Revise a bot answer and train
When to use
For specific imperfect bot answers:
- Customer asked a question.
- Bot answered, but the answer was wrong or incomplete.
- Agent edits the bot's reply to what it should have said.
- AskVault saves the corrected pair for future retrieval.
Different from editing knowledge sources; "revise and train" is conversation-level.
How to revise
There are two places to do this, and they save to the same knowledge base.
From a conversation in Activity:
- Open the conversation and find the AI reply you want to correct.
- Click "Revise & Train" under that reply.
- Edit the answer to what it should have been.
- Save.
From the Help Desk inbox:
- Open the conversation and find the AI reply you want to correct.
- Click "Correct" under that reply.
- Edit the answer to what it should have been. The question the customer actually asked is shown above it, taken from the conversation.
- Optionally list the facts the answer must contain, so a regression test can check for them.
- Click Save correction.
Both write the corrected pair into your knowledge base, where retrieval weights it above ordinary crawled content when a similar question comes in.
Correcting an answer your agent flagged for review
When your agent isn't confident in an answer — low confidence, or an answer it couldn't ground in your knowledge base — it flags that conversation for review. Flagged conversations are the ones most worth correcting, because the agent has already told you it was unsure.
On a flagged conversation the header shows two actions:
- Correct answer — opens the correction editor already pointed at the most recent low-confidence AI answer in that conversation, with the customer's question shown above it, so you don't have to scroll the thread to find which reply was the problem.
- Mark reviewed — for when you've checked the answer and it was actually fine. This dismisses the flag without changing anything.
Saving a correction also clears the review flag. You don't need to click Mark reviewed afterwards — an answer you've already fixed stops being counted as unreviewed straight away, so the review queue reflects work that's genuinely still outstanding.
If the correction fails to save, the flag is deliberately left in place, so nothing drops off your queue on the strength of a change that didn't land.
What happens next time
When a customer asks the same or similar question:
- Retrieval finds the trained Q&A pair with high relevance.
- Bot uses the corrected answer.
- About 90% of revised answers are reused on next similar query.
Audit trail
Every revision logs:
- Original bot answer.
- Revised version.
- Reviewer (agent).
- Timestamp.
Useful for QA and tracking accuracy over time.
Bulk review
Reviewer dashboard under Analytics > Bot Answers > Review:
- Low-CSAT bot answers sorted by frequency.
- Recent conversations with bot replies.
- One-click filter for "needs revision".
About 5 to 10% of bot answers benefit from manual review in early days; drops to under 1% as knowledge matures.
Bot self-improvement
After revision:
- The new Q&A pair is retrievable immediately.
- Bot accuracy on similar questions lifts measurably.
- Compounds over time. Most teams see 15 to 25% accuracy lift in 90 days through ongoing revision.
Limits
- Trained Q&A pairs per workspace. Plan-dependent (up to MB cap).
- Revision rate. Up to 100 per day per workspace.
- Audit retention. 365 days standard.
Common pitfalls
Revising too eagerly. Not every imperfect answer needs revision. Focus on high-frequency questions.
Wrong question phrasing. Revise both question and answer; capture how customers actually ask.
Conflicting Q&A pairs. Two trained pairs for similar questions. Periodically audit and dedupe.
FAQ
Does this re-train the model?
No. We add a Q&A pair to your knowledge index. The base model is unchanged.
Can multiple agents review?
Yes. All workspace-Editor and above can revise.
Will the customer see the revised answer if they re-ask?
Yes within 30 seconds. The bot uses the corrected pair.