Adrian DunkleyAI Risk Management

Scope and authority

Set AI’s authority before you assess the risk

A risk assessment starts with the decision, the service at stake and the actions the system can take.

The same AI assistant has a different risk profile when it drafts a claims message, sends it to a customer or changes a claim status. Start with what the system can do and who can stop it. Capture exceptions and fallbacks in the same scope note its owner will sign.

Name the service and the action

Ask the service owner for the business objective, the normal workflow, affected people, dependencies and the most consequential action. Draw a line from the source record to the outcome and mark where AI reads, recommends, writes or executes. Record who has authority at each step. If the system can hand work to another service, trace that handoff too. Use the approved versions of the policy, operating procedure and delegation schedule. HM Treasury calls for risk management to draw on “the best available information and expertise”. HM Treasury, The Orange Book

In a claims example, the assistant reads a policy extract and suggests the next document to request. If it can send the request, record the channel, which customers are included and who approves the message. If it can change eligibility or payment status, scope those actions independently. They need different evidence and different decision owners.

In an October 2026 survey by the Bank of England and Financial Conduct Authority, 80% of firms responding to the full questionnaire reported using AI agents in production for decision support. The survey describes that use as AI recommending actions while people decide and execute. Ask your own service owner to demonstrate where that boundary sits. Bank of England and FCA, Artificial intelligence in UK financial services, 2026

Put authority in a register row

ActionAuthority to recordEvidence to ask for
Draft missing-document requestClaims officer checks the draftSigned workflow and sample review log
Send customer messageNamed supervisor approves releasePermission list and approval record
Change claim or payment statusAssigned decision ownerAccess test and system audit trail

Capture the maximum volume a defective action could reach before someone intervenes. Ask the owner to demonstrate a pause and an approved manual route. If a different person controls the underlying system, name that person and check that they can act when the usual owner is away. Engagement with affected users can reveal outcomes the implementation team did not consider. NIST, Generative Artificial Intelligence Profile

Copy this prompt

Use the supplied scope note, workflow and delegation schedule. List each action the AI system can take and whether it reads, recommends, writes a record or commits an action. For each, identify the affected service, permission boundary, decision owner, stop route and source clause. Flag any missing authority or fallback. Return a concise scope note with questions for the service owner. Preserve the documents’ exact limits.

What to take to the owner

Bring the scope note and one real workflow trace with confidential details removed. Ask the owner to point to the record that proves the approval step works. Record the test and the person who can accept open risk. Continue with the policy comparison guide when responsibility sits across several documents.

Sources and references