A confident wrong answer is worse than no answer, because the person acts on it. This page is about the edge of what Amvio knows, what it says when it gets there, and the one thing it is forbidden from promising.
The difference is not the disclaimer. It is that the last one names the limit, refuses the guess, and offers the nearest thing it is actually sure of.
“You’ll find that under Settings → Integrations.” — Said with total confidence about a menu that does not exist, because the sentence pattern is more familiar to the model than your product is.
“I’d recommend checking our help centre for more details on that.” — Technically not false. It is the answer of something measured on ending the conversation.
“I’ve passed this to the team and someone will get back to you shortly.” — Nobody was passed anything. This is the one that turns a failure into a betrayal.
“I don’t know that one — I can’t see anything in this workspace that answers it, and I don’t want to guess. What I can tell you is where the import log lives, if that helps.”
Not knowing something, being told not to touch it, and having no standing to answer are three different situations. They deserve three different sentences.
The lookup ran and came back empty, or what came back does not actually answer the question. Nothing in the workspace covers it, so nothing is said as though something did.
A list you write, in your own words, and it is absolute rather than a preference. Pricing conversations, the security questionnaire, anything about a customer’s own data — whatever you decide it has no business discussing.
Price, contract terms, a roadmap promise. Three subjects where being helpful and being wrong are the same act, and where the person who can actually answer is one of yours.
This is the part that is easy to miss. The damage is not the wrong answer — it is that nobody can tell the wrong ones from the right ones without checking, and checking is the thing they came here to avoid.
Once somebody has been confidently misled once, they check everything afterwards — which costs them more time than never having asked.
A wrong answer does not disappear. It arrives at your support queue a week later, wearing the shape of a bug report about something that was never broken.
If it always answers, the answer rate says nothing. A system that refuses is the only kind whose numbers mean anything.
Which is why the refusal is not a limitation being apologised for. It is the thing that makes everything else on this site checkable.
One decides when it considers itself unsure. The other decides what it does about it. You will want different answers for a trial user and for an enterprise admin.
Answers if it has anything relevant at all. Says it does not know only when completely in the dark.
Answers when reasonably confident, and says plainly where it is unsure rather than smoothing over it.
Answers only when it is certain. In any doubt, it says so rather than risk being wrong.
Makes the attempt, flags the uncertainty clearly, still tries to help.
Says plainly that it does not know this one, and offers whatever nearby thing it is sure of instead.
Does not attempt an answer at all. Says this is outside what it can help with.
Which is why this is a setting rather than a personality. The same edge, met by three people who need three different things from it.
They will forgive a hedge and they will not forgive silence. Answer where there is anything relevant at all, flag the uncertainty, keep them moving — the cost of a soft answer here is low and the cost of a dead end is the trial.
They are about to act on what they are told, at scale, on real data. Answer when confident and say so plainly when not, because a hedge they act on is worse than a question they have to ask twice.
Not a confidence question at all. It hands these to you with the conversation attached, regardless of how sure it is, because being helpful and being wrong are the same act here.
Not “someone will be in touch”, not “I have passed this on”, not “the team will get back to you shortly”. It does not say those things, because saying them is a promise about your people that it is in no position to make.
An announced handoff that never arrives converts “the AI could not help me” into “the AI lied to me” — at the exact moment the customer is already stuck.
What happens instead is on the next section: the question is kept, with the person’s own words, and it reaches your team through the dashboard rather than through a sentence said to a stranger.
The promise is what gets removed. The record is what replaces it — and it reaches your team through the dashboard rather than through a sentence said to a stranger.
Not a summary of the session — the specific question that had no answer. Different wordings of the same one are collected together rather than listed separately.
Their own sentence, word for word, and what kind of trouble it was — confusion, frustration, the same question again, a task given up on, or an answer they pushed back on.
Once enough separate customers have hit it, or one customer has hit it enough times. One person asking once is a curiosity, not a work item.
Which means the questions it could not answer are the most useful output it produces. They are the only list in your company written entirely by the people who needed something and did not get it.
Not a ticket somebody has to triage. An entry on a short list, with the question in the words it was actually asked in.
“Can a scheduled export go to more than one destination?”
across four different customers
and it has not stopped
every time, without inventing one
and then it disappears on its own
Illustrative, in the shape the record takes. The numbers here are the reader’s own — six people, four customers, three weeks — not a measurement of anything we have run.
Everything else on this site is checkable eventually. This is checkable in one question, and if it fails here none of the rest is worth anything.
A map with nobody to ask. A conversation with nothing to draw on. A memory of a product it never learned. The system is the point.
A conversation inside the product, on the screen they are stuck on — spoken, and able to see what they are looking at.
read the page →Where it stops. It says it does not know, refuses the three subjects you never want answered, and promises nothing it cannot keep.
you are hereIt walks your product screen by screen and writes down what it found, in a 22-section record per page — including what it could not establish.
read the page →Your maps and docs compiled into things it knows, held at two grains so a single fact inside a long document can still be found.
read the page →What it handled, what it learned, what it could not answer, and which accounts have gone quiet — for the people who own the product.
read the page →Where you set what it is, how it speaks, what it must never touch, and which pages it appears on. Settings that become instructions.
read the page →Your users are already having the conversation. Right now it is with nobody.