1. Record the systems and their owners
Start with an inventory of AI uses, including individual subscriptions, features embedded in business software and internally built services. Record the purpose, users, data categories, supplier and person responsible for decisions. An inventory is not approval; it provides the information needed to decide whether a use is acceptable.
Assign ownership for business outcomes separately from technical administration where appropriate. The person maintaining an integration may not be authorised to decide whether it can handle employee records. Name the person who can stop the system and the person who can approve a change. Avoid responsibilities assigned only to an unspecified “team”.
2. Classify use by consequence
Compare tasks by what happens when the output is wrong. Drafting internal meeting notes differs from producing a customer commitment or influencing a recruitment decision. Consider affected people, reversibility, scale and whether errors can be detected before action. A simple technical design can still create substantial risk through the context in which it is used.
Document prohibited uses and uses requiring additional review. General guidance should not substitute for advice on employment, legal, financial or health decisions. UK data protection requirements may be relevant when personal data is processed. Consult qualified advisers for the specific legal position, particularly where automated processing could affect individuals significantly.
3. Use established references carefully
The NIST AI Risk Management Framework organises risk management around Govern, Map, Measure and Manage. It can help structure questions about context, evidence and responsibility. It is not a product approval or a guarantee that a particular implementation meets UK legal requirements.
The Information Commissioner’s Office provides UK data protection guidance, including material relevant to AI. ISO/IEC 42001 specifies requirements for an AI management system. Referencing a standard does not establish certification. Choose controls that address the actual system and record why they are suitable rather than copying a list without assigning owners.
4. Set data and supplier controls
State which information may enter each approved tool. Separate public material, confidential business information and personal data. Review supplier terms for retention, training use, subprocessors, security commitments and international transfers. Consumer and business offerings can have different contractual arrangements; evaluate the exact service proposed.
Restrict access according to role and review service accounts. Require a documented route for granting and withdrawing permissions. Where a supplier processes personal data on your behalf, address processor obligations in the contract. A data protection impact assessment may be required where processing is likely to result in high risk; obtain appropriate advice on whether that applies.
5. Make review meaningful
Human review is useful only if the reviewer has time, information and authority to challenge the output. Show sources and proposed actions. Explain known limitations and make rejecting a suggestion straightforward. Do not treat a routine click as effective oversight when the person cannot inspect the basis for the answer.
Maintain evaluation cases that reflect the approved task. Reassess after changes to models, prompts, retrieval sources or integrations. Document what changed and why continued use is justified. Monitor reported errors as well as technical failures; a service can be available and still return inappropriate or unsupported responses.
6. Prepare for incidents and withdrawal
Define how staff report an unexpected disclosure, harmful output or unauthorised action. Include a way to pause processing, preserve relevant evidence and involve the right people. Data protection incidents require assessment against applicable notification duties. Avoid retaining every input indefinitely merely to make investigation easier; set a proportionate logging policy.
A governance engagement can be scoped to an acceptable-use policy, system inventory, review responsibilities and incident procedures. Agree deliverables in writing. For an initial enquiry, describe your existing uses and the decisions that need control. Combine this service with AI strategy for a new project or workflow automation when AI is being connected to actions.
