Quality Review is the queue where AI results stop being a black box. It shows dossiers whose automatic processing has finished and which need human sign-off before they move to an external operation or to completion.
The idea is simple: AI reads, extracts and classifies; a person steps in when the process says so or when the reading itself is not safe enough to continue on its own.
Why it exists
A document can be read automatically and still need a human decision. The policy number may be right, but the document may be the wrong type. The date may have been extracted with low confidence. An attachment may hit a rule that, for that process, requires review.
Quality Review avoids two extremes: making people review everything, which cancels out the automation, or letting uncertain values flow into decisions and external systems with no sign-off.
How a dossier gets here
There are two routes:
- The process requires review. If the process has Requires quality review turned on, dossiers for that process stop in Quality Review after automatic processing.
- AI raised an exception. Even when global review is off, a dossier can stop here when at least one attachment needs a human decision — for example, unexpected document type, confidence below the configured threshold, missing confidence or a result that should not advance by itself.
A technical failure is not Quality Review. If the AI service, OCR, timeout or integration fails, the dossier goes to Failed, where it can be recovered or repeated through the right operation.
What the queue shows
The queue lists the dossiers in Quality review that you are allowed to see and review; the menu entry itself only appears to people who hold that permission. At the top you see three figures:
| Awaiting review | How many dossiers are queued |
| Low confidence | How many have results the AI marked as uncertain |
| Longest wait (days) | How long the oldest dossier has waited |
Rows show the dossier, the process, the analysed files, the lowest confidence found and how long it has been waiting.
Lowest confidence matters more than the average. One badly read field can be exactly the point that compromises the decision.
Reviewing attachments
Inside the dossier, each analysed attachment shows the AI result and the effective review status.
- If AI raised no exception, the result can be treated as automatically approved for the queue.
- If AI raised an exception, a person must decide that attachment.
- A rejected attachment needs a clear reason, because that reason guides the correction.
The human decision always wins over the AI suggestion. AI can propose; the person decides.
The three dossier decisions
Approve
Use Approve when the results that needed validation are correct. Before the review closes, every attachment that requires a human decision must be decided.
After approval, the dossier follows the process design: if an external operation is configured, it enters Triggering; otherwise it becomes Completed.
Return for correction
Use Return when attachments are wrong, incomplete or unsuitable and somebody needs to correct them. The reason is mandatory and should say what to do, not just that something is wrong.
The return moves the dossier to Returned and carries the correction scope with it: the rejected attachments. That tells the person correcting it exactly where to act.
Reject
Use Reject when the AI result should not be used and this is not just an attachment correction. It also requires a reason. The dossier becomes Failed, with the reason visible for recovery or analysis.
What to check
- Start with attachments that have exceptions or the lowest confidence.
- Compare the extracted value with the document, not only with what seems plausible.
- Confirm that the document matches the expected type.
- When you reject something, write an actionable instruction: "the expiry date was read from the wrong page" helps; "wrong" does not.
When the queue is empty
It may mean nothing has finished processing, everything has already been reviewed, no process requires global review, or AI found no exceptions that need a person. That is the intended design: the queue exists to show work that needs judgement, not to force permanent manual review.
What you should have now
- A process with Requires quality review turned on, and the reason it was that one.
- A dossier that stopped in the queue on an AI exception, with global review off — and the exception read on the attachment.
- A dossier approved from the queue, and where it went: Triggering or Completed.
- One returned, with a reason written as an instruction and the correction scope on the rejected attachments.