Expect a retail bank to test whether you can turn a process into bounds an agent can work within: what it may read, draft or do, which cases it must hold for a person, and how you would know it is working. The best answers describe a real requirement you wrote and a decision it changed. A template follows.
Who should use this answer set?
It is for business analysts interviewing for change, operations or digital roles at banks that are putting agents into customer servicing, lending or onboarding. The adverts may say business analyst, process analyst, product analyst or AI change analyst. What matters is the duty list: mapping processes, writing acceptance criteria, defining exceptions and approvals, measuring outcomes.
An analyst on an agent project is asked something new. A traditional requirement says what a system must do. An agent requirement also says how far it may act, per capability, and what evidence will show it is safe to act further. If you can talk about that, you will sound like someone who has done the work rather than read about it. The guide to the AI skills employers will ask you to prove sets out the wider picture.
How should you shape each answer?
Use a three-part shape that suits analysts:
- The process and the decision inside it. Name the point where a person used to decide, such as approving a fee refund above a limit.
- The bound you wrote. State what the agent may do at that point: read, draft for a person to send, or carry out routine cases within a limit. Say which cases you sent to a named person and why.
- The evidence. Say how you tested the requirement against real past cases and what you measured afterwards: how many cases were held, how often reviewers changed the agent's draft, which red lines were never crossed.
Avoid answers that stop at "I gathered requirements from stakeholders". Interviewers hear that often. They rarely hear an analyst explain which bound decided a case.
The template
Copy the list and write two or three lines of your own under each.
- Walk us through how you would scope an agent for card dispute intake. Tests decomposition into capabilities. Answer from: a process you broke into steps with different risk.
- How do you decide what the agent may do on its own? Tests understanding that authority is earned on evidence. Answer from: a rollout you staged, starting narrow.
- What would you never let it do? Tests red lines. Answer from: a control you insisted on, such as no change to a payee based on an emailed instruction.
- How would you test the requirements before go-live? Tests replay against real cases. Answer from: a user acceptance test you designed from historical cases.
- Which measures would you report after launch? Tests operational thinking. Answer from: a dashboard or report you owned.
- A reviewer says the agent's drafts are wrong half the time. What do you do? Tests root-cause analysis. Answer from: a defect you traced to a source, a rule or a training gap.
- How would you explain a held case to a customer complaint team? Tests whether decisions can be reconstructed. Answer from: an audit trail you relied on.
A worked example
Sample answers only. Nadia is an invented analyst at a mid-sized retail bank.
How would you scope an agent for card dispute intake? "I would split it into capabilities. Reading the transaction history and the customer's message is low risk. Drafting the dispute form for a colleague to submit is next. Issuing a temporary credit moves money, so I would keep that with a named person until the agent had a track record on our own cases. In my last project we did exactly that with fee refunds: the agent drafted, a team lead approved anything above 50, and we reviewed the held cases every week."
What would you never let it do? "Act on a change of payee or contact details that arrived by email or from a web form without verification. That information comes from outside the bank, so it can start a case but cannot authorise anything. I wrote that into the requirements as a hard rule, not a preference."
A reviewer says the drafts are wrong half the time. "I would pull a sample of the changed drafts and sort them by cause. On my last project most errors traced to one out-of-date fee schedule the agent was reading. We fixed the source, not the agent, and the change rate dropped back within a fortnight."
How would you explain a held case to complaints? "Every verdict should be filed with the reason and the rule that decided it, so the complaints handler can read why the case waited and who released it. If we cannot reconstruct a decision, the requirement was incomplete."
The fraud and KYC sides of the bank ask sharper questions about evidence, covered in the interview answers for fraud analysts and the interview answers for KYC analysts. If you are applying in Britain, AI operations careers for business analysts in the UK adds the market detail. Pryme Intelligence Academy's free Foundations track covers depth and the decision chain if you want the vocabulary before the interview.
What should you leave out?
Leave out claims you cannot support. Do not say a project cut costs by a figure you cannot explain, and do not name tools the interviewer may not use. Describe the decision, the bound and the evidence, and let them ask for more.
Questions people ask
Do I need to know how models are trained?
Not for this work. Banks want analysts who can define what an agent may do and prove it on cases. Knowing that a business changes an agent's brief, sources and bounds rather than its underlying model is more useful than training theory.
How do I answer if my projects did not involve agents?
Use the same structure for any automation or workflow change you scoped. Say where a person decided, what the system was allowed to do, and how you measured it, then explain how you would apply that to an agent.
Is a certificate worth mentioning?
Mention it if it is current and relevant, briefly. An employer can check it, but the interview is still decided by how clearly you explain your decisions.