What an AI Assistant Can See of Your Customers’ Feedback

The line, and exactly where it runs

A survey report is made of two different things. There are the numbers: how many people answered, what each of them chose, how the choices spread across a scale. And there are the words: what somebody typed into the open question at the end, in their own sentence, about a visit they had.

An assistant is given the first and never the second. It can tell you that question three has forty-one responses averaging 3.2 with most of the weight at the bottom of the scale. On the same question it can tell you that nine people also wrote something. It cannot tell you what any of the nine wrote, and there is no scope, key or permission that changes that — the text is removed on its way out, so there is no version of the request that returns it.

• Returned: response counts, what each person chose, the spread per question, the trend
• Returned: how many people wrote a comment on each question
• Never returned: the comment itself, to any key, connector or assistant

The question, not the average

The useful thing an assistant does with feedback is the thing nobody has time for: reading every question separately instead of the one number at the top.

A place with a score of 4.1 is usually not uniformly 4.1. It is a five on welcome, a four on the food and a two on how long the bill took, and the average hides all three. Ask an assistant to compare the distribution question by question over the last month and it will name the one that is dragging and the one that is carrying, which is an answer you can act on this week. That is arithmetic somebody could do by hand, and precisely for that reason it never gets done.

It reads the same numbers your dashboard shows, so there is nothing to reconcile afterwards. It is the reading that was missing, not a second source of truth.

• Compare the spread per question, not one average per place
• Name the question that moved and the period it moved in
• Same numbers as the dashboard — a summary, not a second source

Why the sentences stay behind

Somebody who writes two lines about a bad evening is writing to the business they had it in. They are not publishing, and they are not addressing whatever systems that business has connected this quarter. Handing that sentence to an assistant makes it a third party to a conversation it was never part of.

It also goes somewhere. A reply sent to a credential lands wherever the holder of that credential keeps things, which is a question about somebody else’s infrastructure rather than about yours. The customer cannot see that decision, cannot object to it, and would have no way to know it happened. The honest way to hold text written by another person is to keep the number of places it can travel to small, and an integration is exactly the kind of door that makes that number grow quietly.

So the seam is not a caution flag on a tool description, which a model can reason its way around. The text is dropped before the response is built. A limit that depends on an assistant behaving well is not a limit.

What that costs you, said plainly

It costs you the feature everybody demonstrates first: paste in six hundred comments, get back five themes. That is a genuinely useful thing and you do not get it from an assistant connected this way.

What you get instead is a decision that stays yours. Comments can be exported from the inbox, and if you choose to put them through a tool you have chosen, you do that knowingly, once, for a set of answers you looked at — rather than leaving a standing connection through which everything anybody writes, from now on, arrives somewhere by default. The first is a choice; the second is a habit nobody remembers forming.

Judge the trade honestly. If reading the comments is the whole job, read them, or export them deliberately. If the job is knowing which question is slipping across nine places since the menu changed, the numbers were always the right input and the comments were never going to answer it.

• No theme extraction or sentiment over comments through the connection
• Export from the inbox when you want that, as a decision you make once
• A standing connection is not the same as a considered export

What anonymous survives

A survey that collects no contact details stays that way through an integration, because the integration cannot reach further than the survey does. Nothing about connecting an assistant re-identifies an answer that was anonymous when it arrived.

Where a survey does collect a contact detail, the same rule applies as everywhere else in the workspace: the assistant sees what its scopes allow, the connection is listed where you can see it, and one press ends it. The useful discipline is the ordinary one — give a first key read-only scopes, watch what a week of questions actually looks like, and widen it only once you would not be surprised by anything it did.

Before you connect anything to customer data

These questions are worth asking of any feedback product, and the answers are more telling than a feature list.

Does the text people wrote leave the product through integrations at all, or only the numbers? Is that boundary enforced where the response is built, or is it a note in the documentation? Can you list what is connected right now and cut any of it off in one press? And when you ask the product what it will not do, does it tell you plainly — or do you find out halfway through a task, from a refusal?

• Does written feedback leave through integrations, or only the numbers?
• Is the limit in the code, or a note in the documentation?
• Can you see every connection and end one immediately?