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

# Findings

> Detect recurring problems in your conversations and investigate them with AI

Findings reviews completed WhatsApp conversations with AI, records what it sees as structured events, and groups the recurring ones into findings you can investigate.

## How it works

1. A managed workflow evaluates each conversation when it ends and emits [project events](/docs/platform/events) such as `conversation.user_frustrated`.
2. Kapso compares the last 7 days against the previous 28 and opens a finding when a signal rises, falls, or keeps recurring.
3. The Kapso Agent Investigator reads the evidence behind the finding, inspects your workflows, executions, logs, and connected repositories, and reports the cause with suggested fixes.
4. You apply a fix and mark the finding as addressed. Kapso then watches new conversations to see whether the problem comes back.

## Turn it on

Open **Findings** in the sidebar and complete the setup. It creates two components in your project:

* **Default event evaluator** — a workflow triggered by `whatsapp.conversation.ended` that reads the transcript and emits events. It never replies to the customer and takes no outbound action.
* **Kapso Agent Investigator** — a Kapso Agent mode that investigates a finding when you start an investigation.

During setup you can also review up to 50 conversations from the last 7 days, so you get findings without waiting for new traffic.

Both components use AI credits when they run. Findings needs project events on your plan, and only evaluates conversations on production numbers.

## Default events

The evaluator emits these events. Each one carries a short `reason` property with the evidence for it.

| Event                                            | Emitted when                                                                               |
| ------------------------------------------------ | ------------------------------------------------------------------------------------------ |
| `conversation.user_frustrated`                   | The user clearly expressed frustration, dissatisfaction, or loss of confidence             |
| `conversation.positive_outcome`                  | The user said the problem was solved, or expressed gratitude                               |
| `conversation.agent_response_wrong`              | The agent made a specific wrong claim and the user corrected it                            |
| `conversation.unresolved`                        | The conversation ended without answering the user's request or giving a concrete next step |
| `conversation.security.prompt_injection_attempt` | The user tried to override agent instructions or manipulate the runtime                    |

A conversation can emit zero, one, or several of them. The evaluator optimizes for precision: with ambiguous evidence it emits nothing.

The evaluator treats the transcript as untrusted data and never follows instructions found inside it.

## How findings are detected

Kapso refreshes findings once a day. It compares a **current window** of the last 7 days against a **baseline** of the 28 days before it, counting distinct conversations.

| Direction   | Meaning                                                                         |
| ----------- | ------------------------------------------------------------------------------- |
| `rising`    | The signal happens in a larger share of conversations than the baseline         |
| `falling`   | A positive signal happens in a smaller share of conversations than the baseline |
| `recurring` | The signal keeps appearing, with no baseline to compare against                 |

A rising finding needs all of:

* At least 30 evaluated conversations in both the current and the baseline window
* At least 5 affected conversations
* An affected rate of at least 10%
* A rate at least 5 points above the baseline
* Evidence on at least 2 different days, with something in the last 2 days

A falling finding needs a rate at least 5 points below the baseline on at least 2 days, and a shortfall of at least 5 conversations.

When there isn't enough history for a baseline, a signal seen in at least 5 conversations over the last 2 days opens a recurring finding instead. Prompt injection attempts are always tracked as recurring.

Recurring findings have no baseline comparison. What matters is that the problem keeps appearing.

A finding that stops qualifying goes quiet and leaves the list: after 7 days without new qualifying evidence, or 14 days once it's under investigation or monitoring.

## Investigate with AI

Open a finding to see the evidence behind it: the affected conversations, the events that qualified it, and the workflow they run through.

Click **Investigate with AI** to run the Kapso Agent Investigator. It reads the finding evidence, compares affected conversations against successful ones, follows them into workflow executions and execution events, searches your logs, and inspects any repositories connected to the project.

The investigation is read-only. It never messages customers, changes configuration, edits repositories, or starts workflows.

When it finishes, the finding shows:

* **Summary** — what is broken and why
* **Causes** — each with an explanation, a confidence of low, medium, or high, and evidence references
* **Suggested fixes** — at most three, each with the target, the change, and how to verify it
* **Coverage limitations** — what the investigator could not verify, so you can weigh it before changing anything

Evidence references link back to the source: the conversation, the workflow, the execution, the project events, or the log entry.

An investigation that fails or times out can be retried, up to 3 attempts for the same evidence. A new investigation also becomes available when new evidence arrives after the last one ran.

## Act on a finding

**Mark as addressed** after you apply a fix. Kapso records the current metrics as a baseline and starts monitoring:

* If the signal doesn't qualify again, the finding resolves.
* If it qualifies again, the finding is marked as not improved and comes back.

You can only mark a finding as addressed once an investigation has completed and covers the current evidence.

**Dismiss** a finding you don't want to see. A note is required. If monitoring later detects the problem again, the finding can reopen.

## Custom events

Findings works on your own [project events](/docs/platform/events), not just the default ones. Any active event definition with recent data is included in the daily refresh.

Events linked to a conversation are measured as a rate over conversations. Events without a conversation are measured by volume, and qualify when the current window exceeds the expected count by at least 50% and 5 events.

## From the Kapso Agent and MCP

Findings are also available to agents through the `findings` tool on the [Kapso MCP server](/docs/whatsapp/mcp), with the actions `list`, `get`, `read_evidence`, `start_investigation`, `dismiss`, and `mark_addressed`.