AI Agents & Automation
Observability of AI Agents
Which tools an AI agent called, with which data and with which result
Tip
Every action of an AI agent is traceable in the ticket: which tool was called, with which input, which output came back and whether the call was successful. This makes it possible to check why an AI agent made a decision, and to pinpoint errors in prompts, tools or connected systems.
Where this information can be found depends on the channel:
| Channel | Where the tool calls can be seen | Display |
|---|---|---|
| Email and letter | In the resolution plan in the middle column of the ticket view | Agent steps list with status, duration, input and output per call |
| Chat and phone (voicebot) | In the conversation history of the ticket, at the point in the conversation where the tool was called | Collapsed system messages with request and response |
Email and letter
For emails and letters, the smart AI agent creates a resolution plan for the ticket. The plan consists of individual steps – such as a deposit change or new bank details – and a prepared reply to the customer. How the plan and its steps change over time is described in What happens to an intent.
The resolution plan is located in the middle column of the ticket view, in the AI agents area below Re-run AI agents.
Reading the plan at a glance
The header of the plan shows its name and a counter:
- "3 / 4 ready" – this many steps are ready to be executed. The remaining ones still need a decision.
- "All resolutions applied" – the plan has been executed.
- "Closed without reply" – the plan has been executed and the ticket closed without a reply to the customer.
Each step carries a coloured label with its state:
| Label | Meaning |
|---|---|
| Ready to confirm | The step is complete and will be executed with the plan. |
| Plausible | The values have been checked and appear consistent. |
| Needs review | The AI agent is not certain – an employee should look at the step before it is executed. |
| Ask customer | Instead of executing the step, the reply asks the customer a follow-up question. |
| Skip | The step will not be executed. |
| Applied | The step has been executed. |
| Clarification requested | Instead of executing the step, the customer was asked a follow-up question. |
| Failed | The execution failed, for example because the connected system reported an error. |
| Skipped | The step was not executed. |
A step that needs attention is expanded automatically. Any other step can be opened with a click.
Why the AI agent suggests a step
An expanded step shows what it is based on:
- Based on – the passage from the customer message from which the step was derived. The values the step will write are highlighted in it.
- Resolution option – the selected option. The option preferred by the AI agent is marked Recommended; as long as the step has not been executed, it can be changed. After execution, this shows the option that was actually applied.
- Notifications – messages from the AI agent about this step, such as check results or warnings.
- Not validated – a value still has to be confirmed before the plan can continue.
The prepared reply to the customer is shown below the plan. Until the plan is confirmed, it carries the label Not applied yet.
The tool calls: Agent steps
Below the steps is the collapsed list Agent steps · N tool calls. It shows, in chronological order, every tool call the AI agent made while creating the plan – including those that only read data and therefore do not produce a step of their own in the plan.
Each row shows:
- Name of the tool – for a rule-based AI agent used as a tool e.g. "Deposit Agent executed", for integrated and user-defined tools their label
- Duration of the call
- Status: a check mark on success, Staged for an action that is only executed when the plan is confirmed, or Failed
Clicking the arrow of a row opens the details:
| Field | Content |
|---|---|
| Tool | The technical name of the tool, as it appears in the configuration |
| Input | The parameters the AI agent called the tool with, as an expandable JSON structure |
| Output | The result the tool returned. Very long results are shortened and marked Output truncated. |
Note
Staged is the core of the resolution plan: reading tools (e.g. retrieving contract data) are executed by the AI agent immediately. Tools that change something (e.g. changing a deposit) are only staged and executed when the plan is confirmed – manually or through automation.
Answering typical questions
Chat and phone
In chat and on the phone, the AI agent works within the ongoing conversation. Its tool calls are therefore written directly into the conversation history of the ticket – exactly at the point where they happened in the conversation. This makes it possible to follow, message by message, what the customer said, which tool the AI agent then called and which result it received before replying.
What the entries look like
Each tool call appears as two entries titled System message, shown as a thin divider between the messages:
- Request – the tool called and the parameters the AI agent called it with
- Response – the result of the tool, or the error message if the call failed
The entries are collapsed. Clicking the arrow opens the content as a JSON structure. Chat and phone look the same here; for phone calls the entry appears in the call transcript.
Customer: I would like to lower my deposit to €80.
─── System message ──────────────────────────────── ▸ (Request: tool + parameters)
─── System message ──────────────────────────────── ▸ (Response: result)
AI agent: I have adjusted your deposit to €80 as of 1 November.Who sees the entries
- Employees see the system messages in the ticket view.
- Customers do not see them – they neither appear in the chat window nor are they read out.
- When a ticket is forwarded, they are not included.
After the call
After a phone call, Enneo adds the voicebot's tool calls to the ticket as AI agents. This also makes visible in the AI agents area which actions were executed during the conversation.
Answering typical questions
Permissions
- The system messages in the chat and phone history can be seen by anyone allowed to open the ticket.
- The full list of Agent steps additionally requires the permission to view Neo sessions. Without it, the resolution plan only shows the steps stored in the plan.
- The names of the AI agents in the agent steps only appear if the role is allowed to list AI agents.
Roles and permissions are managed under Roles and permissions.