AI Advisor
The AI Advisor looks back once a week at the conversations where your AI auto-reply could not handle the customer on its own. It then proposes concrete fixes — a knowledge entry to add, a lesson to save, a reply setting to adjust — that a company admin can apply, edit, dismiss or undo from one page. Nothing changes until an admin applies a suggestion, unless you switch on Auto-apply.
Weekly review
The Advisor runs once a week, early on Monday (02:00 in UTC+7, which is Sunday 19:00 UTC), and looks at the previous 7 days. The page has no button to start an analysis by hand.
Suggestions, not changes
Each suggestion waits as Pending until a company admin applies or dismisses it. The only exception is Auto-apply, which you switch on yourself.
Company admins only
Only company admins can view and manage the Advisor. Other roles see the tab with a lock and a message saying so.
See the evidence
A suggestion based on conversations shows how many it is based on, with links to open up to five of those chats.
Turn On the Advisor
The Advisor is off by default. A company admin turns it on from the AI Agent page.
- Click AI Agent in the sidebar, then open the Advisor tab
- Switch on Weekly Advisor analysis
- Under the switch you see the cost: about 80,000 tokens per run (16 credits), and a run is skipped when your balance is below 100,000 tokens (20 credits)
- Optionally switch on Auto-apply suggestions (see Auto-Apply below)
When the first analysis starts, company admins get an in-app notification titled "AI Advisor is on" that repeats the token cost. Admins are also notified when new suggestions are ready, or when the Advisor applied changes automatically. Each notification opens the Advisor page.
What the Advisor Looks At
- How your AI auto-replies ended during the past 7 days. Every AI reply is recorded with its outcome (sent, handed to a person, held back, error) and a reason
- A sample of up to 50 chats where the AI could not handle the customer on its own, with a short stretch of messages before and after the failure. This includes chats where the AI replied but a person stepped in within about two hours
- Questions that your knowledge could not answer
- Corrections your team made: thumbs-down reactions on AI messages, and AI suggestions that an agent rejected, corrected or edited
- Your current setup: AI auto-reply settings, what your knowledge already says, your Teach answers policies, the sync status of your knowledge sources and any failed Data Connection lookups, so suggestions do not repeat what you already have
An AI reply counts as a failure when the AI could not answer on its own: it had nothing to answer from, it handed the chat to a person, it held back because its confidence was below your threshold, it sent a fallback message, it held a reply for review because it lacked knowledge, or it hit a technical error. A chat where a person took over before the AI replied is not counted as an AI failure.
Card numbers, email addresses and phone numbers are masked in the chat text the Advisor reads. When you apply or edit a suggestion, text that contains personal details is rejected.
The Run Summary
Above the suggestions, the Advisor page shows the date of the last run and its status: Running, Completed, Failed or Skipped. For a run that was not skipped it also shows four figures:
AI handled on its own
The share of the week's AI auto-reply attempts that ended with a reply that was not a failure.
Handed to a person
The share of attempts that the AI handed to a person.
Top reasons
The three most frequent failure reasons, with a count. The reasons are:
- No knowledge to answer from
- Customer asked for a person
- A policy rule escalated it
- Confidence was below the threshold
- The AI chose not to answer
- Fallback message because nothing matched
- Message type not supported
- Technical error
- Other
Tokens used
The tokens that run spent on its analysis. One credit on the Billing page equals 5,000 tokens.
When a run is skipped
Kinds of Suggestions
Every card shows a badge for its source — Chat analysis or Repeated corrections (from your team's own corrections) — and a badge for its kind.
Add knowledge
A customer question with a drafted answer.
Apply adds it to your Knowledge Hub as a conversation example. If a conversation example with the same customer question already exists, the answer is added to it. You can edit the question and answer first.
Lesson
A customer question, what the AI said, and a better answer.
Apply saves it as a lesson that the AI uses from then on. You can edit all three texts first.
Settings
A change to one of your AI auto-reply settings, shown as Current → Proposed.
Apply changes the setting. Only four settings can be applied this way (see below).
Source health
A knowledge source whose sync is unhealthy, with its last sync status and error.
For your information only. The card links to the Knowledge Hub.
Data connection
A suggestion to add or fix a Data Connection, with a table of failed lookups (source, error, count).
For your information only. The card links to Data Connections.
Policy review
An observation about one of your Teach answers policies.
For your information only. The Advisor never edits policies. The card links to the question in Teach answers.
Cards that cannot be applied carry a For your information badge and offer only Dismiss and a link to the page where you can act on them.
Settings the Advisor can apply
- Never-say phrases — adds up to 3 phrases per suggestion, each up to 100 characters. Phrases already on your list are not repeated. Topics such as legal terms, pricing, fees, refunds, warranty and tax are not proposed, because they are business decisions
- When the AI has no answer — can only move toward a person: from Silent to Reply then route to agent or Hand off to human agent, and from Reply then route to agent to Hand off to human agent. Reply then route to agent is an older option that the Auto reply tab no longer shows
- Reply tone — one of the tones on the Auto reply tab: Professional, Friendly, Empathetic, Concise or Direct
- Confidence threshold — a value from 50 to 95, changed by no more than 10 points in one suggestion, and never while a delivery mode is set
Settings it only mentions
For these settings the Advisor can show a suggestion, but it never changes them. The card says "The Advisor never changes this setting. Change it yourself if you agree."
- Conversation stages
- Analysis guidance
- Delivery mode
- Disabled channels
A Confidence threshold suggestion that breaks the rules above is also shown for information only, with the reason on the card.
Review and Act on Suggestions
Use the Pending, Applied and Dismissed filter above the list to switch between suggestions. Pending suggestions with the most supporting conversations come first. A card shows its title, the reason for the suggestion, the content (for example the question and answer, or the current and proposed value), and the evidence.
Apply
Applies the suggestion as shown. A message confirms "Applied. The AI will use it from now on." and offers an Undo link.
Edit
Available on Add knowledge and Lesson cards. Opens "Edit before applying" so you can change the texts, then Apply with edits. A question can be up to 500 characters and an answer up to 2,000.
Dismiss
Moves the suggestion to the Dismissed filter without changing anything.
Undo
Shown on cards under Applied. Reverses the change; the suggestion then leaves the Applied list.
What Undo reverses
- Add knowledge — the conversation example the suggestion created is removed, or, if the answer was added to an existing example, only that answer is removed
- Lesson — the lesson is deleted
- Settings — the previous value is restored. For Never-say phrases, only the phrases the suggestion added are removed
Messages you may see
"This setting changed since the suggestion was made. Wait for the next analysis."
Someone changed that setting after the Advisor wrote the suggestion, so the suggestion no longer matches your live value.
"The text is empty, too long, or contains personal details like a phone number or email. Please edit it."
The texts must not be empty, a question can be up to 500 characters and an answer up to 2,000, and personal details are not accepted. Use Edit to fix the text, then Apply with edits.
"This suggestion was already handled. Refresh the page."
Another admin applied or dismissed it first.
"Another action on this suggestion is in progress. Try again in a moment."
Two actions on the same suggestion arrived at once.
"This setting was changed again after you applied it, so it cannot be undone here."
Undo for a setting only works while the setting still has the value the suggestion gave it. Change it back by hand on the Auto reply tab if you want to.
Auto-Apply
Auto-apply suggestions is a second switch on the Advisor page. It is off by default and can only be switched on while Weekly Advisor analysis is on. With it on, the Advisor applies its new suggestions right after each run.
- Only suggestions that can be applied are auto-applied. Informational cards, and any suggestion that cannot be applied at that moment (for example because the setting it targets has changed), stay Pending for an admin
- Every auto-applied change appears under Applied with an Applied automatically badge, and you can undo it like any other applied suggestion
- Instead of the usual "new suggestions" notification, admins get one titled "AI Advisor applied N changes automatically"
How Suggestions Are Kept
Per run
At most 10 new suggestions, and at most 5 of any one kind.
Pending for 30 days
A Pending suggestion that is not touched for 30 days expires. If the same pattern shows up again first, its evidence is added to the existing card.
No repeats for 90 days
A suggestion you applied or dismissed is not proposed again for 90 days.
Troubleshooting
The page says no analysis has run yet?
- Check that Weekly Advisor analysis is switched on. The Advisor is off until an admin turns it on
- The first analysis happens at the next weekly run after you turn it on, and the page cannot start one by hand
The last run shows Skipped?
- The reason is shown next to the status: AI auto-reply is not set up, not enough AI failures to learn from, or the token balance is below 100,000 tokens (20 credits)
- A skipped run creates no suggestions. The Advisor tries again at the next weekly run
The last run shows Failed?
- The page says the last analysis failed and that it will try again next week
A suggestion I expected is missing?
- A suggestion you applied or dismissed is not proposed again for 90 days
- A single run creates at most 10 suggestions, and at most 5 of each kind
- A Pending suggestion that nobody touches for 30 days expires and leaves the Pending list
The Auto-apply switch is greyed out?
- Auto-apply can only be switched on while Weekly Advisor analysis is on
- Both switches are also disabled for a moment while a change is being saved
The Advisor works from your AI auto-reply. See AI Auto-Reply for the settings it can change, and the Knowledge Hub for where applied knowledge appears.