A bank wants to deploy an always-on AI agent for its procurement team. Design the rollout.

Strong answer

Define the job narrowly first: e.g. process purchase requests under a threshold against policy, chase missing quotes, prepare POs for approval. Give the agent its own non-human identity with least-privilege access to the procurement system, read-only to the vendor master, and no payment rights. Write a tiered approval policy: reading and drafting allowed, external emails and PO creation require a named approver, anything touching vendor bank details is never automated. Stream every action with the agent identity, approver and outcome to the bank's SIEM. Build an eval set from a few hundred historical requests and require it to match human outcomes before moving from shadow mode to supervised mode. Run a 90-day pilot with an owner in procurement and a platform-team owner. Expect model risk management and internal audit to review it, so the evidence pack (policy, eval results, logs) is a deliverable.

Weak-answer indicators

• Agent uses a human's account
• No approval tiers
• No eval before go-live
• Forgets audit and model risk review