A security and privacy checklist for an AI agent

Separation and data ownership

• Whether your data is separated from the vendor’s other customers, and how.
• Whether data is used to train models - the vendor’s or a third party’s.
• Where data is stored and where backups are held. Both, not one.
• Who at the vendor can access your content, and in what circumstances.
• What happens to content and data at the end of the contract, and in what export format.

Personal data in the conversation

The most underrated point. The risk is not only what is stored, but what the person is invited to type in the first place.

• Whether the system warns against entering ID or payment details.
• Whether identifiers such as a WhatsApp phone number are masked, and exactly where the masking happens.
• How long conversations are retained, and who can view them.
• Whether a specific conversation can be located when a privacy request arrives.
• Whether a written retention policy exists, not only a verbal assurance.

Control

• Audit log: who changed content or permissions, and when.
• Role-based permissions, including separation between departments.
• Whether fixed responses can be predefined for sensitive topics.
• Whether the system can be stopped immediately, and who is authorised to do it.

Evidence worth requiring

The principle: ask for something checkable, not a statement. Note the difference between an ISMS built to a standard and a valid certificate - a vendor who blurs that should be asked to explain it.

• The policy and procedure pack, for review.
• External security test results, with a link that lets you re-run them.
• An explicit statement: ISMS to the standard, or a valid certificate. Do not accept vague wording.
• A description of monitoring, backup and continuity - and what has actually been tested rather than planned.