Putting the Assistant Where People Already Talk

Why chat is the right place

An internal assistant on a separate website gets used in week one and forgotten by week four. One in the tool people already have open all day gets used continuously.

That single fact determines the adoption curve more than the quality of the answers does.

Design rules

1. Summoned, not ambient — respond when asked, not to every message
2. Reply in threads so channels stay readable
3. Cite sources with links people can open
4. Respect permissions — answer per the asking user's access
5. Offer a human when it does not know

A bot that comments on everything is muted within a week. One that answers well when asked becomes part of how the team works.

What it is genuinely good at

• Policy and process questions from your documentation
• “Where is the current version of X”
• Summarising a long channel or thread on request
• Looking up a record from a connected system
• Capturing an action item into your task tool

Permissions in a chat context

The assistant answers as the asking user, with their access. In a shared channel, that raises a design question: whose permissions apply when several people can read the reply?

The safe answer is to restrict channel replies to content everyone in that channel can see, and put anything else in a direct message.

Measuring it

Questions asked per week, answer rate, citation clicks and the number of the same questions still going to people directly. The last is the real measure.

Also track what it could not answer — that list is your documentation backlog, generated for free.

Frequently asked questions