Short answer: Before using AI with business data, identify what information is involved, who can access it, how the provider handles it, how long it is retained, and what actions the system may take. Get the organisation’s security, privacy, and legal owners involved for sensitive workflows.
A practical review checklist
- Classify the data: public, internal, confidential, personal, or otherwise restricted.
- Send only the fields needed for the task; remove or mask identifiers when possible.
- Review the provider’s terms, retention settings, training use, region, subprocessors, and deletion process.
- Use separate credentials with the minimum permissions required.
- Protect secrets and tokens; do not place them in prompts, logs, or shared documents.
- Decide how outputs and action logs are stored, who can view them, and when they are deleted.
- Restrict actions the model can request; require approval for material or hard-to-reverse changes.
- Test prompt injection and malicious or misleading content in documents where relevant.
- Define incident ownership, access revocation, and a manual fallback.
- Reassess when the model, vendor, data type, or workflow changes.
Fictional example: summarising customer requests
A fictional service team wants to summarise incoming requests for its staff. It begins with a limited sample using masked identifiers, checks the AI provider’s business data terms, and tests whether quoted text in a request can manipulate the summary. The pilot stores a link to the source request rather than copying more personal information into a separate system. This is an example design, not legal or security certification.
An AI output should not be treated as secure merely because it is generated through a business account. Review actual product settings and contractual terms. This guide is general operational information, not legal advice or a substitute for your security assessment.
For process controls, read human-in-the-loop AI and AI accuracy testing. Formula Works scopes AI-assisted operations around the task, data access, review, and exception requirements.
Written by
Formula Works
Published
Sat Nov 14 2026