Skip to main content

How the AI works

The assistant is not a chatbot bolted onto a help desk. It runs on the same pipeline every message goes through, and every step it takes is visible in the conversation afterwards.

The pipeline

1. The message becomes a conversation

An incoming message either starts a new conversation or joins an existing one. Email threading follows the mail headers, chat conversations follow the session, so a customer replying to their own thread stays in one place.

2. Classification

The assistant sorts the message into one of your categories. It sees the catalog you defined with each category's description, the conversation so far (channel, language, subject, recent messages), the nearest conversations you have already classified, and knowledge that looks related.

Two things make this better over time. Conversations a person classified by hand count as verified examples and carry more weight than ones the AI classified itself. And NONE is a valid answer: when nothing fits, the assistant says so and records why, rather than forcing a message into the closest category.

The result carries a confidence between 0 and 1, shown as a percentage in the conversation. Below the workspace threshold, no automatic reply and no automatic close happens, whatever the category says.

3. Knowledge retrieval

Before writing anything, the assistant searches your knowledge base for relevant passages. It searches with the customer's latest message and the subject, a translation of both when the conversation is not in your workspace language, and the earlier messages in the thread.

Search is hybrid: a semantic (vector) search and a keyword (full-text) search run in parallel and their results are fused. That combination finds the article that talks about "range in cold weather" when the customer wrote "how far in winter", and still finds the one that contains an exact part number.

Each passage reaches the assistant with its title, its source link and its text. Those links are what let it point a customer at the right help article.

4. The reply

What happens next is decided by the classification:

Category settingWhat happens
Automatic reply onThe answer is written and sent, and the conversation may be closed
Draft onlyThe answer is written and left in the editor for review
NeitherNothing is written; the conversation waits for a person

The assistant writes in the customer's language and follows the writing rules you set for the workspace and, if the category has one, its prompt.

5. Hand-over

When the knowledge base does not cover the question, the assistant does not improvise. It escalates: the conversation goes to high priority, no reply is sent, and the reason is recorded.

On website chat the visitor is told plainly that a person will take over, and anonymous visitors are asked for a name and an email so the team can reach them.

This is the behaviour to lean on. An assistant that answers 60% of messages correctly and hands over the rest is worth far more than one that answers 100% with a fifth of them wrong.

What stays visible

Every decision is recorded on the conversation:

  • The classification, with the AI's reason and its confidence.
  • Whether a reply was sent by the AI or by a person.
  • Who closed the conversation and why: manually, by the AI, or by inactivity.

A classification a team member set by hand has no AI reason attached, which is how the interface distinguishes the two. Changing a classification by hand clears the AI's reason and confidence, and the corrected classification then serves as a verified example for future messages.

Where to tune it