Find the AI already entering your business
Ask teams about tools, personal accounts and AI inside familiar applications. Keep discussion supportive so unapproved use can be disclosed. Record purpose, owner and information involved. A ban without a usable alternative may drive activity out of sight rather than stop it.
Prioritise tools handling client records, employee information, sensitive material or business-system access. A simple register helps consistent decisions. Review features before enabling them; an existing supplier relationship does not automatically cover every new AI use. Identify who can activate functions in the tenant.
Focus on realistic routes to harm
Common concerns include disclosure, incorrect output reaching clients and excessive connector access. AI-assisted impersonation can make fraud more convincing. A familiar voice is not proof of identity; retain independent callback and payment verification rather than asking staff to detect every synthetic message.
Ask what happens if an answer is wrong, a file exposed or a record changed unexpectedly. Use your operational consequence to choose safeguards before building an exhaustive threat catalogue. Include reliance on a single provider and the fallback if a service becomes unavailable or its terms change.
Set a few enforceable rules
Define approved tools, permitted information and uses needing review. Check retention, training, account security and terms for the actual subscription. Name a contact for questions. A short practical rule people can follow is better than a long policy that leaves common tasks unanswered.
Restrict connector permissions and keep consequential actions behind approval. Verify important outputs. Prompts and awareness cannot replace access controls or independent payment checks. Make the tool's permitted actions explicit, particularly where it can alter records, communicate externally or access customer information.
Assign ownership and plan for incidents
Name someone to maintain the register, review changes and coordinate expertise. They need not solve every technical or legal issue alone. Seek appropriate advice for sensitive uses and UK GDPR questions; this is not legal advice. Define who can refuse or pause a proposed use.
Create a short response for accidental uploads, suspicious instructions and unintended actions. Identify who disables access and contacts suppliers. Thank people who raise concerns and review arrangements as use grows. Compass is available now; no specific control capability should be inferred from this guide.