AI Vendor Contract Checklist: 8 Clauses to Fix Before You Sign

The most expensive clause in an AI contract is the one that isn't there. We say that because most of the vendor agreements we review in practice are derived from a standard software template and never mention the three risks specific to AI: training on your data, hallucination, and the chain of sub-providers behind the vendor. The contract gets signed, the system goes live, and when the first problem appears everyone searches for the same page. It does not exist.

This is an AI vendor contract checklist: eight clauses, each with why it matters, how it should read in the agreement and which body of rules it touches. Chatbot, document assistant, workflow automation or the enterprise subscription of an AI tool; the same eight clauses apply to all of them. For the decision framework around the AI investment as a whole, our end-to-end guide to AI for small business is the wider read; here we look only at the stretch before the signature.

One thing up front: this is not legal advice. Show the contract to a lawyer before you sign; this list exists so you know which questions to bring them.

Why does an AI vendor contract checklist differ from a software one?

A classic software contract covers a product whose behaviour is fixed: the code does what the code says. An AI system behaves probabilistically, may learn from data, and usually runs on a third party's model in the background. Those three properties create three risk areas the classic contract is silent on: where your data goes, how accurate the output is, and who in the chain is responsible for what.

Regulation adds a layer on top. If your vendor is a local firm but relies on an API from OpenAI, Anthropic or Google, customer data is leaving the country, and whichever cross-border transfer regime applies to you, GDPR-style standard clauses in Europe, Turkey's standard contracts, or equivalents elsewhere, comes into play. Custom development agreements also tend to fall under "work contract" rules in many civil-law systems, which makes acceptance testing and defect-notice periods something the contract has to spell out. The eight clauses below close both layers together.

Clauses 1 to 4: where does your data go, and who keeps it?

The first four clauses follow the data's journey: whether it is used for training, where and for how long it is held, who owns outputs and code, and who the sub-providers in the chain are. Without these four, the other four lose their meaning, because you have not drawn the boundaries of what you are protecting.

1. Your data is not used to train models

Providers now price training rights explicitly: a cheaper plan in exchange for the right to learn from your inputs. At the time of writing, the API and enterprise plans of OpenAI, Anthropic, Microsoft Azure and Google do not train on customer data; their consumer plans are a different story. An agency processing your data through an individual ChatGPT or Claude account changes the policy entirely. Two sentences belong in the contract: "The vendor and its sub-processors will not use customer inputs or outputs to train, fine-tune or improve any model," and "The vendor delivers the service exclusively through API or enterprise plans."

2. Retention period, location and deletion confirmation

The major providers retain inputs for abuse detection for a set period, commonly 30 days. In August 2026 both OpenAI and Anthropic announced zero-retention and customer-cloud storage options for enterprise customers; those are for "eligible customers" and a small business will rarely get them directly, but you can ask which plan your vendor is on. Contract wording: "Customer data is processed only on servers in [region]; processing logs are retained for no more than 30 days; on termination all copies, including backups, are deleted within 30 days and written confirmation is provided."

3. Intellectual property in outputs and code

"We paid for it, so the code is ours" is not automatically true. In most jurisdictions an assignment of economic rights has to be in writing and, in many civil-law systems, has to list the rights individually; if the contract is silent, the rights stay with the vendor and you hold a licence to use. For custom development, the agreement should name which repository, which version and which dependencies are delivered, and state that the rights to reproduce, adapt, distribute and communicate the work are assigned without time or territory limits. The vendor's pre-existing libraries get listed separately with a perpetual licence. On AI outputs, one caution: the contract can say "the vendor claims no rights in outputs," but it cannot say the outputs are copyrighted. Whether AI-generated content qualifies for copyright at all is unsettled in most countries.

4. Sub-processors are listed and changes are notified

Your contract is with the local vendor, but the data actually travels to the model provider, the cloud host and the vector database. Under GDPR-style rules that chain is controller (you), processor (vendor) and sub-processor (model provider), and it has to be documented. Wording: "The vendor lists all sub-processors and their processing locations in the annex; notifies the customer 30 days before adding one; the customer has a reasonable right to object; the vendor is liable for its sub-processors' acts."

Clauses 5 to 8: what happens when the system fails or the relationship ends?

The last four clauses govern the bad day: the system giving wrong answers, who carries the loss, parting ways with the vendor, and the invoice growing unexpectedly. What they share is that nobody wants to discuss them on signing day, and everybody looks for them on the first day something goes wrong.

5. An SLA that includes an accuracy metric

A 99.5 percent uptime commitment does not cover a chatbot that is running and answering wrongly. An AI-specific SLA defines an accuracy rate measured on a test set you prepared. Example: "The system maintains at least 90 percent correct answers on the 200-question test set in the annex, measured monthly. If the threshold is missed in two consecutive periods the vendor remedies within 15 days; failing that, 20 percent of the month's fee is credited, and after three periods the customer may terminate without penalty. Model version changes are notified at least 30 days in advance." The numbers are examples and vary by sector; the rule that does not vary is that you prepare the test set, not the vendor.

6. Liability cap and its exceptions

Vendors usually cap liability at the last twelve months' fees and add a broad disclaimer that "output may be inaccurate." Twelve months' fees do not cover a data-protection fine or a third-party copyright claim. What to ask for: either no cap, or a separate higher cap such as three times the fees, for confidentiality breaches, data-security incidents, regulatory fines and intellectual-property claims. Exclusions for intent and gross negligence are void in many legal systems anyway, but writing that down closes the argument before it starts. Hallucination now appears in contracts by name: the vendor will require human review, so you should write down which outputs go through human approval.

7. Exit plan and data portability

When the contract ends, conversation history, knowledge-base documents, embeddings, prompt templates and any fine-tuning data should be handed over in an open format. Example: "On termination the vendor delivers all customer data in JSON or CSV at no additional charge within 30 days; provides transition assistance for up to 90 days; and gives at least six months' notice before retiring a model or feature." Custom development needs one more step: source code held with an independent third party and released to you if the vendor becomes insolvent or stops the service. Where a formal escrow market is thin, shared access to the git repository or a notarised deposit does the job in practice.

8. A ceiling on usage-based pricing and a currency clause

Most AI services bill per token, message or call, and the underlying model cost is priced in dollars. The contract should define the billing unit, the monthly estimate, the overage rate and a ceiling you set; reaching the ceiling should trigger a notification and approval, not a cut-off. Price increases once a year, capped, with 60 days' notice and a right to terminate. If your local currency differs from the provider's, say who carries the exchange risk. And if the vendor switches to a cheaper model, the accuracy metric in clause 5 stays in force; otherwise the saving goes to the vendor and the quality loss to you. We worked out which sourcing model is cheaper over time in our three-year cost comparison; the pricing clause is the legal counterpart of that arithmetic.

Cross-border data transfer, with a Turkey spotlight

If your vendor uses a foreign model provider in the background and the chatbot handles customer names, phone numbers or order details, that is a cross-border transfer of personal data. In Europe that means standard contractual clauses and a transfer assessment. Other jurisdictions are converging on the same mechanism, with the added twist that some require the regulator to be told.

Turkey is a sharp example of that twist. Its data protection authority published standard contract templates in July 2024 and requires the signed contract to be filed through an online module within five business days; with no adequacy decisions issued for any country, the standard contract is in practice the only route between unrelated companies. Advisory sources put the 2026 fine for missing the filing at roughly the equivalent of a few thousand to several tens of thousands of dollars; confirm the exact figure on the authority's own site. Three questions for the vendor anywhere: which transfer mechanism was signed, was it filed where filing is required, and what is the reference number. Alternatives exist too: masking personal data before it leaves, or using a locally hosted model. For what Turkey's generative-AI guidance means for companies, our piece on its ten rules for businesses goes into the detail.

A concrete scenario: an e-commerce firm's negotiation

A 25-person e-commerce company bought a customer-service chatbot from a local vendor. The first draft ran to four pages: price, term, confidentiality, jurisdiction. The only sentence about AI was "the system uses artificial intelligence technology." When the firm ran this checklist against the draft, seven of the eight clauses were missing; only confidentiality was partly there.

The negotiation took three rounds. In the first, the vendor disclosed which model provider it used and documented that it was on an API plan; the training ban and the sub-processor list went in. In the second, the firm built a 150-question test set; the first measurement came out at 86 percent accuracy, the SLA threshold was written at 85, with service credits and a termination right. The third round was the liability cap: the vendor defended twelve months' fees, and the parties settled on a separate cap of three times the fees for regulatory fines and data breaches.

On data transfer, the firm asked for customer phone numbers to be masked before reaching the chatbot at all; the vendor implemented it in a week and the need for a transfer contract largely fell away. The monthly fee did not change; the contract grew from four pages to eleven. Eight months later the vendor wanted to move to a cheaper model. Thanks to clause 5, the test set was run first; the new model came in at 83 percent and the switch was postponed until it improved. Without that clause the firm would have learned about the change from customer complaints.

Frequently asked questions

The vendor says "we don't train on your data." Is that enough?

No. The model provider behind the vendor must not train on it either, and that depends on the plan: API and enterprise plans default to no training, consumer plans may not. Get "API or enterprise plan only" into the contract.

With a SaaS subscription, doesn't the data stay with us anyway?

No. In SaaS the data sits on the vendor's and its sub-processors' servers. Unless retention, location, deletion and export are written down, control is not yours.

Why should acceptance testing be in the contract?

Because in many legal systems custom development counts as a work contract, and if you do not inspect and report defects within a reasonable period after delivery, the software may be deemed accepted. Write down the acceptance period and criteria; your test set serves here as well.

Can you really negotiate the liability cap?

Yes, and it is routine. Twelve months' fees is reasonable as the general cap; carve-outs for data breaches, regulatory fines and intellectual property are a standard request. Vendors refuse in round one and usually accept in round three.

Local law governs my vendor contract. What about the provider's own terms?

The model provider's terms of use stay in force down the chain and usually fall under US or Irish law. Ask whether the vendor passes those terms through to you and whether it accepts liability for its sub-processors' acts.

So what should you actually do?

• Scan the draft against the eight clauses: mark each as present, partial or missing. Seven missing is normal; the contract starts from there.
• Build the test set before signing: 100 to 200 real questions with correct answers. Both the SLA and acceptance testing rest on it.
• Get the chain documented: model provider, cloud, database; which plan, which country. A gap here hides a transfer-filing obligation.
• Send no personal data if you can avoid it: masking is the cheapest compliance measure, about a week's work for the vendor.
• Take this list to your lawyer: eight clauses cut the review to half a day, and the bill with it.
• Win the meeting before the contract: most of these clauses grow out of questions asked in the pitch; our ten-minute AI literacy prep for executives lists them.

A contract is not written for the days when the relationship is going well. When the first wrong answer, the first model switch and the first price rise arrive, the page protecting you is the page you add today. If you have a draft in hand and want to see which of the eight are missing, send us the AI-related sections; we will help you sharpen the questions before they reach your lawyer.