Rating answers and data controls

Rate the assistant's answers, optionally share a conversation with Alchex, and govern all of it with one workspace switch that starts off.

Every answer the assistant gives can be rated, and a conversation that went wrong can — if your workspace allows it — be shared with the Alchex team. This page explains both, and exactly what leaves your workspace in each case: for the feedback signal, no content at all; for a share, only what you explicitly agree to send.

Telling the assistant an answer was wrong

There is nothing to click. Say it — "that's wrong", "that's not what I asked", "you misunderstood" — and Alchex records that the reply missed. There are no thumbs to hover for and no control to find: you were going to type that sentence anyway.

It only hears you complaining about the answer. Correcting the subject — "no, use the other supplier", "change it to quarterly" — is ordinary work, not a complaint, and is never recorded as one. Nor is talking about things going wrong in your own organisation: "if a change goes wrong we revert it" is your procedure, not a verdict on the reply.

What gets recorded is a content-free signal: which answer and that it missed — never what was said, and never which records the answer cited (see References for how citations render and where their chips lead). It is what tells Alchex where the assistant underperforms without exposing what you were talking about.

There is no "that was good" to record, because nothing would read it. Say it if you like — the assistant is listening — but only the misses are kept.

What happens after you say an answer was wrong

The signal is saved. Then, only if your workspace admin has turned on conversation sharing, a small offer appears under the message: Share this conversation with Alchex to improve the product, next to a Dismiss button.

  • The offer is just an offer. Nothing opens on its own, and dismissing it leaves only the saved rating.
  • If sharing is off for your workspace — the default — no offer appears at all.

The steps behind an answer

While the assistant works you see what it is doing, step by step — reading a document, checking the register, looking up a clause — each one ticking off as it finishes. When the turn ends those steps fold into one line: how long it took and how many steps it ran. Open that line any time, including after reopening the chat weeks later, to read the steps again.

The full detail of the run is kept with the answer and is not shown on screen. That is what travels with a shared conversation, and it is what lets the Alchex team reproduce a wrong answer rather than guess at it — see the next section for exactly what a share sends.

Sharing a conversation

Choosing the offer opens a dialog that lists, before you confirm, exactly what sharing sends:

  • the messages in this conversation so far;
  • what the assistant did to answer — the tools it used, what it looked up, and what came back (the full record of the run kept with the answer, in more detail than the steps shown on screen);
  • the note you add in the dialog (optional).

Nothing else is shared — no other conversations, no documents, no records. Share and No thanks are equal choices, and closing the dialog any other way sends nothing. Once shared, the copy the Alchex team receives expires automatically after 90 days.

If the admin switches sharing off between the offer appearing and you confirming, the share is refused and the dialog says so plainly — nothing is sent. Sharing covers the conversation only: a document the assistant edited during it is never sent with it, whatever state that document is in.

The workspace switch

Sharing is governed by one switch in Settings → General → Data controls, visible to workspace admins: Allow members to share individual conversations with Alchex to improve the product.

  • It is off by default for every workspace. Nobody is opted in.
  • While it is off, members never see the share offer, and the server refuses any share request outright.
  • Turning it off later stops all future sharing immediately; it is enforced on the server, not just hidden in the interface.

Where your conversations live

Assistant conversations stay in your workspace, and each one is private to its members. A chat you start has one member, you: other members cannot open it, and neither can Alchex. A session run by a document's agent has one member, the document's team, so everyone on that team can open it from Sessions in their Chat rail and nobody else can. Membership is never inferred from a role.

People in the bar under the message box lists who can read it — each person by name, a team by its name — and Add people offers everyone in your workspace who is not in it yet. Once more than one person is in a chat, the control shows their initials and how many they are. A colleague you add sees the chat in their own rail, reads everything in it, and can ask the assistant there themselves. A person can be removed again by anyone in the chat, except the one who started it; a team is the document's, and changes on the document rather than here. You can add someone before you have said anything — opening People on a chat you have not started yet starts it, so the colleague you add is in it from your first message rather than after it.

A conversation reads like a team chat: every message sits under the name of who wrote it — you, a colleague, or Alchex, the assistant, marked Agent — with the time it was sent, and several messages in a row from one person share one name. In a chat with more than one person the thread stays live while it is open: a colleague's question, and the assistant's answer to it, arrive on your screen as they are written, without a reload, under a New line that goes when you next send something. The assistant answers one question at a time in each conversation. While a colleague's question is being answered, the composer says so — Ayşe is asking… — and a message you send meanwhile waits in your queue and goes when their answer lands; there is no Stop on a turn that is not yours.

While a colleague is writing, a line above the message box says so — Ayşe is typing…, two names when two are writing, Several people are typing… beyond that. It appears a few seconds after they start and goes a few seconds after they stop or send; you are never told that you yourself are typing, and a chat with only you in it shows no such line. Edit & resend from here is offered on your own messages only. Resending from a message discards everything after it, a colleague's later messages included, so only the person who sent a message can start over from it; a colleague's message offers copy and nothing else, and the app refuses a rewind from anyone but its sender.

A reply is written into the conversation by the app when the assistant finishes it, not by the tab it was asked from — so it is there, and can be rated, even if the person who asked had moved to another page or closed the tab before it arrived. Only the conversation's members can see it; a turn never writes into anyone else's conversation, whatever asked for it.

Some cards in a conversation show your own records — the card that asks where you keep something lists your registers so you can pick one, each with the same status the library shows: whether anyone has confirmed a record in it. That list is read by the app on your behalf at the moment the card is drawn, from the records your teams can already reach. The assistant is not handed it; it learns which register you meant when you choose one, the same as any other answer you give it.

What Alchex learns from is aggregated and anonymised: counts, categories and trends of where answers fall short, computed across many workspaces, with any pattern specific to a small group of users or organisations suppressed entirely. No one at Alchex reads customer conversations to produce these patterns.

The single exception is the explicit share described above — a deliberate action by the member who owns the conversation, available only in workspaces whose admin has enabled it. If you never share a conversation, no person at Alchex ever sees one.

Records the assistant names in an answer

When the assistant refers to one of your records it links to it, and the link carries the record's name. Two rules decide whether you see that name.

In your own conversation, you see the names of records you could open anyway. Owning a conversation is not the same as having access to everything it mentions: the assistant can answer about a record on a team you are not on. Where that happens the link is replaced by a restricted record — no name, no identifier, and nothing to follow. The sentence around it is left exactly as written, so you can still see what was said.

A public share link never shows a name. A conversation you publish by link is readable by anyone who has the link, with no sign-in, so there is no way to tell what that reader is entitled to. Every record reference in a shared conversation becomes a restricted record, whoever opens it — including you. If a link needs to show what a record is, describe it in the conversation rather than relying on the reference.

A public share link never shows the assistant's working, either. While the assistant answers it keeps its own record of each step it took — the requests it made to your connected systems and what came back. Those entries are for the audit trail, not for reading: they are raw, they can quote system messages, and nobody wrote them as prose. They are left out of the conversation you read in Alchex, and they are left out of a share link too, so publishing a conversation cannot expose them to whoever opens it. What a share carries is what was actually said: your messages and the assistant's replies.

In both cases a hidden record looks identical to one that no longer exists. That is deliberate: telling the two apart would itself say that the record is real.

Related