> This page is for Workforce.

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.merge.dev/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.merge.dev/_mcp/server.

# Insights

> What Insights is, how Dreaming generates findings, what data it reads, the finding types, and how to act on one.

Insights reads the past few weeks of your organization's AI usage and returns a handful of findings: a frontier model carrying prompts a cheaper one already handles, a Connector nobody calls anymore, a tool sequence several people keep running by hand. Use it to decide what to change. For the raw numbers, use the [Dashboard](/merge-workforce/dashboard).

Insights is at [wf-develop.merge.dev/insights](https://wf-develop.merge.dev/insights). **View workforce** can read it. **Manage workforce** can dismiss a finding or run its fix. Nothing is applied until a person acts.

> **Enabled per organization**
>
> Insights is off until Merge turns it on for your organization. Ask your Merge contact.

## How findings are made

The analysis behind Insights is called Dreaming. It condenses each day's telemetry into a summary, then reads across those summaries to write findings. Every figure on a finding (dollars, counts, shares) is computed in code. A language model writes the headline and explanation from those figures and cannot change them.

Insights uses telemetry you already produce, so there is nothing to instrument:

| Source            | What Insights reads                                                                             |
| ----------------- | ----------------------------------------------------------------------------------------------- |
| **Gateway**       | LLM request logs: model, provider, project, tokens, cache use, streaming, routing outcome, cost |
| **Agent Handler** | Tool call metadata: the tool, its Connector, the employee, and timing                           |

It never reads tool call arguments, tool results, or prompt text, so every analysis works with tool call logging off.

A new organization starts quiet. Each analysis waits for enough usage to tell a pattern from a coincidence, so findings appear gradually.

## Finding types

| Finding                      | Category         | Detects                                                                                                                                                                           |
| ---------------------------- | ---------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Overpowered direct calls** | Optimize routing | Direct calls pinned to a model at least twice your average price, where a cheaper routed model already serves similar prompts. Flagged once the model passes 5% of Gateway spend. |
| **Idle Connectors**          | Improve security | Connectors provisioned but unused, including ones that went quiet. Recently provisioned Connectors are skipped.                                                                   |
| **Suggested skills**         | Suggested skill  | The same tool path of three or more steps, in one agent turn, run repeatedly by several people                                                                                    |
| **Flex tier candidates**     | Reduce cost      | Regular, non-streaming overnight work on the standard tier, which the flex tier serves for about half the price                                                                   |
| **Cache opportunities**      | Reduce cost      | Projects re-sending a prompt prefix of 4,096 tokens or more without cache markers, where a similar project caches it                                                              |

New analysis types ship disabled and appear once Merge enables them.

## Reading and acting on a finding

Each card shows its category, a headline, and the estimated impact: the monthly saving for cost findings, or the usage behind a suggested skill. **See details** shows the evidence and how the analysis detected it.

**Create skill** is the only fix that runs from a card. It saves the workflow as a draft in [Skills](/merge-workforce/skills/overview), and the card then reads **Skill created as a draft**. Check the draft against a few real sessions before publishing, since a path several people share can still be one team's habit.

For the other findings, you make the change:

* **Overpowered direct calls** and **Cache opportunities**: add a policy in [Routing](/merge-workforce/models/routing), or add cache markers to the project's prompts. A new routing policy applies from the moment it is saved
* **Idle Connectors**: the **Revoke access** button is not active yet. Remove access in [Connectors](/merge-workforce/connectors/overview)
* **Flex tier candidates**: the **Switch to flex tier** button is not active yet. Move one job at a time, since the flex tier trades latency for price

To **dismiss** a finding that is real but intentional, such as a pilot Connector that has not started, click the X in the card's corner. It stays dismissed on later runs.

## Next steps

#### [Dashboard](/merge-workforce/dashboard)

Check the numbers behind a finding yourself

#### [Tool calls](/merge-workforce/logs/tool-calls)

Debug an individual call behind a finding