When a vendor enables an AI feature by default, check five things within a week: what data the feature reads, where that data goes, whether it is used for training, which admin controls exist, and what your contract says. Then decide to keep it, limit it or turn it off. Do not leave the default in place by omission.
Step by step
- Confirm it is on. Check the admin console and the release notes. Note the date it changed.
- Identify the data. Which fields, files or messages can the feature read? Does it respect existing permissions or work from a broader index?
- Follow the data out. Which model provider processes it, in which region, and for how long is it retained?
- Check training use. Does the vendor use your content to improve models? Is there an opt out, and is it on by default?
- Find the controls. Can admins disable it for everyone, for groups, or per feature? Can they see usage?
- Read the contract. Does the data processing agreement cover the new processing, or has the vendor updated its terms? Look for changes to subprocessors.
- Decide and record. Keep, limit or disable, with the owner and date.
A quick triage table
| Finding | Suggested response |
|---|---|
| Reads only public or non-sensitive data | Keep, note in register |
| Reads sensitive data, respects permissions, no training use, controls available | Limit to a pilot group |
| Reads sensitive data, training use unclear | Disable until answered in writing |
| Broader access than the user has | Disable and raise with the vendor |
| No admin control at all | Escalate; consider the vendor's other options |
The permission question matters most
A feature that answers using documents a user could not otherwise open is a data exposure, even if everyone means well. Test it. Ask the feature a question that only a restricted document could answer, using an account without access.
Ask the vendor in writing
Short questions get short answers. Send: what data does the feature process, which providers are subprocessors, is customer content used for training, how is it retained, and how do we disable it. Keep the reply with your records.
Tell your own people
Staff will find the feature before you finish reviewing it. Send a note saying what it is, what to avoid putting into it, and who to ask. Silence leads to guesswork.
Prevent the next surprise
Subscribe to vendor change notices where you can, and add a standing item to your vendor reviews: new AI features since last review. Add the feature and its model provider to your AI inventory. Add a line about this to your procurement questions so future contracts require notice before such changes.
Quick check
Can we disable a feature the whole company already uses?
Yes, if the risk warrants it. Give notice, offer an alternative where possible and set a date.
Who owns this decision?
The business owner of the vendor relationship decides with security and privacy input.
Safeguard's asset discovery and compliance evidence automation can help keep the register and review records, and the free scan is available for your own code.
Last reviewed: 19 October 2026.
