# Runs & approvals

> Reads run free; every write pauses with the exact action shown. Approve it, or reject with a reason the AI adapts to. Nothing changes silently.

The back office's trust model is one sentence: **nothing changes silently**. Reading is free — the AI can query, list, search, and summarize without asking. The moment a step would write — send an SMS, update a CRM row, create anything in a connected account — the run parks in *waiting for approval* with the exact action displayed.

## The approval moment

- You see **precisely what will happen** — the recipient and full text of the message, the record and the field values, the API call and its payload. Not a summary: the action.
- **Approve** resumes the run from that step.
- **Reject** is an instruction, not a dead end — add a reason ("too pushy, mention the quote number") and the AI revises its plan and continues.
- Everything is kept: the run timeline shows each step, each approval, and who clicked it.

## What counts as a write

Classification is conservative: integration actions are treated as writes unless they're clearly reads (get/list/find/search). A new connector's ambiguous action pauses rather than runs — the safe default when real accounts are attached.

## Reports

Every run ends with a report — findings, actions taken, actions skipped — delivered in the run view. For [scheduled runs](/docs/backoffice/schedules) the report is the product: the Monday email that used to take an hour of tab-hopping.

> **WHY PAUSE-ON-WRITE SCALES TRUST:** Week one you approve everything and learn how it thinks. Week four you're approving batches in seconds because the actions have been right. The trail is always there when a client asks "who sent this?"