customers' ticket histories into the vendor's training set and doubled the bill in year two. They renegotiated the training clause and a 5% annual cap — and kept the vendor.

Related service AI products

Building something worth shipping?

We take on a small number of AI product engagements. Tell us what you are building — we reply quickly.