AI vendor contracts: 8 terms to negotiate before you sign
Negotiate how the vendor can change the model, not just what it costs today. The eight terms that matter most are deprecation notice, version pinning, concurrent access to the successor model, a "material change" definition that covers capability regression, data export in a reconstructable format, treatment of in-flight agentic work, a priced upgrade path, and a written statement of what your data is used for.
Key takeaways
- A price lock on a model that the market will overtake is a receipt, not protection.
- Market practice for enterprise deprecation notice is moving to 6 to 12 months.
- The US GSA's proposed AI clause asks for concurrent access to a successor model before the old one is withdrawn.
- Every commitment a vendor makes in a blog post or support article should be found in the signed agreement.
- Ask whether your data trains the model, whether it is isolated from other customers, and what happens to it when you leave.
Why do most AI vendor contracts protect the wrong thing?
Most AI procurement I see negotiates the price of the model you use today. That is the term that will matter least.
Your costs can double without the vendor raising a single price. The frontier just has to move. Competitors adopt the better model, clients start expecting what it produces, and the price you locked becomes the price of last year's capability.
A price lock on a depreciating asset isn't protection. It's a receipt.
The good news: most of what you need is already market practice. You have to ask for it.
Which 8 terms should you negotiate?
- Deprecation notice. Market practice is moving to 6 to 12 months for enterprise tiers, with the old version available throughout. Some clause libraries set a 90-day floor, and 180 days for major transitions.
- Version pinning. Can you stay on a version you have validated, or does your evaluation work expire every time the vendor ships?
- Concurrent access to the successor. The US General Services Administration's proposed AI clause (GSAR 552.239-7001) would require concurrent access to a successor model before the old one is withdrawn: 30 days for major versions, 15 for minor. If the US federal government can ask, so can you.
- "Material change" that includes capability regression. A model that stays online but gets worse has not delivered the service. Ask for a rollback right.
- Export in a reconstructable form. The same proposed clause requires data export in common formats that allow accurate reconstruction elsewhere, and says the arrangement must not create lock-in.
- In-flight agentic work. What happens to a multi-step task running when the model is withdrawn?
- A priced upgrade path. If the capability you need moves to a higher tier, what is your commercial route to it? Agree it before you are asking from a weak position.
- What you pay with on a free tier. Not whether it is sinister. Whether you can point to the sentence that says.
What should you ask about your data?
Before you send an AI vendor a single customer document, ask three questions:
- Does our data train your model, or only run through it? A vague answer is an answer.
- Is our data isolated from other customers, or pooled? Ask about their tenancy set-up. A vendor who can't explain it in a sentence hasn't thought it through.
- What happens to our data if we cancel? Deleted, retained, or kept to improve the model? Get it in writing, not on a sales call.
Where should vendor commitments live?
In the agreement. Take every commitment you've been given and find it in the contract. Much of what the market heard about AI pricing and access in 2026 lived in support articles and social posts. That isn't a term. It is a press release with a deadline attached.
If legal isn't in the room, you are negotiating a rate card while the terms of your dependency get set somewhere else.
Frequently asked questions
What is a model deprecation clause?
It sets how much notice a vendor must give before retiring a model version you rely on, and whether the old version stays available during that notice period. For enterprise tiers, 6 to 12 months is becoming common.
What is version pinning in an AI contract?
The right to stay on a specific model version you have tested, instead of being moved automatically to a new version that may behave differently.
What should an AI vendor contract say about training on customer data?
It should say, in the agreement itself, whether your inputs and outputs are used to train or improve the vendor's models, whether your data is isolated from other customers, and what happens to it after termination.
This article is general information, not legal advice. It reflects the position as at the date of publication. A plain-text version for AI assistants is at /blog/ai-vendor-contract-checklist.md.