> ## Documentation Index
> Fetch the complete documentation index at: https://docs.keloa.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent performance

> How each AI agent is doing — volume handled, deflection, escalations.

When you run more than one AI agent — say a sales agent on the website and a support agent in email — the **Agent performance** breakdown tells you which one is pulling its weight, which one is escalating too much, and which knowledge needs attention.

## Where to find it

**Analytics → Agents**. The section appears on every plan; the per-agent confidence and escalation-reason histograms below are gated to **Business+** (paywall key `advanced_analytics.use`).

## Per-agent table

| Column                | What it tells you                                                        |
| --------------------- | ------------------------------------------------------------------------ |
| **Agent**             | Name and the LLM it's running on (`claude-sonnet-4`, `gpt-4o-mini`, ...) |
| **Handled**           | Conversations the agent participated in over the selected window         |
| **Deflection**        | Conversations the agent resolved end-to-end with no human reply          |
| **Escalation rate**   | Share of handled conversations that handed off to a teammate             |
| **Median confidence** | Median of the model's self-rated confidence per reply (0.0–1.0)          |
| **Credits used**      | Reply credits this agent consumed                                        |

Sort by any column. Click an agent name to filter the entire dashboard to just that agent.

## Reading the numbers

* **High volume + high escalation** — the agent is taking calls it can't close. Usually a knowledge gap. Add Q\&A pairs for the topics that escalate most ([Q\&A pairs](/knowledge/qa-pairs)).
* **Low volume + high deflection** — the agent is being conservative and only answering what it's sure about. Either fine (high precision) or the system prompt is too restrictive — review what it refuses.
* **High volume + low confidence median** — model is guessing more than it should. Tighten the system prompt's *"refuse if not sure"* line, or move the agent to a stronger model.

## Escalation reasons (Business+)

A bar chart of *why* each agent handed off:

* **Low confidence** — model fell below the confidence threshold (default 70%). See [Confidence threshold](/agents/overview).
* **Topic refused** — system prompt instructed the agent not to answer a topic.
* **Tool failure** — a tool the agent needed (Shopify lookup, Custom HTTP) returned an error.
* **Customer asked** — the customer explicitly requested a human.
* **Manual** — a teammate took over via the **Take over** button.

If **Low confidence** dominates, your knowledge base is the lever. If **Topic refused** dominates, the prompt is too narrow. If **Tool failure** dominates, check the integration's status page.

## Confidence distribution (Business+)

A 10-bucket histogram of confidence scores across every reply in the window. Healthy agents show a peak at the high end (0.7–1.0) with a long thin tail. A flat or bimodal distribution usually means the agent is being asked questions outside its knowledge.

## Per-agent conversation drill-down

Click any row → **View handled conversations**. The inbox opens filtered to that agent and date window. Read 5–10 to get a feel for what the numbers mean — analytics tells you *what*, conversations tell you *why*.

## Related

<CardGroup cols={2}>
  <Card title="Analytics overview" icon="chart-bar" href="/analytics/overview">
    The headline KPIs.
  </Card>

  <Card title="AI agents" icon="bot" href="/agents/overview">
    Create, configure, and tune.
  </Card>
</CardGroup>