Plan requirement
| Subscription | Suite Team or higher, Support Team or higher |
| Access | Admin |
Reading what happened. Half an hour in running an AI agent that people skip.
Read them
- In the AI agents workspace, select the agent and click Conversation logs in the sidebar.
- Set the period, then filter. The list opens on the past seven days, so an empty screen usually means the window rather than a quiet agent; widen it before concluding anything.
- The filters that repay the effort are Conversation attributes, which isolates messages not understood, interrupted dialogues and drop-offs, and Conversation status, which separates automated resolutions from escalations and unassisted conversations.
- Open a conversation and read it from the first message. The Details panel beside it shows how it resolved, which segments matched and the session parameters.
- Note the cause of anything that went wrong.
Two filters change what you are looking at rather than narrowing it. Conversation source separates live conversations from your own test-widget runs, which is what stops you drawing conclusions from your own testing. And Resource usage shows whether an answer came from a dialogue, a generative reply, a procedure or a knowledge answer, which is the difference between fixing an article and fixing a flow.
Everyone is called Customer here
The logs show Customer for every conversation, whether the person was signed in or not. The name is on the Zendesk ticket that came out of it, not in the log, so identifying who a conversation was with means opening the ticket beside it.
There is no filter for this, so do not go looking for one.
Which ones to read
- Handovers. Every one is a case the agent could not close, with the reason visible.
- Long conversations. Length usually means confusion.
- A random ten. The filters show you what you already suspect; random sampling shows what you do not.
Read from the start
The failure is visible at the end and caused at the beginning: an ambiguous question the agent guessed at, an article that half fits. Reading only the last exchange leads to fixing the symptom.
Write down the cause, not the case
"Article on returns does not mention the fourteen-day period" is a fix. "Agent gave a bad answer about returns" is a complaint. The first accumulates into a work list; the second does not.
Most causes are content
Missing article, outdated article, two articles disagreeing, an article assuming context. Expect the majority of your notes to be article work, and be suspicious of a review that concludes with only settings changes.
Make it a habit
Twenty conversations a week, by one person, with the notes going somewhere shared. That rhythm is what separates agents that keep improving from ones that are launched and left.
Share the good ones
Send the team an exchange where it worked well. Scepticism about AI agents is usually well founded and always corrosive, and a real transcript is the only convincing answer.
Comments
0 comments
Article is closed for comments.