AI Vendor Contract Checklist: 10 Clauses Companies Should Review Before Buying Generative AI

What should a company review in an AI vendor contract?

Review the approved use case, data and model-training rights, confidentiality, security, IP ownership, output restrictions, compliance duties, warranties, indemnities, liability, product changes, audit evidence, termination, and data deletion.

An AI contract should be reviewed against a documented use case, not as a generic software purchase. A tool that summarizes public marketing copy presents different issues from one that processes source code, customer records, regulated information, or material business decisions.

Start with a short risk map: who will use the system, what data will enter it, what decisions or content will rely on its output, which third parties may be affected, and how easily the company can stop using it. Those facts determine which contract terms deserve the most attention.

1. Does the contract match the company’s approved AI use case?

The agreement should identify permitted users and purposes, prohibited inputs and outputs, usage limits, required controls, and any high-risk activities the service is not designed to support.

• Confirm the product, model, features, integrations, and environments included in the order.
• Identify whether employees, contractors, customers, or automated systems may use it.
• Compare acceptable-use restrictions with the intended workflow and industry obligations.
• Put material sales claims about privacy, accuracy, security, or model isolation into enforceable documents.

2. Can the vendor use company data to train or improve models?

The contract should distinguish prompts, uploaded files, outputs, account data, telemetry, feedback, and derived data, then state exactly how each category may be used, retained, shared, and deleted.

Do not rely on a single statement that the vendor does not train on customer data. Review whether the restriction covers subcontractors, affiliated products, abuse monitoring, human review, fine-tuning, de-identified information, and future product development. If an opt-out or technical setting is required, assign responsibility for enabling and maintaining it.

The agreement should also address where data is processed, how long it remains in logs or backups, and whether the customer can retrieve a complete record of prompts and outputs when that record is operationally or legally important.

3. Are confidentiality and security obligations specific to the AI workflow?

Confidentiality and security terms should cover the actual inputs and outputs, limit access, require appropriate safeguards, control subprocessors, and establish incident-notice and cooperation duties.

• Match security requirements to the sensitivity of data and the consequences of disclosure or alteration.
• Ask for relevant independent assessments, testing summaries, access controls, encryption practices, and continuity plans.
• Define security-incident notice, investigation support, remediation, and responsibility for customer communications.
• Confirm the contract, security exhibit, data-processing addendum, and product documentation do not conflict.

4. Who owns prompts, outputs, fine-tuned models, and custom work?

The parties should separately address ownership and licenses for customer inputs, generated outputs, vendor technology, fine-tuned components, configurations, feedback, and implementation deliverables.

An output-ownership clause does not resolve every IP issue. The vendor may need a limited license to process inputs, and the customer may need rights to use configurations, embeddings, or custom integrations after termination. The contract should also explain whether outputs may be similar to material generated for other users and whether the vendor makes any non-infringement commitment.

If employees or contractors will contribute confidential code, designs, or content, confirm that internal policies and upstream agreements allow that use. Contract rights from the vendor cannot cure a disclosure the company was not authorized to make.

5. How does the contract allocate responsibility for inaccurate or harmful outputs?

The agreement should state the service’s intended level of reliability, required human oversight, customer verification duties, restricted decisions, available safeguards, and remedies when the product materially fails its commitments.

NIST’s Generative AI Profile identifies risks such as confident false content, data privacy problems, information-security concerns, harmful bias, intellectual-property issues, and overreliance. Contract review should translate the risks relevant to the company’s use case into operational controls rather than assume a disclaimer resolves them.

For consequential uses, identify who reviews outputs, what source or audit information is available, how errors are reported, and whether the service can be paused. Avoid commitments to customers that exceed the vendor’s own contractual obligations and product capabilities.

6. Which laws, policies, and customer commitments apply to the use case?

The contract should allocate responsibility for legal compliance, required notices and consents, records, risk assessments, accessibility, sector rules, and assistance with customer or regulator inquiries.

• Inventory the jurisdictions and industries affected by the deployment.
• Confirm whether the vendor supports required documentation, testing, disclosures, or data-subject requests.
• Address changes in law and who pays for required product or process updates.
• Align vendor promises with the company’s privacy notices, employment practices, customer agreements, and AI policy.

7. Do warranties and indemnities address the risks the vendor controls?

Review warranties about authority, service conformity, security, legal compliance, malicious code, and rights in training or service materials, along with indemnities for relevant third-party claims.

AI indemnities vary widely and often contain conditions, exclusions, or required product settings. Examine what claims are covered, whether output-related claims are included, who controls the defense, what remedies the vendor may choose, and whether the protection disappears if the customer modifies an output or combines it with other material.

8. Is liability proportionate to the data and business exposure?

Model the general cap, higher caps, excluded damages, carve-outs, claim period, and available insurance against realistic failures in the proposed AI workflow.

• Calculate the cap using expected fees rather than reading the formula in the abstract.
• Check treatment of confidentiality, security, data use, IP claims, legal violations, and willful misconduct.
• Compare the vendor’s exposure with liabilities the company may owe customers or affected individuals.
• Confirm insurance requirements are relevant and supported by evidence when appropriate.

9. Can the vendor change the model, policies, or risk profile without notice?

The contract should define which online terms may change, require notice of material changes, preserve evidence of prior versions, and provide meaningful rights when a change harms the approved use case.

AI services can change through model substitutions, new subprocessors, revised retention practices, new safety filters, or feature deprecation. Require a practical notice channel and enough time to assess material changes. For important workflows, consider regression testing, documentation access, and a termination right if a change materially reduces security, functionality, or compliance support.

10. Can the company exit, export records, and verify deletion?

Exit terms should address prompt and output export, custom configurations, transition help, continued access, model portability where relevant, retention schedules, deletion, and surviving obligations.

A usable exit plan is especially important when the service becomes embedded in customer-facing products or internal knowledge systems. Determine what the company must preserve for legal, security, or audit purposes and what the vendor must delete. Assign an owner to calendar renewal and non-renewal deadlines when the agreement is signed.

How should legal, security, and business teams review an AI vendor together?

Use one shared decision record that identifies the use case, data, affected people, required controls, contract gaps, risk owner, launch conditions, and a date for reassessment.

Legal review cannot substitute for technical testing, and a security questionnaire cannot decide whether a use is appropriate. A short cross-functional review aligns procurement, product, privacy, security, HR, and business owners around the same facts. Revisit the decision when the model, data, audience, or purpose materially changes.