Plan requirement
| Subscription | Suite Team or higher, Support Team or higher |
| Access | Admin |
The replies the agent uses when things go wrong. Nobody reviews them, and they are what customers see at the worst moment.
Find and edit them
- In the AI agents workspace, select the AI agent you are working on.
- Click Content in the sidebar and select Use cases.
- Click the Categories filter, select System replies and click Apply.
- Click a system reply, then open the Replies tab. On a zero-training agent, click Edit dialogue; on an agentic agent, choose Dialogue or Procedure in the reply method column.
- Change the wording, and test by deliberately triggering each one.
Every advanced AI agent gets these automatically. Nobody writes them for you, though: the defaults are generic until you edit the dialogue behind them.
What they cover
| Welcome reply | The greeting at the start of the conversation. |
| Default reply | No use case matched and no generated answer could be produced. |
| Knowledge reply | No use case matched, the question is a real question, so the agent searches your knowledge sources. Agentic agents. |
| Generative replies reply | The same situation on a zero-training agent, answered from the knowledge you imported. |
| Escalation reply | What it says while handing the conversation to a person. Agentic agents. |
| Failed escalation reply | What it says when the handover does not work. Not used on email. |
| Technical error reply | Something went wrong on our side. |
| No message | The customer sent nothing, or nothing usable. |
| Unsupported language reply | The language was not recognised. |
| Collect BSAT reply | The feedback question at the end of the conversation. |
Why they matter more than the rest
Every other reply is read by somebody being helped. These are read by somebody who is not, and that is when tone either salvages the exchange or ends it.
The default texts are serviceable and generic. Twenty minutes rewriting them is the highest-value twenty minutes in the whole configuration.
The technical error reply deserves the most care, because it is used for more than errors. An AI agent answers only from the sources you connect it to, so a question outside them, including a plain general-knowledge one, has nowhere to go and can fall back here. A customer who asked something harmless then reads that something went wrong on your side. Wording that covers both cases, an error and a question you do not answer, is the safer default.
Each one should offer a next step
"I did not understand that" is a dead end. "I did not understand that. Try telling me in a few words what it is about, or say 'agent' and I will pass you to somebody" is not.
Do not apologise repeatedly
One apology, then the route forward. A reply that is mostly regret reads as evasion, particularly to somebody who is already annoyed.
Be honest about waiting
If handover means a wait, say roughly how long. A customer told it may take a few hours will wait; one told nothing will send three more messages.
Trigger every one
Deliberately, in a test conversation. These are the replies you never see by accident, which is why they are the ones still carrying defaults two years on.
Comments
0 comments
Article is closed for comments.