In a typical accounts payable (AP) setup, putting OCR or AI extraction in front of an ERP shifts manual work to exception handling. When a parser flags an uncertain total or missing purchase order (PO), many teams route a text summary into email or chat with Approve and Reject buttons—leaving reviewers unable to verify figures against the PDF or fix a misread digit without rejecting the bill or re-keying it in the ERP.
We recommend placing a human review interface for invoice extraction before financial approval. Its job is narrow and strict: show four linked elements on one screen—the highlighted source excerpt, the extracted value, the validation error that blocked automation, and an editable correction that re-runs deterministic checks before sign-off.
Drawing on third-party docs (not Insonl clients) from AWS’s Amazon Textract best practices and API references alongside Microsoft’s Power Automate approvals documentation, this post defines what the exception screen should show and where generic approval tools stop.
Separate extraction exception review from financial approval
We recommend separating extraction review from financial sign-off. Financial approval asks whether to pay a valid bill; extraction review asks whether extracted fields match the PDF and pass deterministic rules. Keeping them separate avoids slowing approvers with OCR fixes and preserves a clear field-level audit trail.
Microsoft’s Power Automate approvals show the boundary. The connector automates sign-off—including invoices (modern approvals guide)—where approvers respond from Outlook, a Teams adaptive card, or the Power Automate action center. Per Microsoft’s approval actions guide, Start and wait for an approval creates records in Dataverse, sends notifications, and blocks until completion; posting a Teams adaptive card uses Create an approval, a Teams post action, and Wait for an approval. That fits final sign-off, not field-level correction:
- Decision buttons, not multi-field edits. Per the Standard approvals connector, Basic and Await-all types accept case-sensitive
"Approve"and"Reject"responses plus comments. Custom responses can fail under"Everyone must approve"across many users due to results-field size limits, andStart and wait for an approval of texttargets one suggested text block, not a validated invoice grid. - Client formatting limits. Although
Detailssupports Markdown, Microsoft’s approvals Markdown guide notes inconsistent client rendering, states Markdown is not currently supported in the Teams Approvals app, and notes forced table-cell line breaks and table alignment are unsupported across Power Automate, Outlook Web, and Teams. - Delivery and access limits. The Standard approvals connector notes actionable Outlook emails are unsupported for guest users (who must use the Power Automate portal), multiple attachments cannot share a filename, emails are only actionable from the standard sender account, and timestamps always display in UTC.
Handle extraction exceptions in a dedicated AP review screen, then trigger ERP or Power Automate sign-off once blockers clear.
Anchor every extracted field to its source excerpt
Bind every editable input to its source-page crop so clerks never hunt through a multi-page PDF to confirm one figure.
In Amazon Textract response objects, detected items return as Block objects carrying Confidence and Geometry. Per Locating items on a document page, BoundingBox, and Geometry, Geometry returns an axis-aligned BoundingBox (Left, Top, Width, Height as 0–1 ratios from upper-left (0,0)), a Polygon of normalized Point (X, Y) coordinates for a finer-grain outline, and optional rotation for WORD blocks (the location guide and Geometry API reference describe different ranges).
For invoices, Analyzing invoices and receipts documents AnalyzeExpense (synchronous) and StartExpenseAnalysis / GetExpenseAnalysis (asynchronous), mapping supplier labels to standard types (INVOICE_RECEIPT_ID, PO_NUMBER, SUBTOTAL, TAX, TOTAL, AMOUNT_DUE, and line-item ITEM, QUANTITY, UNIT_PRICE, PRICE, with unmapped fields under OTHER). Per ExpenseField and ExpenseDetection, each field can include optional PageNumber, Type (with its own confidence), LabelDetection, and ValueDetection, with Geometry and Confidence nested inside each ExpenseDetection. Because LabelDetection is optional—for instance, Textract can infer VENDOR_NAME from a logo without an explicit label—the UI must handle a missing label and fall back to highlighting ValueDetection.
Use that geometry in three places:
- Click-to-focus crops: Multiply
BoundingBoxorPolygonratios by rendered pixel dimensions so focusing a field jumps toPageNumberand outlinesValueDetection(plusLabelDetectionwhen present). - Page badges on multipage files: Amazon Textract best practices notes PDF and TIFF files can span multiple pages; show
PageNumberbeside each field so later-page totals stay visible. - Table-band zoom on layout issues: Textract best practices advises upright input ideally at
150 DPIor higher without downsampling existing PDF, TIFF, JPEG, or PNG files, and keeping tables upright and visually separated from images or complex patterns. Separately, AWS notes inconsistent table extraction can occur when merged cells span columns or cells, rows, or columns differ across the table. On line failures, highlight the wholeEXPENSE_ROWor table band so clerks spot shifted columns immediately.
Show extracted value, confidence, and validation error together
An extracted value alone hides why the bill stopped; a confidence score alone misses high-confidence reads that fail accounting rules.
Amazon Textract best practices defines confidence as a 0–100 probability score. For false-positive-sensitive workflows, AWS recommends enforcing a minimum threshold and discarding or flagging lower-scoring results for higher human scrutiny—noting archival tasks like handwritten notes might use 50%, while financial decisions might require 90% or higher.
Expose four validation gates side by side:
| Gate | What the screen displays | Upstream control |
|---|---|---|
| Extraction confidence | Numeric score beside that field’s cutoff | OCR confidence thresholds for invoice review |
| Schema and math | Rule code plus expected vs. actual (null key, bad date, subtotal + tax != total) |
JSON Schema for invoice extraction fields |
| PO and receipt match | Invoiced quantity or unit price beside PO and warehouse receipt | Three-way invoice matching exception queue |
| Duplicate check | Matching vendor key, normalized invoice ID, amount, and link to existing ERP entry | Duplicate supplier invoice detection |
Pin blockers to the top, collapse clean fields behind a toggle, and print each failure in plain text so clerks see what must pass to unblock staging.
Make corrections editable, re-validated, and audited
Exception screens break when edits skip validation or overwrite raw parser output. Enforce four rules:
- Keep raw extraction immutable. Store parser output (detections, confidence, geometry, page number, parser version) separately from
corrected_valueso engineering can measure error rates and retune thresholds. - Re-run deterministic checks live. Every edit to an ID, quantity, unit price, or tax amount should immediately re-run schema checks, math, duplicate detection, and PO/receipt tolerances. Keep
Submit for Approvaldisabled while any blocker remains. - Separate OCR fixes from supplier disputes. High-confidence reads can still reflect overbills or short shipments. Provide
Confirm Extracted Value,Save Correction,Route to Buyer / Receiving, andReject as Duplicatewith mandatory reason codes so clerks never alter a real invoice figure to force a match. - Log edits in UTC. Store
reviewer_id, UTC timestamp,field_name,raw_value,corrected_value, andreason_code. Because Power Automate approval actions create records in Dataverse (approval actions guide) and show timestamps in UTC (Standard approvals connector), UTC logging aligns both timelines.
Illustrative exception screen before approval
(Illustrative mockup—not a real supplier, invoice, PO, or Insonl client. All names, IDs, coordinates, scores, amounts, and rule codes below are made up.)
+---------------------------+-------------------------------------------+
| SOURCE PDF (Page 1 of 2) | Doc #EX-8841 [3 Blockers] | PO-10492 |
| Subtotal: 1,200.00 | Vendor: Example Industrial Supply |
| Tax (10%): 120.00 |-------------------------------------------|
| +-----------------------+ | 1. INVOICE_RECEIPT_ID [74.2 < 95.0] |
| | Total: 1,320.00 | | Raw: INV-2026-441O -> Edit:[_________] |
| +-----------------------+ | 2. TOTAL [81.4 < 95.0 | SUM_MISMATCH] |
| | Raw: 1,300.00 -> Edit:[_________] |
| [Zoom] [Rotate] [Boxes] | 3. LINE 2 QTY [99.1 | THREE_WAY_QTY] |
| | Inv 12 vs Rcp 10 -> [Route Ops] |
+---------------------------+-------------------------------------------+
| Field & source excerpt | Extracted value & confidence | Blocking error | Editable correction / action |
|---|---|---|---|
INVOICE_RECEIPT_ID — Page 1 (Left 0.68, Top 0.14), label Inv No: |
INV-2026-441O (letter O), conf 74.2 |
CONF_BELOW_THRESHOLD (95.0 cutoff) |
Clerk checks crop, edits to INV-2026-4410, and confirms; confirmation clears CONF_BELOW_THRESHOLD and duplicate check passes. |
TOTAL — Page 1 (Left 0.66, Top 0.80), label Total: |
1,300.00, conf 81.4 |
CONF_BELOW_THRESHOLD + SUM_MISMATCH (1,200.00 + 120.00 != 1,300.00) |
Clerk checks smudged 2 in crop, edits to 1,320.00, and confirms; confirmation clears CONF_BELOW_THRESHOLD and math passes. |
Line 2 QUANTITY — Page 1 (Left 0.52, Top 0.56), table row 2 |
12, conf 99.1 |
THREE_WAY_QTY_MISMATCH (PO 12, receipt 10) |
PDF matches extraction; clerk clicks Confirm Extracted Value and Route to Receiving. |
Once blockers clear, freeze the validated snapshot, stage it in the ERP, and trigger sign-off. When routing sign-off to Power Automate approvals, pass a locked HTTPS link to the review record (per Microsoft’s approvals Markdown guide, HTTPS links render across clients) and give any attached PDFs unique filenames, because approvals with multiple attachments of the same name cannot be created (Standard approvals connector).
Audit your invoice review workflow before building a custom UI
Insonl is a Ho Chi Minh City engineering studio building ERP systems and operational AI pipelines. If your AP team fixes OCR errors directly in the ERP—or approves summary notifications without seeing the source document—start with our free AI workflow audit.
The audit is free; work after it is quoted separately.
Sources
- Amazon Web Services (third-party docs), Best practices — Amazon Textract: https://docs.aws.amazon.com/textract/latest/dg/textract-best-practices.html
- Amazon Web Services (third-party docs), Text detection and document analysis response objects — Amazon Textract: https://docs.aws.amazon.com/textract/latest/dg/how-it-works-document-layout.html
- Amazon Web Services (third-party docs), Locating items on a document page, Geometry, and BoundingBox — Amazon Textract: https://docs.aws.amazon.com/textract/latest/dg/text-location.html · https://docs.aws.amazon.com/textract/latest/APIReference/API_Geometry.html · https://docs.aws.amazon.com/textract/latest/dg/API_BoundingBox.html
- Amazon Web Services (third-party docs), Analyzing invoices and receipts, ExpenseField, and ExpenseDetection — Amazon Textract: https://docs.aws.amazon.com/textract/latest/dg/invoices-receipts.html · https://docs.aws.amazon.com/textract/latest/APIReference/API_ExpenseField.html · https://docs.aws.amazon.com/textract/latest/APIReference/API_ExpenseDetection.html
- Microsoft Learn (third-party docs), Get started with Power Automate approvals, Create and test an approval workflow, and Differences between flow approval actions: https://learn.microsoft.com/en-us/power-automate/get-started-approvals · https://learn.microsoft.com/en-us/power-automate/modern-approvals · https://learn.microsoft.com/troubleshoot/power-platform/power-automate/approvals/differences-between-flow-approval-actions
- Microsoft Learn (third-party docs), Standard approvals — Connectors and Use Markdown language to format Power Automate approvals: https://learn.microsoft.com/en-us/connectors/approvals/ · https://learn.microsoft.com/en-us/power-automate/approvals-markdown-support
- Insonl: ERP Systems · Free AI workflow audit · Set OCR Confidence Thresholds for Invoice Review · A JSON Schema for Invoice Fields Your ERP Can Validate · Three-Way Invoice Matching: What to Send to the Exception Queue · How to Catch Duplicate Supplier Invoices Before Posting
