Design an AI customer support system for a SaaS product: 50,000 tickets a month, target 60 percent automated resolution without wrecking satisfaction.
What they are really testing: Design under a business constraint. The trap is optimising the resolution rate, which is trivially gamed by confident wrong answers; the senior answer designs the escalation boundary and measures resolution quality, not just deflection.
A real interview question
Design an AI customer support system for a SaaS product: 50,000 tickets a month, target 60 percent automated resolution without wrecking satisfaction.
What most people say
drag me
“A RAG chatbot over the help docs handles incoming tickets, escalating to humans when it cannot answer, which should reach 60 percent since most tickets are repetitive.”
It answers questions rather than resolving tickets, half of support is doing things, resets, refunds, account changes, and "escalates when it cannot answer" is exactly backwards: models do not reliably know when they are wrong, so the boundary must be designed, not self-assessed.
The follow-ups they ask next
Which single metric would you watch daily and why?
Reopen rate on automated resolutions, it is the earliest honest signal: deflection can look great while reopens climb, and CSAT lags. A reopen spike in one category pauses automation for that category.
Legal says refunds cannot be fully automated. Does the design survive?
Yes, that is what tiered tool risk is for: the model prepares the refund case, eligibility, amount, evidence, and a human approves in one click. Resolution time still drops even where the final action stays human.
What the interviewer is listening for
- Redefines the target as resolution-that-sticks with paired counter-metrics
- Splits informational from action tickets and designs tools with tiered risk
- Escalation designed with full-context hand-off, never a restart for the customer
What sinks the answer
- Chatbot-over-docs as the whole design, no tools, no account context
- Escalation left to the model self-assessment
- Optimises deflection with no reopen or satisfaction guard
If you genuinely do not know
Say this instead of freezing. Reasoning out loud from what you do know beats silence every single time, and a good interviewer is listening for exactly that.
“Define success as [resolved and stayed resolved, with CSAT and reopen guards]. Then [classify by intent, sentiment, tier], route [how-tos to account-aware RAG with citations], [actions to scoped tiered tools], [high-risk straight to humans], with [designed escalation triggers and full-context hand-off], rolled out [category by category on proven numbers].”
Keep going with system design
Senior
Design a document Q&A system over all internal knowledge, wikis, drives, tickets, for a 10,000-person company.
Senior
Design an AI code-review assistant for a 200-engineer organisation. What does it check, and how do you stop it becoming noise everyone ignores?
Senior
Design the evaluation platform for a company running 12 LLM features. What does it provide, and how do teams use it without a central bottleneck?
Senior
Design an agent that automates invoice processing end to end: receive, extract, match to purchase orders, and schedule payment. Finance signs off on it. What does the design look like?
Senior
Design the model-serving layer for a company with 15 LLM features. One gateway or per-team integrations, and what lives in it?
Senior
Support reports the AI assistant got noticeably worse this week. Nothing was deployed. Walk me through the investigation.
Knowing the answer is not the same as recalling it under pressure
Sign in to send the questions you fumble to spaced recall, so they come back right before you would forget them, and learn the concepts behind them with hands-on labs.
Start free