Interviewers will ask what you would let an agent send without you, how you check its drafts, and what you do when it gets something confidential wrong. Good answers rest on one real example each: what was drafted, what you checked, what you decided. This template gives you seven likely questions and sample answers.
What are interviewers trying to find out?
For administrative roles that now sit beside agents, the question is simple: will this person catch the mistake before it reaches a client? Agents in professional services draft letters, book travel, schedule meetings, summarise threads and prepare expense claims. Most drafts are fine. The ones that are not, such as a letter with the wrong client's details, a meeting booked over a court date or an expense claim outside policy, need a person who notices.
Interviewers also want to know that you understand the difference between an AI writing tool and a governed agent. A writing tool produces text. A governed agent acts within limits, holds what it may not do, and records every decision. If you can explain that in a sentence, you stand out.
How do you build each answer?
Keep each answer to about a minute, using this order:
- What happened, in one line.
- What was drafted or proposed, and by whom or what.
- What you checked, and what you decided.
- What changed afterwards.
If your example comes from manual work rather than an agent, that is fine; say so and draw the link.
The question template
| # | Likely question | What it tests |
|---|---|---|
| 1 | What would you let an agent send without checking? | Understanding of limits |
| 2 | How do you check a document someone else drafted? | Review method |
| 3 | Tell me about an error you caught before it went out. | Judgement and attention |
| 4 | What would you do if an agent included confidential information in a client email? | Confidentiality and escalation |
| 5 | How would you handle a partner who wants everything sent faster? | Holding a line under pressure |
| 6 | What is the difference between an AI writing tool and an AI agent? | Basic understanding |
| 7 | How are you preparing for this kind of work? | Learning and evidence |
Sample answers
Question 1: What would you let an agent send without checking? "Very little at first. Internal meeting confirmations and room bookings, perhaps, once I had seen it get them right for a while. Anything to a client, anything about fees, and anything with an attachment I would check every time. I would want the limits written down, so everyone knows what is held for a person."
Question 3: Tell me about an error you caught. "A partner's assistant drafted a proposal from an earlier one. The client name was updated on the cover, but the fee table still showed the previous client's rates. I caught it on a final read before sending, corrected it and told the partner. Since then I check every table and footer on reused documents, not just the cover."
Question 4: Confidential information in a client email. "I would stop it, and if it had already gone, I would tell the partner straight away and follow the firm's data breach procedure. Then I would find out why it happened. If the agent drew on the wrong matter file, that is a source problem the person who runs the agent needs to fix, or it will happen again."
Question 6: Tool versus agent. "A writing tool gives me text and I decide what to do with it. An agent can act, like sending a booking or filing a document, but only within limits it has been given. Anything outside those limits is held for a person, and every decision it makes is recorded."
A worked example
An invented assistant, Hana, adapts question 5 to her own experience:
"One partner liked letters out the same day. When an assistant tool started drafting them, he asked me to stop checking routine ones. I agreed a short list with him: routine acknowledgements could go after a quick scan, but anything with fees, dates or a new contact I would read in full. A week later a 'routine' letter quoted the wrong hearing date. After that he was happy with the list."
The answer works because it shows Hana negotiating a limit, not refusing outright, and the example proves why the limit mattered.
What do interviewers listen for in all of these?
Three things: that you protect the client, that you know where your authority ends, and that you fix causes rather than just symptoms. If an answer of yours does not show at least one of those, find a better example.
Avoid saying an agent "will never make mistakes" or that you "trust the AI". Both suggest you would stop checking. Avoid the opposite too: refusing to let an agent do anything sounds unworkable. The right position is clear limits, held decisions and a record.
Questions people ask
What if I have never used an AI agent at work?
Use examples from checking colleagues' drafts and explain how the same checks apply. A short introductory course in governed agents gives you the vocabulary for question 6.
How do these differ from answers for virtual assistants or paralegals?
Virtual assistants face more questions about multiple clients; paralegals about sources and legal documents. See interview answers for virtual assistants and for paralegals.
Where can I read about roles and preparation?
AI operations careers for administrative assistants in the UK and in the UAE cover the markets, and the AI skills for job seekers guide covers what employers ask you to prove.