Most companies have feedback. A survey after purchase, a rating after a resolved request, an annual study. They collect for years and in the overwhelming majority of cases no change comes out of it.
Not because the feedback is useless. Because between a thousand answers and a decision sits work nobody has time for.
A score names the problem, it does not explain it
Satisfaction fell from 4.4 to 4.1. That is information you cannot act on — until you know why. The reason is always in the free-text answers, and those are exactly what goes unread, because there are too many of them.
This is where automation is most useful: not in the collecting, but in turning three hundred sentences into five themes with counts and examples.
Sort, do not judge
The same line as elsewhere. The agent should place answers into themes (price, delivery speed, communication, quality, user interface), capture tone, and flag what needs a response now — a complaint threatening to leave, a description of a situation likely to repeat, a mention of a system fault.
What it should not do: decide which theme is important. That is a management judgement resting on context the agent does not have.
For every theme it is essential to be able to open the original answers. A summary with no route to the source is just another way of not reading the feedback.
Timing decides the response rate
A question asked at the wrong moment goes unanswered no matter how it is worded.
- Right after a completed experience — after delivery, after a request is resolved. Not after ordering, when the customer has experienced nothing yet.
- Short — one question and a free-text field. Every additional question cuts the response rate noticeably.
- No sign-in — a click straight from the email.
And do not ask the same customer after every interaction. A cap of your own, say once a quarter, protects both the response rate and the relationship.
Responding is part of collecting
This is the most common mistake. The company asks, the customer writes an objection — and nothing happens. Next time they will not bother.
The minimum is acknowledging receipt. Better is automatically creating a task for a named person, with a deadline, on a negative answer. The agent can prepare the case here: what the customer wrote, what their history is, what it concerns.
Every so often it is worth going back with what changed as a result. It is the cheapest way to raise willingness to answer next time.
Do not forget internal feedback
The same mechanism works inwards — reports from operations, suggestions from colleagues, notes from exit interviews. The volume is smaller there but the value often higher, because people in operations see causes rather than only effects.
Watch anonymity, though: if a survey is anonymous, the agent must not group answers in a way that lets the author be inferred from theme and context. Related principles are in the piece on AI agents and GDPR.
Connecting it to reporting
Feedback means something when it can be compared with other numbers — order volume, delivery times, complaint counts. That is when it becomes visible whether a drop in satisfaction relates to a specific change.
So it belongs in the ordinary reporting view, not in a separate tool nobody opens. How to build that is covered in the piece on automating reporting.
What to measure
- Response rate by moment of sending — shows when it is worth asking.
- Share of answers with free text — numbers without text have limited value.
- Time from a negative answer to a response.
- Number of changes that came out of feedback — the only figure that says whether any of it is working.
Where to start
With one point in the process — typically after a resolved support request. Let answers be sorted into themes for a month and compare them with what the team believes. Usually one theme appears that everyone knew about, and one that surprises.
If you would like to work out what can be extracted from your feedback, write to us. More on AI and automation and examples in our case studies.