# Human review Some decisions are best confirmed by a person. When a flow reaches a point where it needs human judgment — a low-confidence extraction, a value that failed validation, an exception that needs a person to confirm — the **Flag for Human Review** node hands the item to a reviewer. The reviewer checks the source documents, corrects what needs correcting, and submits the item, which returns it to the flow to finish. This page is the complete guide to that experience: how review work is created, how reviewers find and open it, what they can inspect and correct, and how a completed review returns the item to the flow. **Permissions.** Reviewing items is the **Reviewer** role's primary job — reviewers can open items, inspect and download the source documents, correct field values, and submit reviews, but they don't build flows or see run history. Admins have full access as well. See [Users and access](../onboarding/users-access.md). ## How an item reaches review Human review is created by a node in the flow. When you place a **Flag for Human Review** node (shown in the Nodes library as **Review**, in the Utilities category) at the point in a flow where a person should weigh in, that node pauses the item and opens a review for it. See [Utility nodes](../build-ai-agents/utility-nodes.md). The node carries an **exception reason** that tells the reviewer why the item was flagged — for example, the default *"Fields are below the confidence threshold."* — and it specifies which fields to include in the review. While an item waits for a reviewer, it sits in the **Review** status. (If a flow has no Flag for Human Review node, its items are processed without human review and complete automatically.) ## Find work in the Items list Reviewers work from the **Items** list, which shows the items produced by your flows. Each row shows the item's **Name**, an **Agent Summary**, its **Status**, the assigned **Reviewer** (or *Unassigned*), the **Flow Name**, and when it was **Created**. To focus on what needs attention: - **Filter by status** to show only items in **Review**. - **Filter by reviewer** to see items assigned to you (or to anyone), and **by flow** to focus on one AI Agent. - **Search** items — the search box accepts plain text and `key:value` tokens (for example, `flow:fnol status:REVIEW`). You can sort the columns and page through the list. An item currently open by someone else shows an **Editing disabled** badge with a lock, so you know not to wait on it. If there are no items, the list shows **No items yet** with a note that items appear once they are created. Click an item to open it. ## The item screen An open item has a header showing its name (or *Item from [flow] flow*), its **Status** badge, and — when the item has an open review — a completion bar reading **"N % complete"** that tracks how much of the review you've worked through. Below the header are two tabs: - **Insights** — an overview of the item: its key details (Item ID, Flow Name, number of Documents, Reviewer, Created date) and the agent's summary of what the item contains. Use it to get oriented before you review. - **Review** — the working surface where you inspect and correct the item's fields. A badge on the tab shows how many fields still **need review**. The actions for the review — **Save**, **Cancel**, and **Submit Item** — sit at the top right of the item screen. ## Inspect the item On the **Review** tab, the item's extracted fields are listed, organized into groups you can **Expand All** or **Collapse All**. Several tools help you find what matters: - **Field filters.** Switch between **All Fields**, **With Data**, **Needs Review**, and **Empty** — each shows a count. **Needs Review** is the important one: it lists the fields the flow flagged as needing a human, such as low-confidence or failed-validation values. - **Source filter.** When more than one document is attached, filter the fields down to a single source document. - **View Documents.** Toggle the document viewer to see the source file alongside the fields. Fields are linked to the place they came from on the page, so you can confirm a value against the original. A field's source-document badge jumps you to that document. - **Confidence.** Each field shows a confidence indicator with a tooltip (**Confidence: X%**). Values at or above the confidence threshold read as **High**; lower ones are flagged so you can prioritize them. Fields that need review are marked visually and carry a short **validation message** explaining what's wrong (for example, a value that didn't match the expected format). ## Correct the information When a review is open and you hold the item (see locking below), you can edit field values directly: - **Text fields** are edited inline in a text box. - **Choice fields** present a dropdown of the allowed options. - **Tables** are edited as a grid, cell by cell, with a popout for larger tables. As you change a value, the field is highlighted as edited and an indicator appears; hovering it shows the **original extracted value**, so you can always see what the agent first produced versus your correction. Values the flow derived — such as a record it matched — appear among the fields the same way, and you correct them in the same place. ### One reviewer at a time An item under review is **locked** to the person reviewing it, so two people don't overwrite each other: - If someone else holds the item, you'll see **"This item is currently being reviewed by [name]. Editing is disabled."** and the fields are read-only. - If you're inactive for about 15 minutes, your lock is released and a banner offers **Reclaim editing** to pick the item back up. - If the lock status can't be confirmed, editing is disabled with **"Could not verify lock status."** ## Save, submit, and return the item to the flow Two actions move your work forward: - **Save** stores your edits without finishing the review, confirming with **"Fields updated."** Use it to step away and resume later — your corrections and progress are kept. - **Submit Item** completes the review, confirming with **"Review submitted."** Submitting returns the item to the flow, which continues from the Flag for Human Review node using your reviewed values. A **Cancel** action also appears; it discards the open review and is a restricted action that isn't part of the Reviewer role. If a save or submit can't complete, a banner reports it (for example, **"Save failed"** or **"Submit failed"**). ## How a completed review appears afterward After you submit, the item carries your corrections forward: - The **item status** advances past **Review** as the flow continues, ending at **Complete** when the flow finishes its path. - Corrected fields keep both your value and a record of the **original extracted value** (shown on the edit indicator), so it's clear what was changed during review. - The **completion bar** reflects the review as worked through, and the assigned **Reviewer** is shown on the item. - In [Run history](./run-history.md), the review appears as a step in the run, so an Admin or Partner can see where the flow paused for review and resumed. Run history itself isn't available to the Reviewer role. ## Empty states and edge cases - **No items yet** — the Items list is empty until flows start producing items. - **No fields match the filters** — the Review tab shows *"There are no extracted fields matching the selected filters."* Switch the filter (for example, back to **All Fields**). - **Item not found** — if an item can't be loaded, the screen explains it may have been deleted or be outside your access. ## Where to go next - [Item status reporting](./item-status-reporting.md) — see review work and other items summarized by status. - [Utility nodes](../build-ai-agents/utility-nodes.md) — add and configure the Flag for Human Review node. - [Run history](./run-history.md) — where the review step appears in a run (Admin or Partner). - [Monitor and Review](./overview.md) — how review fits the production operating model. - [Users and access](../onboarding/users-access.md) — the Reviewer role and what it can do.