I ran switch interviews for years and they work. Reconstruct the timeline of a customer's decision, find the forces, and you get the real job instead of the feature request they opened with. Bob Moesta built a career on it for good reason. Here is what I could never fix: they are slow, they sample almost nobody, and they ask a human to accurately remember what they were thinking three weeks ago. Two weeks of work to learn from eight people what they thought they wanted. Then we shipped AI features and I realized the interview I kept working to schedule was already happening thousands of times a day, in writing, for free.
What a prompt actually is
When a user types a request into an AI feature, look at what they just handed you: the job, in their own vocabulary; the trigger, because they typed it the second they hit the wall; and the failure, captured the instant your product could not do the thing. That is the exact payload a switch interview tries to reconstruct from memory, except the customer is doing it live while you watch. No recruiting, no 45-minute call, no walk me back to the moment. The moment is in the log.
Where logs win and where they do not
Logs win on what and where. What is the customer trying to do, and where does the product fall down. That is unbeatable in the logs because it is literally the text: the request your product cannot serve, the rephrasing when it fails, the workaround narrated in frustration. Interviews still win on why. A log shows you a user asked the same thing three ways and gave up. It does not show you they gave up because their boss was waiting and they were embarrassed. That is still a conversation. So the model is not logs instead of interviews. It is logs for breadth, interviews for depth, in that order.
The four signals to read first
When I open a prompt dataset I hunt four things, in priority order. Requests the product cannot do yet, which are unmet jobs stated in the customer's own words, no translation needed. Reformulations, because a user who rephrases the same ask three times is not confused, they are blocked. Vocabulary mismatches, where customers consistently name the job differently than your UI does, which is a positioning problem hiding as a discovery one, and the words in the logs belong on your landing page. Rising intents, where a request type grows week over week, which is demand forming months before it reaches a sales call.
To run it you need consent, a de-identification step, and a clustering pass, not a data team. Start with aggregated, de-identified prompts, strip anything personal, and cluster by intent. Then, and only then, book interviews, now surgical: call four people who triggered a specific rising intent to learn the why behind a pattern you already measured. Pick one AI surface in your product this week, pull a de-identified sample of what people typed, and cluster it by intent. The top three clusters usually contain a job you did not know was on your roadmap and a word your marketing should have been using.