When to use review
- You want a human gate before delivering results to a downstream system. Connect the node directly: reaching it pauses the run.
- You only want humans involved on uncertain runs. Route into the node with a condition edge on the upstream signals:
{{metadata.reviewAdvice.needsReview}}is true (extract’s Review advice),{{metadata.unmatchedFields}}is not empty,{{metadata.contestedFields}}is not empty, or a field’s{{metadata.fieldConfidence.<field>}}below a value you choose. - You want a fallback path when validation fails on the
Warnaction: route on{{metadata.validationFailures}}is not empty into review for correction. - You want to catch numeric disagreements between two sources. Add an upstream Reconcile node and route its mismatches into review.
- Skip review when an automated validation check or transform can fix the data deterministically. Reserve human time for cases that genuinely need judgment.
Configuration
The node has no trigger setting of its own: a run pauses whenever it reaches the node, and the routing that decides which runs reach it lives on the incoming edges.Reminders nudge the assigned reviewer while the task is still pending. If the task is unassigned (anyone can claim), no per-person reminder is sent.
Inputs and outputs
Allowed inputs: Extract, validation, reconcile, if, switch, HTTP, loop, variable, store, prompt, call workflow, the connector nodes (Business Central and QuickBooks in any operation, and callback), and the sub-workflow trigger (review the caller-supplied data directly). Maximum one input connection. Review pauses on extracted data, so it sits after the node that produced that data, not directly after filter pages or split. Output: Approved (and possibly edited) data, passed to the next connected node.Common pitfalls
Review connected without a condition on a high-volume workflow
Review connected without a condition on a high-volume workflow
Every run that reaches the node pauses. On a busy workflow this floods the review queue and stalls throughput. Put a condition edge in front of it: extract’s Review advice, unverified or contested fields, or validation failures.
Unverified fields mistaken for a confidence gate
Unverified fields mistaken for a confidence gate
Unverified means grounding, not certainty. A field is unverified when the matcher could not place its value on the page, so a field can carry a high confidence score and still be unverified. Route on
metadata.fieldConfidence when you want a score, or turn on Review advice to let the extraction check weigh both.No fallback when reviewers reject a run
No fallback when reviewers reject a run
A reviewer can mark a run failed. Make sure the failure path is intentional: either delete the document, route to a fallback workflow, or notify a system. Don’t let rejections sit silently.
Related
Human review guide
Set up and manage human review workflows
Review queue
Learn how the review queue works
Extract action
Extract data before review
Transform action
Transform data after review
Reconcile action
Compare two sources before review