Plan requirement
| Subscription | Suite Team or higher, Support Team or higher |
| Access | Admin |
Publishing puts the agent in front of customers, and it is not a one-off launch event: every later change to its settings needs publishing again before customers see it. You need to be an admin in both Support and Knowledge.
Publish it
- Confirm the handover works and goes somewhere staffed.
- In Admin Center, click AI, then AI agents and AI agents.
- Click the agent, then Publish AI agent.
- Select the channels it should be available on. One agent can go to several channels, and several channel types, at once.
- Read any warnings, such as conflicting triggers, before continuing.
- Click Publish, tell your agents it is live, and read transcripts the same day.
Republishing is how you remove it too
To take an agent off a channel, publish it again without that channel selected. There is no separate unpublish, which means the same button both widens and narrows the deployment.
Zendesk warns that republishing restores the agent's triggers to their default configuration, so if somebody has customised those, check them after every publish rather than assuming they survived.
A deactivated trigger takes the agent offline
On an email or web form channel, publishing creates one trigger and two automations for the agent. Deactivate that trigger and the agent is unpublished automatically and returns to draft, so somebody tidying up business rules can take it off the air without touching the agent at all.
It works the other way too, and less tidily: when an agent is unpublished the trigger is deactivated but the two automations are not. Turn those off by hand, or they keep acting on tickets for an agent that is no longer answering.
Start where the risk is lowest
One channel, ideally messaging, during hours when people are around to pick up handovers. A launch on every channel at once removes your ability to tell what is working.
Tell the team first
Agents will receive conversations that started with the AI. They should know that is expected, that the transcript comes with it, and where to report a bad answer. An agent who thinks the AI is broken will tell customers so.
Read transcripts on day one
Not the dashboard. Twenty actual conversations, in full. Every deployment produces surprises in the first fifty exchanges, and they are the cheapest ones you will ever fix.
Fix articles, not the agent
Most bad answers trace to content: a missing article, or one that does not state the answer plainly. Changing the article fixes the AI agent and helps every reader, and it does not need a republish.
Widen slowly
Add subjects and channels when the transcripts on the current ones are consistently good. Each addition is a new set of failure modes, and adding them one at a time is how you keep knowing which change did what.
Comments
0 comments
Article is closed for comments.