The most expensive sentence in an AI deal is the one a salesperson says out loud and the contract never repeats.
A governance professional described a vendor review at an industry conference. The vendor had said, verbally, that it did not use customer data to train its models. The contract said the opposite: the vendor could use customer data to improve its product. Legal and IT had both approved it. It was only caught when the contract reached the governance team.
This is getting more common because AI is being added to tools you already own. A vendor review written before AI existed does not ask the questions that now matter.
This is general information, not legal advice. Have counsel review contract language.
1. Does the vendor learn from our content?
- Is our data used to train or improve any model, the vendor's or a third party's?
- If it is anonymized first, who decides what counts as anonymous, and can we verify it?
- Can we opt out, and is the opt-out written into the contract?
2. Where does our data go?
- Which cloud, which region, and which subcontractors touch it?
- Which AI model provider does the vendor send our content to?
- Are prompts, responses, and chat history stored, and for how long? Some commercial AI services keep history for a period even when they do not train on it.
3. How does the AI get access to our files?
There are three common patterns, and each has different risks:
- As the signed-in user. The AI can see whatever that person can see.
- As its own identity. Cleaner, but may need a separate license.
- Through a connector. Easy to set up, but it needs credentials. Ask whose credentials, and how broad they are.
Ask for least-privilege access and a way to see exactly what the AI can read. Our guide to how AI exposes data nobody labeled explains why this matters.
4. What happens when we delete something?
- When we delete a file, is the searchable copy the AI uses removed too? How quickly?
- When the contract ends, is everything removed, and can we get written confirmation?
- Can the vendor show us how to test that deleted content is really gone?
5. What evidence can we get?
- Can we see logs of what the AI accessed and produced?
- Do answers cite the source documents, so a reviewer can check them?
- How would we reconstruct an answer six months from now?
6. What does the contract say?
- Are the answers above written into the agreement, not only a sales deck or a call?
- Are we notified before AI features are added or changed in a product we already use?
- What are the breach-notification terms for data sent to AI features?
- Do we have the right to audit?
Add AI to your existing vendor review
You do not need a new process. Add an AI section to the questionnaire you already send, and run it again when a current vendor announces an AI feature. Keep a list of who has said what, and file the answers with the contract.
How private deployment changes the answers
Many of these questions get simpler when the AI runs inside your own cloud account. Your content stays in your environment and under your permissions, and you choose the retention. We build that way on Amazon Bedrock, where the content you send is not used to train the underlying models. See the AWS AI assessment for how we evaluate a deployment, and Is your data ready for AI? for the data side.
Illustrations in this guide are drawn from talks at an industry information-governance conference and from general practice. They are anonymized and are not Horus client results.
Part of our guide: Information governance for AI