Plan requirement
| Subscription | Suite Professional or higher, Support Professional or higher |
| Also required | Copilot |
| Access | Admin |
Three things shape this before you touch a setting: scoping runs on a specific tag applied by a trigger, all groups have access by default, and auto assist cannot be configured on the chat or voice channels. It works on messaging and email.
Turn it on
- In Admin Center, click AI in the sidebar, then select Agent copilot and Auto assist.
- Switch on Show auto assist replies and actions in the agent composer.
- In Who has access, select the groups that should have it. By default every group does, so narrowing is an active step.
- Write at least one procedure before going further.
- Create a ticket trigger that adds the _agent_copilot_enabled_ tag to the tickets your procedure covers.
- Test it, then watch real tickets go through.
The tag is the scope
Auto assist only acts on tickets carrying _agent_copilot_enabled_. You add it with a trigger, and the trigger's conditions are where the real scoping happens: Ticket > Channel to limit it to one channel, Ticket > Intent to limit it to one kind of request.
That is why "turn it on for one ticket type" is a trigger question rather than a setting. Broad from day one means every oddity arrives together and you cannot tell which procedure caused what.
Give the trigger a condition that the tag is not already present. Without it the trigger runs again on every update, and a rule that fires repeatedly on the same ticket is harder to reason about than one that fires once.
If you also use suggested first replies, know which one wins: where both apply, auto assist goes first. So an agent who expected the older suggestion and sees something else is not looking at a fault.
Write the procedure first
Auto assist with no procedures does very little, and the team concludes the feature is empty. Zendesk's own order is the same: turn it on, write the procedure, then tag the tickets it covers. One good procedure for your most common multi-step case is a better launch than the switch on its own.
Testing by email needs a non-agent address
Auto assist does not respond to agent comments, so an email ticket created from an address belonging to an agent produces nothing at all. Test from an address that is not an agent's, or you will spend an afternoon debugging a working system.
Start with one team
Use Who has access to pick the group that handles that ticket type. They know the process well enough to notice when a procedure is subtly wrong, which is exactly the feedback you need in week one.
Say what changes for agents
Proposed replies and actions they approve, in the composer. Not instructions they must accept, and not a measure of their work. Both need saying out loud, because both are assumed otherwise.
Approval stays with the agent
Every suggestion must be approved by the agent before it is sent, which is what makes trying it on live tickets low-risk. Worth being explicit about with anybody nervous, and it is what makes it acceptable on cases involving money.
Watch the first fifty
Real tickets, start to finish. Every gap in a procedure shows up there, and they are cheap to fix while the scope is still one ticket type.
Comments
0 comments
Article is closed for comments.