“We don't train on your data” doesn't tell you when an AI supplier deletes your uploads. It doesn't explain security logs, support tickets or backups either.
Before buying, ask the supplier to follow one document from upload to deletion. Get answers for the subscription and connections you'll use. If an important answer is missing, limit what goes into the service while you resolve it.
Send a question the supplier can answer precisely
Here is a fictional procurement request you can adapt:
We are considering your business subscription to draft replies from our delivery guide. We would upload the guide and enter customer enquiries. For each copy created, please identify its purpose, location, retention period, deletion process and any other company receiving it. Please distinguish model training from operating the service.
Name the product, plan, region and optional connections in your real request. A supplier's answer about its enterprise edition doesn't cover a free trial automatically. If a reseller is involved, establish which company receives the information and whose terms govern it.
Use one table to expose the gaps
Ask the supplier to complete this suggested table. These are questions, not assumptions about what every product keeps.
| Possible record | What to establish |
|---|---|
| Uploaded file and extracted text | Are both retained, and does deleting the file remove the text? |
| Prompts and answers | When does retention start, and can an administrator shorten it? |
| Search index | Is a separate record created, and when does it stop returning deleted content? |
| Security and support records | What may be inspected, by whom and for how long? |
| Backups and subcontractor copies | When do they expire, and what happens after a restore? |
“Thirty days” is incomplete without a starting event. Thirty days from upload, last use or account closure can mean different things. Ask what happens when a user leaves, when a trial expires and when your business requests deletion.
The ICO's storage limitation guidance doesn't give AI chats a universal retention period. Personal data should be kept for a justified purpose and appropriate period. The ICO page is under review following UK legislation changes. Check its current status and other relevant retention duties before removing business records.
Keep “no training” in its own column
Ask separately whether inputs, outputs or feedback can be used for model improvement, evaluation or other product development. Find out whether this is automatic, optional or triggered by a feedback button.
A no-training assurance may coexist with logging, abuse investigation or support access. You need to assess those uses too. The NCSC's public language model guidance explains why sensitive information in prompts needs care. Use fictional material until the arrangement is suitable for the information involved.
Where the supplier acts as your personal data processor, the ICO's contract guidance describes required terms, including return or deletion at the end of the service. That guidance is also under review. Establish the roles for each activity. A supplier's separate use for its own purposes may require a different assessment.
Try deletion, then decide what the answer permits
In an approved test workspace, upload a fictional delivery guide containing a distinctive harmless phrase. Delete it using the ordinary controls. Check new searches, old conversations, exports and any connected application for that phrase.
This can expose a visible problem. It can't prove that every backend or backup copy has disappeared, so retain the supplier's written explanation as well.
An imaginary furniture retailer might reach these decisions:
- Public delivery guide: proceed if the terms and retention fit this use.
- Customer complaints: wait while support access and deletion remain unclear.
- Passwords or access keys: keep them out of the drafting task entirely.
Record the permitted information and unresolved answers in the existing AI inventory. That makes the procurement result usable: staff can tell which material is cleared today, rather than interpreting a vague promise that the tool is safe.