How have you adopted AI in your workflows? Walk through examples

[INTERVIEWER] How have you adopted AI in your workflows? Walk through examples. This question sounds like small talk, but it absolutely is not. When they ask how you have adopted AI in your workflows, they aren't asking whether you have a ChatGPT tab open. Everyone has one of those open. What they're really checking is whether you use AI with judgement.

Did you actually check the output before you trusted it? The winning answer names one specific workflow, one specific tool, and one specific number. The losing answer lists five tools and proves nothing. Let me be blunt about what's actually being tested here. This is an AI product manager role. If your whole relationship with artificial intelligence is pasting a prompt and shipping whatever comes back, you're exactly the kind of PM who ships a confidently wrong feature.

They want to see that you treat the output of a model as a hypothesis to verify, rather than an answer to copy. Over the next few minutes I'll give you a scaffold to fill with your own story, followed by a fully worked example and the follow up questions they'll throw at you. The structure is STAR, and the trick is picking one workflow you genuinely changed rather than ten shallow ones.

Depth reads as senior. Start with the Situation for about thirty seconds. Describe the workflow before AI and explain why it hurt. Make it concrete with something like reviewing two hundred support tickets a week to tag themes by hand. You need a real task and a real volume. Next is the Task for twenty seconds. What were you trying to improve and what was the constraint?

Say it out loud. You wanted to cut the time without losing tagging accuracy because those tags fed the roadmap. The constraint matters since it's what stopped you from just automating blindly. Now for the Action. This is the core of the answer, taking about two and a half minutes, and it's where your AI judgement shows. Name the tool and say what you fed it.

Here's the crucial part. You need to explain how you checked it was right before you trusted it. If you built a prompt template, an evaluation script, or a small checker, say so. Name your guardrail out loud by stating that you manually labelled fifty tickets as a golden set and measured the model against that before rolling it out.

That single sentence separates you from the crowd. Finally, the Result takes about a minute. Give the number for time saved, error rate, throughput, or adoption. Then add the crucial honest caveat. Where did it still need a human? If AI solved your problem perfectly with zero limits, they simply won't believe you. Here's how that sounds when it's all joined up.

As a PM on a support tooling team, we tagged inbound tickets by theme by hand. It was about two hundred a week, and it took an analyst most of a day. I wanted that under an hour, but without the tags getting noisy, because we used them to prioritise the roadmap. I built a classification prompt with GPT 4o.

It took the ticket text and returned one of our twelve existing theme tags, plus a confidence score. Before trusting it, I manually labelled fifty tickets as a golden set and ran the prompt against them. The first pass was seventy-eight percent agreement, which honestly wasn't good enough. I added three worked examples per tag to the prompt, along with a rule to return unclear rather than guess.

That pushed it to ninety-one percent. We shipped it with one rule. Anything the model marked low confidence or unclear went to the analyst, and everything else got automatically tagged. That turned out to be about fifteen percent of tickets. The result was that tagging dropped from roughly six hours a week to about forty-five minutes. The analyst spent that time on the hard fifteen percent, which is a good use of a human.

Our theme accuracy, measured against a monthly manual audit, held above ninety percent. The honest limit was that it drifted whenever a new product launched and new themes appeared, so I set a monthly recheck of the golden set. That's the whole answer. One workflow, one tool, one check, two numbers, and a caveat. You measured the model against a golden set before trusting it, rather than waiting until a complaint landed.

That evaluation rigour signal is the entire game. They also notice that you built a confidence threshold routing uncertain cases to a human instead of automating blindly. This shows you designing for the model being wrong. Rounding out their assessment is the fact that you gave a real number on both sides. You provided time saved and accuracy held, alongside the drift caveat proving you are still watching it.

Let's look at how people sink this answer. The biggest mistake is simply listing tools. Listing Copilot, ChatGPT, and Notion AI without a workflow, check, or number sounds like enthusiasm but scores like nothing. Another trap is describing an output you pasted straight through with no verification step, because that tells them exactly how you'd ship a product. A third error is having no honest limit at all.

A perfect result with zero caveats reads as either lucky or completely made up. Expect a follow up question as well. They'll ask what you did with the uncertain fifteen percent, or how you'd know if it started drifting. Have an answer ready. The golden set recheck is your answer to both of those. To pull the whole picture together, pick one workflow you genuinely changed and walk through the STAR method.

Name the tool and the check you ran before you trusted it. Give a number on time and a number on quality, then end with the honest limit. The verification step is not a minor detail, as it is the entire point of the question. If you want a final thought to anchor your preparation, just remember to name one workflow, one tool, one check you ran, and one number.

If you can't explain how you checked it, you didn't really adopt AI. You just used it.