AI does the reading. People check what matters.
The Review Queue shows your team only the fields the model is unsure about, side by side with the cropped part of the original document. One keystroke confirms, typing corrects, and the record moves on.
Fields below your confidence threshold, PO mismatches and failed checks from every reader.
A reviewer sees the value and the document crop together, then confirms or corrects.
A fully trusted record, with who checked what and when written to its history.
Review the field, not the whole document.
Most pages need nothing from a person. The ones that do usually need one or two values, so that is all the queue shows.
Crop beside value
The exact region of the page appears next to the extracted value. No hunting through a PDF.
Keyboard first
Enter confirms, E edits, S skips, J and K move between fields. Reviewers rarely touch the mouse.
Thresholds per field
Set a higher bar for totals and GSTINs than for descriptions. Different rules per document type.
Assignment
Route by document type, vendor, branch or language to the people who know them best.
Full history
Original value, corrected value, reviewer and time are kept for every field.
Learns your vendors
Corrections on a vendor’s layout improve future reads for your account only.
How a field gets checked
Score
Every extracted value receives a confidence score between 0 and 100.
Hold
Values below the threshold, or failing a check, are held in the queue.
Confirm
A reviewer confirms or corrects with the document crop in view.
Release
Once every held field is resolved, the record continues to approval or export.
Nothing uncertain posts on its own.
The Review Queue is where the system admits what it does not know. These rules apply across every reader.
Confidence on every field
Each extracted value carries a score. Anything below your threshold goes to a person before it is used.
Consent and masking
ID documents are processed only with the person’s consent. ID numbers are masked by default and access is restricted by role.
Encrypted and time-limited
Documents are encrypted, kept only for the period your company sets, then deleted.
Never shared for training
Customer documents are never used to train models for other customers.
The details
How the queue is configured.
- Thresholds
- Per field and per document type, from 50 to 100
- Reasons held
- Low confidence, failed validation, PO or GRN mismatch, duplicate suspicion, balance break
- Assignment
- By document type, vendor, branch, language or round robin
- Shortcuts
- Enter confirm, E edit, S skip, J/K next or previous field
- Audit
- Original value, final value, reviewer, timestamp and reason on every field
Review Queue questions
How many fields usually go to review?
Can two people review the same field?
Do corrections train a shared model?
Next on the table
See how little actually needs a person.
Process 500 of your own pages and count the fields that reach the queue. That number is the real measure of the time you get back.
- Stained or folded
- Handwritten notes
- Skewed scans