How I built a customer-feedback loop with AI

Northwind has collected customer feedback for years. Collecting it was never the hard part — a quarterly form went out, responses landed in a sheet, and there they sat.

The hard part was what came next: turning a few hundred free-text answers into something the team could actually act on. Someone had to read every response, find the patterns, and bring them to a meeting. That meeting was a whole afternoon, so it kept getting pushed. By the time we finally talked it through, the feedback was a quarter old and half of it no longer mattered.

It is worth being honest about what that means. Collecting feedback you never analyse is just a polite way of ignoring your customers. The data existed; the synthesis never happened. So I rebuilt the loop around the one step that always stalled — and handed that step to Claude.

How the loop works now

The whole thing runs as a simple chain, and the team only shows up for the last step:

1. A quarterly form goes out to every active client.
2. Responses collect automatically in a linked Google Sheet.
3. A script sends the open-text answers to Claude, which clusters them into a handful of named themes.
4. Each theme comes back with a few real customer quotes attached.
5. The themes land in a Roadmap sheet, ready for the quarterly retro.

This works because clustering hundreds of answers into a few themes is exactly the kind of task an LLM is good at and humans find tedious. Nobody on the team reads raw responses any more — they open a one-page summary and spend their time deciding what to do, not sorting through rows. What used to be an afternoon of reading is a summary that is already waiting when the meeting starts. (If you want the clustering script itself, I go through it in Build an AI survey-response analyzer.)

What I let AI do, and what I kept

The split matters more than the code. Claude does the synthesis — the slow, tedious aggregation of hundreds of answers. The team keeps the judgement: deciding which themes are actually worth acting on is still a human call, made together in the retro.

I cut one thing entirely — reading every response by hand. That single change is what collapsed the cycle from a quarter to a week. The meeting still happens; it just starts from a summary instead of a pile.

And I added one thing deliberately: every theme carries real customer quotes. It would have been easy to let the themes drift into tidy abstractions, but a theme labelled “onboarding is confusing” hits very differently when there is an actual sentence from an actual client sitting underneath it. The quotes keep everyone grounded in what people really said, not in my paraphrase of it.

Was it worth it

Plainly, yes:

• The feedback cycle went from a quarter to a week.
• The retro is now a decision meeting, because the sorting is already done.
• Feedback gets acted on while it is still fresh enough to matter.

Most feedback systems do not fail at collection — they fail at synthesis. Hand the clustering to an AI, keep the quotes for honesty, and the loop finally closes. The people who took the time to tell you something actually see it lead somewhere.