The right level of oversight depends on what the AI use does, who it affects and what could happen if it is wrong. A drafting assistant usually needs a different review process from a system that influences a decision about a customer, employee or service recipient.
Start with the decision
Write down the business decision or action that the AI influences. Then decide whether a person must review the output before it is used, whether they can change or reject it, and what information they need to make that judgement.
Make review workable
A named reviewer is not enough if they have no time, authority or practical way to challenge an output. Give them clear prompts, access to the relevant information and a path for escalating unusual results.
Keep a record when it matters
For uses with greater impact, record the review arrangement, the person responsible and any limits on automated action. Revisit it when the use expands, the information changes or an incident shows the controls are not working as expected.
Questions worth asking
- What can the AI do without a person checking it?
- Who can stop or correct the outcome?
- What information does the reviewer need?
- How will the organisation know when the process needs review?