Your goal: show that you can ship an AI feature responsibly: say why AI fits, define what
good means, test it before launch, watch it after, and decide where a human stays involved.
How this works (about 30 minutes)
- One question, gone deep: you're a PM at Google, and your support agent launches with a
made-up customer. Ben answers questions about the case; he won't volunteer facts.
- Think out loud. Ask for what you need.
- A notepad is available. Ask Ben to open it if you want to sketch, and tell him when you've
written something.
- Lots of follow-ups and one curveball, then a trade-off turn: one number improves while
another falls. Widen, hold or change?
- No hints and no scores during the call. Your report comes after.
What scores well
- Saying which tickets an agent fits and which it doesn't, and why
- Defining success in terms of the customer's problem, not the model's output
- An offline test you could actually build, and a live test that checks it
- Metrics with guardrails: quality, business value and safety, not just a score
- Naming how the agent fails, what you do about each, and where a human stays in the loop
- A clear call when the evidence disagrees, and what would change your mind
- Your plan said out loud, and answers that land on a point
Tips
- Take PM Coach · AI Product Sense first.
- Say what you would decide and what you would measure; avoid listing tools.