Learned replies
See what the AI has picked up from your team's answers, and decide which of it the AI is allowed to keep repeating.
When an operator answers a customer and the conversation is resolved, Chatasist can keep that answer: the customer's question and your colleague's reply are stored together as a learned reply. Later, when someone asks something similar, the AI can find that pair and answer in your team's own words instead of a generic one.
That is the feature at its best. It is also how a one-off answer, a wrong answer, or an answer that was only true last month ends up being repeated to everybody who asks.
Learned replies is where a person decides which is which. Find it under Knowledge → Learned replies.
What gets learned#
Only answers a human typed, in conversations that were resolved. Chatasist skips replies that carry no information — greetings, "one moment", "you're welcome" — and scrubs personal details out of both the question and the answer before storing them.
Two settings control whether anything is learned at all:
- The workspace learning mode. Off learns nothing. Auto stores new pairs and lets the AI use them immediately. Review stores them but holds them back until someone approves.
- Learn from agent replies on each channel, so you can learn from your website widget without learning from, say, an internal test channel.
If learning is off, the page tells you so rather than showing a list nothing is using.
Reviewing#
Start with Needs review
Everything the AI has learned but nobody has judged. Pairs the AI has already been marked down for are shown first, because those are the ones most likely to be wrong.
Read the question and the answer together
A reply that was perfectly good in its original conversation can be wrong out of context — "yes, we can do that for you" means nothing on its own. Select Open conversation to see where it came from.
Approve, edit, or reject
Approve keeps it. Reject takes it out of use for good. Edit fixes the wording — usually the right call, since most bad pairs are a good answer with one stale detail in it.
Use the checkboxes to decide on a whole page at once. A workspace that has been running for months can have hundreds of pairs waiting, and reviewing those one at a time is not realistic.
Rejecting keeps the row rather than deleting it. It is the record that a human ruled this answer out — and if it were deleted, the same reply would simply be learned again the next time conversations are processed.
"Not searchable"#
The AI finds a learned reply by matching the meaning of the customer's question against the stored one, which requires the pair to have been indexed. A pair that was saved but never indexed is invisible: it looks perfectly healthy in any list, and the AI can never retrieve it.
Those pairs are badged Not searchable, and a banner at the top counts them. It matters because approving one changes nothing until it is indexed — the approval is real, the effect is not.
If you see a large number here, tell support: it usually means indexing was failing for a period, and the affected pairs need to be re-indexed in bulk.
Editing and re-indexing#
Editing the answer takes effect immediately — it is the text the AI reuses.
Editing the question changes what the pair is matched against, so the pair has to be re-indexed before it will be found by the new wording. The page says so while you are editing. It usually takes under a minute.
Editing changes future answers only. A customer who was already given the old wording keeps what they were sent — learned replies are not retroactive.
How this differs from the knowledge base#
The knowledge base is what you deliberately wrote for the AI to use. Learned replies are what it picked up from your team on its own.
Both are used when answering, but they are not treated the same: knowledge base entries are authoritative facts, while learned replies are guidance on phrasing and edge cases the knowledge base does not cover. If a fact matters, put it in the knowledge base — do not rely on the AI having overheard it once.