Skip to main content
AI Manager is a second check on each reply. Before a reply reaches the customer, a reviewer scores it against your instructions, the Knowledge Base and the tools the agent has; a reply below your target score is discarded and the agent tries again with the reviewer’s notes. Use it on agents where a wrong answer is costly, such as one that quotes policies or takes orders, and accept that each retry makes the reply slower.

​
Before you start

  • Open the agent from Agents → Agents; the switch is under Messaging.
  • Switch on Handover under Capabilities → General if you want the agent to hand over after failed retries.

​
Switch it on

1

Enable AI Manager

Under Messaging, switch on AI Manager (“Check quality of agent responses and retry if below target score”).
2

Set the target

Fill in Target quality score (0.0-1.0). The hint reads “Higher values demand better responses but may cause more retries. Recommended: 0.5-0.7”. The default is 0.6.
3

Set the retry limit

Fill in Maximum retry attempts, the number of times a reply is regenerated when it scores below the target. The default is 3. The hint warns that higher values may cause delays.
4

Decide what happens after the last retry

Tick Handover to human after max retries to hand the conversation over instead of sending a reply that still scored below target. If Handover is off, ticking it shows Handover capability must be enabled to use this feature.
5

Save and publish

Click Save & review, then Publish changes.

​
How a reply is scored

The reviewer sees the same instructions and tool list as the agent, the conversation up to the customer’s last message, and the reply the agent just drafted. It gives a score between 0 and 1 from three criteria: Each criterion is rated 1.0 (perfect), 0.7 (minor issues), 0.5 (acceptable), 0.3 (poor) or 0.0 (failure), and the weighted total is compared with your target. A target of 0.6 therefore lets through replies with minor issues and blocks replies with a significant problem in any one area. Only replies in conversations assigned to the AI are scored. Replies your team writes are never scored.

​
What happens when a reply scores below target

  1. The reply is not sent. It is kept internally as a rejected attempt together with its score and the reviewer’s note.
  2. The agent is told “The manager reviewed your last response and said it needs improvements” with the note, and writes a new reply.
  3. The new reply is scored again. This repeats until a reply reaches the target or Maximum retry attempts is used up.
  4. After the last retry:
    • With Handover to human after max retries off, the last reply is sent as it is.
    • With it on, the agent is instructed to hand the conversation over instead of answering. The conversation moves to Needs Attention, the customer gets the usual handover confirmation, and your handover notifications are sent. The reason is recorded in the conversation’s activity history.
Replies that pass are sent normally; their score and note are stored with the message. AI Manager does not use extra credits: a customer message the agent answers costs one credit whether the reply was accepted first time or retried.

​
Choosing the settings

  • Target quality score: start at the default 0.6. Raise it toward 0.7 only after checking the Inbox for replies that still got through with problems; every increase means more retries on borderline replies.
  • Maximum retry attempts: each retry is a full new reply, so a customer can wait several times the normal response time when replies keep failing. Two or three is usually enough; a reply that fails three times rarely passes on the fourth.
  • Handover to human after max retries: turn it on for agents whose wrong answers you would rather not send at all. Without it, AI Manager can only improve replies, not stop them.

​
AI Manager and Spotlight

AI Manager works on each reply before it is sent. Spotlight scores whole conversations after the fact on six dimensions and lists conversations to review. Use Spotlight to see whether the agent’s replies are good enough in general, and AI Manager to enforce a minimum on every reply for a specific agent.