Key takeaway
Start with an inventory and named owners. Link evidence of access and action boundaries to responsible owners, and test approval gates and control operation.
Why AI governance matters
Some AI assistants and agents can act on systems as well as read data, depending on their granted permissions. During a security questionnaire or audit, teams need to explain the tool’s purpose, actual access and accountability for decisions. Extend existing access, supplier and incident processes to cover AI use.
1. Maintain a current inventory
Keep one maintained list of assistants, agents, model APIs, plugins and automations across departments. Record purpose, users, technical owner, supplier and pilot or production status. Update it when an integration is added rather than rebuilding it immediately before a review.
2. Document data permissions
Document which files, mailboxes, repositories and applications each agent can read, change or send to another system. Include data sources and sensitivity. Reconcile this record with actual permissions whenever configurations change.
3. Apply role-based least privilege
Give the agent role only the permissions its task requires. Separate service identities from human accounts, and authorize specific resources and actions instead of broad administration. Set expiry for temporary access and remove permissions when the task ends.
4. Keep an action audit trail
Keep searchable records of what the agent did, when and on whose behalf. Link actions to approvals, outcomes and errors so an investigator can reconstruct events. Define retention, access to the logs and protection for sensitive content.
5. Require human approval
Require human sign-off before sensitive or difficult-to-reverse actions such as deleting data, moving funds or changing access rights. Show the approver the purpose and affected resources. Test that unapproved actions stop at the gate.
6. Isolate teams and clients
Prevent an agent on a shared platform from using one department’s or client’s data in another’s work. Define boundaries for accounts, knowledge stores and integrations. Test access across environments, including both reading data and taking actions.
7. Approve new tools through a written policy
Before adoption, require the team to submit the purpose, data, integrations and owner. Define reviews by security, the data owner and other relevant functions. Reassess approvals when the model, use case or integration changes, using a short policy employees can follow.
8. Prepare an AI incident plan
Prepare for data leakage through prompts and unintended agent actions. Name an authorized incident owner and escalation path, including proportionate stopping of the agent and revocation of access. Preserve evidence and establish the cause before restoring the workflow.
9. Review access regularly
Set a review cadence based on risk. Owners reconfirm needed permissions and remove unused integrations or inactive identities. Review role and team changes promptly rather than waiting for the next cycle, and retain the decision and completion date.
10. Train employees on acceptable use
Teach employees which tools are approved for their role, which data they may enter and when to check an output. Use practical examples and provide a clear way to report unapproved tools or unexpected actions. Refresh training when the permitted workflow changes.
11. Assess supplier risk
Review how the supplier of an AI service or embedded AI feature processes, retains and reuses data. Assess contractual terms, access, integrations and incident notification. Feed supplier changes into the tool approval process instead of treating the initial review as permanent.
12. Name a governance owner
Name a person or team accountable for the whole program. Define IT, security and compliance responsibilities, decision authority and reporting. Assign an owner and due date to open gaps, and bring overdue actions and unresolved risk decisions to management.
Close gaps incrementally
Complete the inventory first, then apply access controls, approvals and logging to agents handling sensitive data or consequential actions. Reuse supplier reviews and incident processes. When selecting a platform, demonstrate role scope, centralized visibility, isolation and approval gates with your own use case, then maintain inventory, training and permission reviews on a schedule.