Extraction & review
After Assure Pro classifies a document, the next step is extraction — pulling structured field values out of it. For a W-2, that’s the employer name, EIN, federal wages, federal withholding, and state wages. For a 1099-DIV, dividend totals. For a 1098, mortgage interest paid.
You review the extracted values side-by-side with the original document. If anything’s wrong, you correct it. Once you confirm the job, the values flow into the engagement.
Full coverage of the Documents module is in Phase 5. This page is the intake-perspective: what extraction does, how it ties to the organizer, and where to review.
Where extraction lives
| Surface | Use it for |
|---|---|
| Documents → AI Review Queue | Work through the inbound queue. New uploads land here when AI confidence is low or extraction needs human review. |
| Documents detail drawer | Click any document — the drawer shows extracted fields below the metadata. |
| Organizer review → document item | When a document is uploaded for a checklist item, the extraction status shows next to it. |
The AI Review queue
Open Documents → AI Review Queue from the sidebar. The top of the page shows three stat cards:
| Card | What it counts |
|---|---|
| Needs Review | Extractions that completed but AI flagged for human verification (medium confidence). |
| Low Confidence | Below the threshold; AI is uncertain about at least one field. |
| Failed | Extraction couldn’t complete — bad scan, unsupported document type, and so on. |
Below: the queue table grouped by client. Each row shows the document filename, document type, classification confidence, and a Review button.
[Screenshot: AI Review Queue]
Reviewing one extraction
Click any row to open the Review workspace:
┌──────────────┬───────────────────────────────────────────────┐
│ │ │
│ Client list │ Document preview (left half) │
│ sidebar │ + extracted-field overlay │
│ │ │
│ Rivera J. │ ┌─ Extracted fields (right half) ──────────┐ │
│ W-2.pdf │ │ Field │ Value │ Conf │ OK │ │
│ 1099.pdf │ │ Employer name │ Acme Inc. │ 98% │ [x] │ │
│ Patel S. │ │ Employer EIN │ 12-3456789│ 95% │ [x] │ │
│ W-2.pdf │ │ Federal wages │ 87,250.00 │ 99% │ [x] │ │
│ │ │ Federal w/h │ 12,876.00 │ 87% │ [ ] │ │
│ │ │ ... etc │ │
│ │ └─────────────────────────────────────────┘ │
│ │ │
│ │ [Reject all] [Confirm extraction] │
└──────────────┴──────────────────────────────────────────────┘Left sidebar — queue navigation
- Items grouped by client (collapsible).
- Click any item to load it into the right panel.
- After confirming one, the workspace advances to the next item.
Document preview
The PDF renders in the left half of the right panel. Detected fields are highlighted with editable boxes — rectangles showing where each value was pulled from.
Click any box to highlight the matching field in the table. Click any table field to highlight the box. You can move or resize a box if AI picked the wrong region.
Extracted-field table
| Column | What it shows |
|---|---|
| Field | The field name (for example, “Federal wages”). Fields come from the document type’s field definition. |
| Value | The extracted value. Editable inline. |
| Confidence | 0–100% for this specific field. Color-coded: green at 90+, amber 70–89, red below 70. |
| Confirm | Checkbox per field. |
You can:
- Edit a value inline. The row marks as Corrected.
- Confirm a field (check the box). The row marks as Confirmed.
- Confirm all at once via the footer action.
Footer actions
| Button | What it does |
|---|---|
| Reject all | Marks the extraction as rejected — values aren’t used. Useful when the document was misclassified. |
| Confirm extraction | Confirms all fields and closes the job. Marks the document fully reviewed. |
| Retry (failed only) | Re-runs the extraction. Useful when the original failed for a transient reason. |
After confirming, the workspace advances to the next item in the queue.
Confirming many fields at once
In the workspace, you have two shortcuts for power users:
- Confirm all — confirms every field on the current job in one click.
- Reset confirmations — clears all confirmations on the current job so you can start over.
Both mirror what the per-row checkboxes do.
How corrections improve the model
If you change a value, Assure Pro records both the original AI answer and your correction. Over time, this builds a labeled dataset the AI can learn from for future extractions on your firm’s documents.
Corrections are firm-scoped. Your firm’s history doesn’t leak to other firms.
Fields per document type
Each document type has a defined field set. The default field set for W-2 is: employer name, employer address, employer EIN, employee SSN, federal wages, federal withholding, social security wages, social security withholding, medicare wages, and state info.
You can add or remove fields per document type at Settings → Document Types → (any type) → Fields. The classifier and extractor try to find every field you’ve defined.
For document types with no field set (like Bank Statement or Other), extraction doesn’t run — there’s nothing structured to pull.
Confidence thresholds
Three thresholds affect behavior:
| Threshold | What it does | Default |
|---|---|---|
| Auto-match minimum | Confidence required for the document to auto-match a checklist item. | 90% |
| Review recommended | Below this, the document lands in the AI Review Queue. | 70% |
| Per-field warning | Below this for a single field, the row is highlighted amber. | 80% |
You can change these at Settings → Documents → AI confidence thresholds (covered in Phase 7).
What extraction doesn’t do
- It doesn’t push field values into tax software (Drake, ProConnect, UltraTax). You export manually.
- It doesn’t validate values against IRS rules — that’s the tax prep step.
- It doesn’t reclassify the document. If the document type is wrong, reclassify first (see Document classification) and the extractor re-runs against the new type’s field set.
Tips
”The extractor missed a field that’s clearly on the document”
When AI isn’t confident, it returns nothing rather than guess. Click into the field, type the value, and mark it confirmed. Your correction is recorded.
”It pulled the wrong year’s W-2”
Year detection is part of classification, not extraction. Re-check the classified document. If AI picked the wrong year — for example, the document is a 2024 W-2 but values came from a 2023 side annotation — reclassify or click Reject all and reclassify.
”I want to skip extraction for some document types”
Open Settings → Document Types → (the type) → Fields and remove all the fields. Extraction will skip these documents from then on.
Permissions
| Action | Who can do it |
|---|---|
| View AI Review Queue | Anyone with View extraction access |
| Confirm or correct fields | Edit extraction access |
| Retry an extraction | Create extraction access |
| Edit field sets per document type | Edit firm settings access |
Next
- Document classification — the step that precedes extraction.
- The document checklist — how extracted documents close out checklist items.
- Documents module — full coverage in Phase 5.