How would you design an AI assistant that can take actions in company systems, like issuing refunds, so it stays safe?

What the interviewer is looking for

• This question tests ai safety & security. Understands prompt injection, data leakage and unsafe actions, and applies least privilege, human approval and audit logs.
• Strong answers: Least privilege, limits and thresholds, human approval for risky actions, injection defences, audit logs and staged rollout.
• Likely follow-up: What would you monitor after launch to catch abuse?

Another way it may be asked: Support wants the assistant to refund customers automatically. What controls would you put around that?

A structure to follow: Requirements, sketch, deep-dive, trade-offs

1. Step 1

   Requirements

   Ask who uses it, what it must do, and which qualities matter most: speed, reliability, cost, consistency.

2. Step 2

   Sketch

   Draw the main parts and the path one request takes through them before adding detail.

3. Step 3

   Deep-dive

   Pick the riskiest part (storage, queueing, failure handling) and go deeper there.

4. Step 4

   Trade-offs

   Name what you gave up for each choice, and what you would do first if it needed to grow.

This is guidance on shape, not a script. Use your own real experience and facts; never invent details to fill a gap. See the STAR guide and answer frameworks.

Common slips

• Naming tools before establishing requirements.
• Drawing every box at the same depth instead of going deep where it matters.
• Presenting one design as the only answer, with no trade-offs.

How to practise it

Time yourself: two minutes on requirements, five on the sketch, five on one deep-dive, two on trade-offs.

Related questions

• What is prompt injection, and how might it affect an assistant that reads customer emails?
• In simple terms, what is retrieval augmented generation, and why would you use it instead of just asking a language model directly?
• A chatbot you built keeps giving long rambling answers. What would you change first?
• Your document search assistant often answers using the wrong document. How would you debug and improve retrieval?
• How do you evaluate whether a change to a prompt or model actually made your application better?

Browse all questions