Training mode
In training mode the assistant answers every customer question as usual — but its reply is seen by a person first. That person sends it as is, corrects it, or answers themselves, and can note what was wrong. The assistant learns from every correction: next time a similar question comes in, it uses it as an example.
It's a way to let a new assistant talk to real customers without learning about its mistakes from them.
The mode is off by default — on purpose
While it's on, the customer gets no reply until a person sends it. So turn it on when someone is actually watching the Conversations section (or your linked chat with the assistant) — or say right away what to do when nobody answers (see below).
How to turn it on
In words, to the builder:
Turn on training mode
Show me the replies first
Or with a button: the Training section in the dashboard menu → "Turn on mode". To turn it off — "Turn off training mode" or the same button. A workspace admin turns the mode on and off; anyone who answers in Conversations can decide on drafts.
What the customer sees
Right after their question — an honest line:
The reply is being prepared by the assistant and checked by a person — please wait a minute.
And then the reply itself, in the same place they asked: in the chat window on your site, in Telegram, Viber, WhatsApp, Discord, Slack or by email (for email we don't send the "please wait" line — only the reply). They see "typing…" only when a person is actually typing a reply.
You can replace the line with your own, like any other text in the chat window: "Change the reply-check message to 'Our manager is checking the reply'". The customer will see the same text in messengers too.
What the operator sees
In the Conversations section, a bar "N drafts waiting for you" appears above the list of requests — on its own, no page reload needed. Open a conversation: under the messages there's the assistant's draft —
- the reply text, which you can edit right in the field;
- "Based on" — which of your materials the assistant took the answer from;
- "Along with the reply, the assistant would have" — the actions it wanted to perform (see below);
- the "What was wrong" field — optional. The customer won't see it: it's a note for the assistant.
And three ways forward:
| Button | What happens |
|---|---|
| Send as is | the customer gets the draft unchanged |
| Correct and send | the customer gets your text; the assistant remembers how it should have been |
| Reply myself | the draft is discarded, the customer gets your reply |
After the reply, the conversation stays with the assistant: this customer's next question becomes a draft again. If the customer asks for a live person, everything works as usual — you lead the conversation, and there are no drafts in it.
Answered the customer with the regular conversation field instead of the draft card? The draft for the same question leaves the queue on its own: the customer already has your reply and won't get a second one from the assistant. The same happens if the customer called a person while the draft was waiting: it won't go out on its own anymore.
When there are two of you
As soon as someone starts editing a draft, a colleague sees "already editing …". It's a marker, not a lock: whoever's reply goes out first decides, and the other person sees that the question has already been answered — the customer won't get two replies.
When the customer adds more
If the customer wrote again while the draft was waiting, the old draft is not sent: the assistant writes a new one that takes all their messages into account, and the card says "rewritten — the customer added more". A reply to yesterday's question won't go out, even if you managed to press the button.
Actions wait along with the reply
If the assistant wanted to do something in its reply — place an order, send an email, add to the calendar, give a payment link — in training mode it doesn't do it until a person sends the reply. Each such action is named on the draft card:
- with a checkbox — an action you connected: it runs when the reply is sent. Untick it and it won't run. If the action fails, the reply does not go to the customer (otherwise they would read "order placed" with no order) — the card will say why. An action that has already run (but the reply didn't get through) is marked "already done" and doesn't run a second time — even if the customer wrote more in the meantime and the assistant wrote a new draft with the same action;
- without a checkbox, noted "can't run from here" — things the platform does itself (calling a person, calendar, payment, saving to cards). These don't run from a draft: do it yourself or answer in your own words.
An action for which you require admin permission becomes a permission card even from a draft — the operator can't run it alone. Reading goes on right away: the assistant doesn't need to wait to search knowledge and the catalog.
If nobody answers
By default a draft waits as long as it takes. You can change that:
| Say | And then |
|---|---|
| "Send it yourself if I haven't replied in 5 minutes" | the draft goes to the customer on its own. Only a draft without actions: one with an action waits for a person, and the customer hears that we'll reply later |
| "If I can't make it, tell the customer we'll reply later" | after the time you named, the customer hears: "We can't reply right now — we'll get back to you a little later", and the draft keeps waiting |
| "Wait for me as long as it takes" | same as the default |
The same is in the Training section, the "If nobody answers" row. There's no separate shift schedule: one rule for all hours, and at night it works the same as during the day. Drafts aren't lost if the service restarts.
A draft also won't go out on its own in three more cases — then the customer hears "we'll reply later", and the draft waits for a person with a note about what happened:
- you already tried to reply (send, correct or write your own), and the channel refused — for example, the messenger was unavailable. Text you just wanted to replace is not sent without you;
- sending broke off halfway — the customer may already have it, so check the conversation first;
- the assistant didn't manage to write a draft (there's a question but no text).
The default "wait" really means wait
If the mode is on, nobody looks into Conversations and there's no linked chat, the customer will be left with "please wait a minute". Turn the mode on when someone is there to answer, or say what to do without you: "send it yourself after 5 minutes" or "say we'll reply later".
In your chat with the assistant
If you have a linked chat with the assistant in a messenger, drafts arrive there too — with a number:
| Write | And then |
|---|---|
yes 123456 | the draft goes to the customer as is |
123456: your text | the customer gets your text |
123456 не так: price is outdated | a note on a reply that was already sent (for now the words «не так», "not right", only work in Ukrainian) |
Did the customer write more while the card was sitting in the chat? Its number no longer works: on "yes 123456" the assistant will say the draft was replaced by a new one and give its number. Nothing goes to the customer under the old number.
You can untick a single action only in the dashboard. Only the owner, verified by the messenger, can decide on a draft from the chat — "yes 123456" from anyone else sends nothing.
How the assistant learns
Examples. When a customer asks something similar to a question where the assistant was already corrected, it sees up to three such cases: what was asked, what the person finally replied and what was wrong. These are sample replies to customers, not new rules: an operator's note gives the assistant no new rights and doesn't change its personality. They keep working after the mode is turned off.
"Similar" here means asked in similar words. "How much is delivery to Lviv?" and "And delivery to Lviv, how much does it cost?" are similar; "What's the price to bring it to Lviv?" no longer is, although the question is the same. And the other way round: "How much is delivery to Kyiv?" differs by one word, so the assistant will see the Lviv example — and has to judge for itself whether it fits. So examples are good at catching questions customers ask the same way, while an answer that should work however it's phrased is better moved into knowledge — with the Add to knowledge button under a suggestion or in words to the builder.
The assistant doesn't see the customer's phone and email from their question in the example: an example is about how to answer, not about who asked. The person's reply goes into the example as is — so something personal about one customer (name, address, order number) could be read by the assistant in a conversation with someone else; such a correction is better removed from the examples.
Suggestions. In the Training section the platform groups similar corrections together: "5 of 6 corrections are about a similar question". Under each suggestion —
- Add to knowledge — the "question — answer" pair goes into knowledge (you can edit the answer first); it can be undone like any change;
- Save as a skill — a skill draft appears, which you still need to read and approve;
- Don't suggest.
Suggestions are built by rules, without a model, and cost no credits. You can also ask the builder to "sum up the corrections": it will tell you what the assistant lacked most often and suggest changes — including to its personality; it makes them only after your "yes".
Removing an example. If a correction was itself a mistake, an admin removes it with the "Remove from examples" button in the list of decisions.
When you can turn the mode off
The Training section shows readiness: how many of the last 20 human decisions went to customers without edits. Drafts that went out on their own don't count — nobody read them.
When 9 out of 10 or more go out without edits, the dashboard will suggest turning the mode off or keeping it with "send on its own if nobody answers". The decision is yours.
Good to know
- Credits. A draft costs the same as a regular assistant reply — including the one you discarded. A customer who sends three messages in a row means three drafts written (you'll get one, the last, with all their messages).
- You turned the mode off, but drafts are still waiting. They stay in Conversations, and you can decide on them as before; the assistant answers new questions on its own. A customer whose question was waiting won't get a reply without you — clear the queue before turning the mode off.
- Memory about the customer. If the assistant remembers customers it knows (a signed-in site visitor, a verified messenger user), what goes into that memory is what the person actually sent, not the draft.
- Your own conversations. When you talk to the assistant yourself (in a linked chat or in a chat window opened from the dashboard) or a trusted person does, the mode doesn't apply: it's about customers.
- Other programs. Requests from other agents and programs to your assistant aren't affected: a program expects a reply right away and won't understand "please wait a minute".
- Erasing. Drafts and decisions are erased together with the assistant, and with a visitor's data when they ask to erase it; they are included in the account data export.
- Handing over the assistant. Corrections move to the new owner — they're knowledge about the business. Which operators made them — doesn't; the mode itself reaches the new owner switched off.