> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.cloudraker.com/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.cloudraker.com/_mcp/server. # Review AI work When Florent reads a document and writes values onto a record, every value arrives **unverified**. Reviewing AI work is how you check each value against its source and decide, one value at a time: this is correct (**Approve**), or this is wrong (**Reject**). Florent drafts; you decide what becomes trusted data. This page covers two things you do constantly: verifying a value against its source with a **citation**, and locking it in with **field review**. Reviewing AI values is a permission. **Admin**, **Manager**, and **Builder** can approve and reject; a **Member** sees the values and their citations but the controls are read-only. See [Roles and permissions](/admin/roles-and-permissions). ## Two gates, don't confuse them You meet two different "approval" moments around AI work, and they answer different questions. * **Task approval** — "May these drafted records be written at all?" A one-time gate inside a [run](/projects/run-a-playbook): an approver lets the drafted values land. This page does not own that model — see [Assignments and approvals](/playbooks/assignments-and-approvals). * **Field review** — "Is each individual value human-verified?" Ongoing, per-value state on the record, done **after** the values have landed. This is what this page is about. The normal sequence: the task is approved, the values land on records as **Pending review**, then you approve or reject each value. Approving the task does not approve the fields, and approving the fields does not approve the task. ## Verify before you approve: citations Every value Florent extracts carries a **citation** — a pointer back to exactly where it came from: the file, the page, the highlighted region on that page, and the quoted text the AI relied on. Click a value to see the original passage instead of taking Florent's word for it. Two small chips sit next to each AI value: * **The confidence chip.** A small badge that rates how sure Florent is about the value (an icon out of 5 — green for 4–5, amber for 3, red for 1–2). Hover it to see the score and the AI's reasoning in a tooltip. For an empty value, the chip reads either **Confirmed absent** (Florent is confident the field genuinely isn't in the document) or **Not found — review** (Florent couldn't find it and you should check). You may also see **Scoring…** while scoring is in progress, or **Not scored**. * **The source-count chip.** A small quote-mark badge with a number — how many citations back this value. A value with one citation opens its source directly; a value with several opens a short **Cited sources** list first so you pick which one to inspect. The reasoning in the hover tooltip lets you tell a confidently-absent field from a missed one **before** you even open the source. A **Confirmed absent** value is usually safe to approve; a **Not found — review** value is worth opening. ### What the source viewer shows Click a citation to open the cited source. Press the confidence chip or the source-count chip on the value. The source viewer opens — filling the document pane in the record panel, or as a modal titled **Source for \{field}** from a standalone chip. The original file renders inline, scrolled to the cited page (shown as a **Page N** badge) with the cited region highlighted. Read the **Cited text** banner underneath — the exact quote Florent used — and check that the highlighted passage actually says what the value claims. The viewer repeats the confidence chip and shows the AI's reasoning for the score (or, for an empty value, why it judged the field absent). A **Download** button pulls the original file if you want to open it yourself. For a transcribed recording, the viewer shows a **Transcript** pane with speaker and timestamp lines instead of a PDF, scrolled to and highlighting the cited passage. The header reads **Transcript** with the recording's duration. A non-PDF, non-transcript source can't be shown inline. The viewer reads **This source can't be previewed here** and prompts you to use **Download** to open it. You still get the quoted **Cited text**. If the cited file was deleted from the project, the viewer reads **This field cites a source file that couldn't be found.** The value is still there; only the linked file is gone. ## Field review: approve, reject, lock Once you've verified a value, you record your decision. Each AI value shows a status badge that doubles as the menu trigger. | Badge | Meaning | Locked? | | ------------------ | -------------------------- | ------- | | **Pending review** | Awaiting your decision | No | | **Approved** | You confirmed it's correct | Yes | | **Rejected** | You flagged it as wrong | Yes | A fresh AI value lands as **Pending review**. You pick **Approve** or **Reject**, and the value locks. To re-open a locked value, use **Dismiss**, which returns it to **Pending review**. Dismiss is the only way to unlock. **Approved** and **Rejected** both lock the value — they protect it identically and differ only in meaning. **Rejecting does not delete or blank the value**; it flags it as wrong and surfaces it as follow-up work in [My work](/work/my-work). To actually fix a wrong value, **Dismiss** it first, then edit it (see [The project Data tab](/projects/data)). ### Review one value at a time Open a record from the [records table](/projects/data) to slide out the record panel. Each AI-written field shows its citation chip and its review status badge next to the label. Use the citation chip to check the value against its source, as above. Press the status badge to open its menu — **Approve**, **Reject**, **Dismiss**. The item matching the current state is greyed out (when a value is already **Approved**, **Approve** is disabled). Only **AI-written fields** carry review controls — a value has something to verify only when there's a citation behind it. A value a person typed by hand shows no review badge. If a field you expected to review has no badge, it was entered by a person, or Florent didn't ground it. ### Review a whole record or a whole column at once At the bottom of the record panel, **Approve all on this record** approves every AI-written value on that record in one click. It appears only when you have review permission and the record actually has AI fields. In the records table, the column header for an AI-written field has a small dropdown (a down-arrow). Its menu — **Approve this column**, **Reject this column**, **Dismiss this column** — reviews that one field across every record in the current table that has an AI value for it. Use it to bulk-approve a field you trust across a whole import. ## What locking protects against Approving or rejecting a value locks it, which matters in three concrete ways. If a playbook re-runs and would have updated that field, it skips your locked value — neither the value nor its citation changes. Approve the values you trust to freeze them before any re-run. When Florent proposes changes to an existing record, it strips out any locked fields from its proposal and tells you which fields it left untouched. A direct edit to a locked field is refused with a message telling you to **Dismiss the review first**. There's no silent overwrite path — unlocking is always deliberate. A run can finish (**Done**) with values still **Pending review** or **Rejected** — review is decoupled from run completion. Don't expect the run to wait for you. Chase outstanding review from [My work](/work/my-work) instead. ## Reviewing inside a run When a [Document or AI task](/playbooks/task-types) drafts records, the approver opens a **Drafts to review** panel inside the run before anything is written. Each drafted record shows its fields, the confidence chips, and citations — and a per-draft state of **Ready** or **Needs changes**. Use the citation chips to confirm the drafted values against their sources, exactly as on a record. Edit a draft and **Save changes**, or **Discard** a draft you don't want. A draft marked **Needs changes** must be fixed or discarded before the rest can go through. Partial approval is allowed — some rows save, the others stay parked with a per-row reason. **Approve values** writes the drafts into the record type as records. **Reject** sends the task back — rejecting requires a reason, so type why before you confirm. For generated documents, the in-run panel reads **Documents to review** — open each one to confirm it reads correctly, then **Approve documents**. Approved documents become part of the [project's files](/projects/files). Approving the task is one gate; the values it writes still land as **Pending review** for field-level review. The rules for who can approve a task — including the assignee's manager — live in [Assignments and approvals](/playbooks/assignments-and-approvals). ## Reviewing a batch When an action runs over many files or records at once, every result is reviewed in a grid rather than one record panel — same citations, same Approve/Reject, shown in a two-column view per row. That flow is covered in full on [Batch action runs](/actions/batch-runs). One difference to remember: rejecting a **field** never deletes anything, but rejecting a generated document in a batch run **deletes** the generated file. See [Batch action runs](/actions/batch-runs). ## Where to go next Who can approve a task, and when approval is forced. Run an action over many items and review every result in a grid. Browse, filter, and edit the records your reviews produce. How Document and AI tasks draft the records you review.