Troubleshooting
Messages are not arriving
The inbox is marked unhealthy. The connection is failing. For email, the credentials stopped working — a rotated password, an expired app password, or the provider disabling IMAP. Re-enter the password and use Test connection. For OAuth channels, access was revoked on the provider's side; reconnect.
The inbox is healthy and nothing arrives. Confirm the mail is reaching the mailbox at all. A rule that files support mail into a folder hides it, because only the inbox folder is read.
Some messages arrive and some do not. Usually a filter upstream. Check your mail provider's rules and spam handling.
A conversation is missing rather than the whole channel. Check the Spam view: it is excluded from every other view and from search elsewhere.
Replies are not sending
Email arrives but replies fail. IMAP works and SMTP does not. Check the SMTP port — 587, or 465 for implicit SSL — and whether your provider uses a different host for sending.
A social channel rejects the reply. Meta and LinkedIn allow free-form replies only within 24 hours of the customer's last message. Outside that window an approved template is required.
The reply is too long. Every channel except email has a limit. The editor shows it.
The workspace is read-only. A lapsed subscription blocks sending. See Subscription.
The AI is not replying
Work down this list in order:
- Is the classification set to reply? Automatic reply is per category, not per workspace. Check the category the conversation was filed under.
- Is the confidence above the threshold? The percentage is on the conversation. Below the threshold nothing is sent, whatever the category says.
- Was it classified at all? An unclassified conversation follows no category's rules.
- Does the knowledge base cover it? If not, the assistant hands over on purpose. This is the most common answer.
- Is automatic classification on for the workspace?
The AI replies badly
| Symptom | Fix |
|---|---|
| Confident and wrong | The article behind it is out of date, ambiguous, or contradicted elsewhere |
| "I don't have information about that" on something you documented | The article does not retrieve. Rewrite the title in customer language, put the answer in the first paragraph |
| Right facts, wrong tone | The identity prompt or a category prompt |
| Answers the wrong question | Classification descriptions are too vague, or two categories overlap |
| Too long or too short | The identity prompt. Give it a number |
| Wrong language | The behaviour prompt should say to answer in the customer's language |
The rule: facts in the knowledge base, tone in prompts, routing in classifications. See Reviewing quality.
Classification is wrong
Everything lands in one category. Its description is too broad, or the others are too thin. Descriptions are what the assistant reads.
Similar messages land in different categories. Two categories overlap. Write into each description what belongs to the other one.
Lots of unclassified conversations. The catalog has a hole. The unclassified pile names the missing category.
Correcting by hand is not just a fix for that conversation: it becomes a verified example that improves later ones.
Knowledge sources
The source shows ERROR. The message on the source names the cause: rejected credentials, an unreachable site, or paths that no longer exist. Existing content is kept, so the assistant keeps working on what it already had.
A sync returns nothing. The existing rows are kept deliberately, because an
empty crawl is far more often broken than genuine. Check the base URL, the
paths and whether robots.txt blocks crawlers.
New articles are not being used. Give it a minute or two for indexing, then test with Find answer in a real conversation. If the article still does not come back for a question it clearly answers, the title and first paragraph are the problem.
The website chat widget
It does not appear. Check the snippet is on the page and the widget key is correct, and check the browser console for a blocked request.
It appears but will not send. The domain is not on the allowlist. Add it in
the inbox's settings, including the www. variant if you use it.
The visitor sees no reply. If the assistant handed over, no automatic reply is sent by design — the conversation is waiting for your team, at high priority in the AI Queue.
Sign-in
"Session expired". By convention a 401 in the app means the session ended. Sign in again.
Two-factor code rejected. Check the clock on the device with the authenticator app; TOTP codes fail when it drifts.
Lost the authenticator device. Use a recovery code. Without one, recovery goes through support.
Invitation link does not work. Invitations expire. Ask an Admin to resend it.
Still stuck
Have this ready and it will be a much shorter conversation:
- What you expected and what happened.
- The workspace name and, if relevant, the conversation.
- When it started.
- Whether it affects everyone or one person, one channel or all of them.