Skip to main content

Documentation Index

Fetch the complete documentation index at: https://explore.airia.com/llms.txt

Use this file to discover all available pages before exploring further.

​
Overview

The Form Review Step pauses your agent and presents reviewers with a document-aware split view: the source document on the left and an editable, AI-prefilled form on the right. Reviewers verify, correct, and submit the form. The submitted values become the step output, and the workflow continues down the branch matching the action the reviewer chose. Use this step whenever an AI model extracts data from a document and a person needs to confirm that data is correct before it goes to a downstream system.

​
How It Works

The most common setup is a three-step sequence:
  1. A document is passed into the workflow as an input.
  2. A model step with structured output extracts data from that document.
  3. The Form Review Step pauses execution so a reviewer can verify the extracted data before the workflow continues.
When the agent reaches the Form Review Step, execution pauses and reviewers are notified (in-app and, optionally, by email). They open Productivity > Approvals, see the split view, verify and correct the AI-prefilled form, and click an action button. The edited values become the step output and the agent continues.

​
How to Use It

  1. Add the Form Review Step In Agent Studio, find the Form Review Step in the Steps menu and drag it onto the canvas.
  2. Connect a model step Wire the output of an upstream model step (with structured output enabled) to the Form Review Step. Fields are generated automatically from the model’s output schema: objects become groups, arrays become repeatable rows, and primitive values become inputs.
  3. Assign reviewers In Form Review Settings, pick specific users, user groups, or both. Email notifications are optional. In-app bell notifications are always on.
  4. Set the source document Point the step at a file produced earlier in the workflow. PDF, Word, and image formats are supported. If no document is specified, reviewers see the form without a source panel.
  5. Configure action buttons Approve and Reject are included by default. Add any additional buttons your process requires (Request Changes, Escalate, Route to Legal, and so on). Each button creates its own branch on the canvas, so different answers lead to different downstream paths.
  6. Publish and test Run the agent. When execution reaches the Form Review Step, a notification is sent to the assigned reviewers and the run pauses until they act.

​
Reviewing a Request

Reviewers open the request from Productivity > Approvals or from the notification they received. The interface shows:
  • Left panel: The source document, with scroll, zoom, and full-screen controls.
  • Right panel: The editable form, pre-filled with the values the AI extracted.
Reviewers can edit any field directly, expand array fields (like a list of beneficiaries) to add rows when the document contains more than the model caught, and leave a comment for the requester. When they click an action button, the form locks to read-only, the workflow resumes, and the audit log records who acted, when, and what changed.

​
Configuration Reference

​
Fields

By default, fields are generated from the connected model step’s structured output schema. To define fields manually, enable Custom Fields in Form Review Settings and add each one with a label, a type (text, number, date, dropdown, list, or yes/no), and an optional required flag.

​
Default Values with Scriban

When Custom Fields is enabled, each field’s default value slot accepts a Scriban expression. The autocomplete opens on {{ and resolves against the full execution context: the upstream model output, earlier step results, workflow inputs, prompt variables, and the standard Scriban library for strings, dates, numbers, and collections. For example:
  • {{ Steps.Extract.output.trustor.name }} pulls the trustor name directly from an upstream extraction step.
  • {{ Execution.now | date.to_string "yyyy-MM-dd" }} normalizes the current date.
  • {{ Inputs.requestedBy ?? "unassigned" }} falls back to a placeholder when no requester is set.
This is the same expression language used throughout Agent Studio.

​
Action Buttons

Approve and Reject are always available. Any additional button you define creates a new branch on the canvas. Name them to match your process. Common examples include Request Changes, Escalate, and Route to Legal.

​
When to Use It

​
Document Processing With Required Human Verification

Any workflow that extracts structured data from a PDF, Word document, or image before writing it to a downstream system where accuracy matters. Use Case Examples:
  • Mortgage and deed processing: verify extracted names, dates, and amounts before recording
  • Vendor contract intake: confirm extracted terms before creating a purchase order
  • Insurance claims: review extracted field values before initiating a payout

​
AI Extraction That Needs a Sign-Off

When a model step produces a structured output and the data needs a person to confirm it was correct before it moves downstream. Use Case Examples:
  • Medical records digitization: a clinician verifies AI-extracted diagnoses
  • Legal document review: a paralegal confirms extracted parties and dates
  • Financial data entry: an accountant approves AI-pulled figures before posting