Customer Feedback After the Purchase: AI Interview Instead of a Questionnaire

Short answer: a rating is a measurement, not an explanation. Four out of five tells you where you stand and nothing about what to change — you cannot tell whether it was the delivery time, the onboarding call, the packaging or the price. That is why so many companies watch a satisfaction score for years without it ever producing a decision: the number was never capable of carrying a reason.

Every established metric has the same hole, and the explainer pages admit it openly. NPS sorts people into promoters and detractors without saying why anyone is a detractor. CSAT records a satisfaction level at a moment. CES measures perceived effort. All three are useful as a trend line and useless as a to-do list. The classic workaround is the optional free-text box at the end — which the people with the strongest opinions skip, because by then they have already spent their patience on the rating grid.

You define the questions and the goal. The AI conducts the conversation, one question at a time, and evaluates each answer as it arrives. When an answer is complete, it moves on. When an answer stays vague — or when a rating comes back low — it asks one specific follow-up before continuing. A two out of five never passes without the question of what would have made it a four.

This is the one setup decision that changes what you learn, and it follows a simple rule: choose anonymity when the feedback concerns a person, and attribution when it concerns a case you want to fix. A customer rating the consultant who advised them will not be candid under their own name. A customer whose delivery arrived damaged wants you to know exactly which delivery it was.

The cost of anonymity is that you cannot close the loop with the individual, and closing the loop is the single most effective retention move there is — a customer whose specific complaint gets fixed and acknowledged usually ends up more loyal than one who never had a problem. A workable compromise for larger rounds: run the interview anonymously, and end with an optional question offering contact details for anyone who wants a personal reply. Most stay anonymous; the ones with a real, fixable problem identify themselves.

One honest note on the market, because the term AI is doing a lot of work in this category: most tools that advertise AI use it to draft your questionnaire or to summarise results afterwards, while the response itself stays a static form. Asking a follow-up during the response is a separate capability that Typeform, Qualtrics and Alchemer also offer — we are not claiming to be alone in it. SurveyMonkey does not, and the German-market classics easyfeedback and LamaPoll are form builders. The useful test when comparing tools: if someone gives two out of five and writes nothing, does anything happen?

Send within 24 hours of the moment you are asking about. Memory of a specific irritation is precise on day one, general by day three and gone by the following week — and a general answer is exactly the kind you cannot act on. Three things then decide how many people finish: length, whether the first question is easy, and whether the invitation tells them how long it takes. Promise two minutes and keep the promise; an interview that quietly runs to eight minutes trains people to ignore the next invitation.

The interview is worth exactly as much as what follows it. Pick the one theme that came up most often, change it, and tell the customers who raised it that you did. That last step is skipped almost everywhere, and it is the one that turns a feedback round into a retention instrument — people forgive a problem that was fixed far more readily than one that was merely recorded. Everything else is a dashboard.