Pricing
Book a Demo Register for a Trial
AI Invoice Review · Live

See exactly what the AI saw.

Bounding boxes show where every field was extracted from. Confidence colours show how sure the AI was. Drag-to-redraw any region - the AI re-extracts on the spot. Trust comes from showing your work.

Book a Demo Watch Reels Calculate Your Savings
99%+
Field-level accuracy
3
Confidence tiers
9
Reviewer feature flags
AI Invoice Review LEDGE
INV-4821 · £2,847.40 96%
Math check Pass
VAT rates valid 20% std
No duplicate found 90d clean
Salmon price 23% above avg Anomaly
3-way match PO + GRN
Why this matters

Most invoice tools hide their work.

Black-box extraction tells you what was found, not where, not how confident and not what to double-check. That asks reviewers to trust on faith - and quietly lets errors through.

The black-box approach

"Here's what we extracted."

  • Fields shown as a flat list - no link back to where they came from on the document.
  • No confidence indicator - every field looks equally trustworthy.
  • To fix an error, you re-type - the AI never learns from the correction.
  • Anomalies surface as a separate alert, disconnected from the document context.
● The OmniPATH approach

"Here's what we extracted and where and how sure we were."

  • Bounding boxes on the document show exactly which pixels became which field.
  • Three confidence tiers colour-coded - green, orange, red - so you know exactly what to scan.
  • Drag to redraw any region - the AI re-extracts and the field updates instantly.
  • Anomaly hover labels explain why a value was flagged, right next to the value itself.
01 / Bounding box overlay

Every field traces back to the pixels.

When you click a field on the left, the corresponding region on the document highlights. When you hover a region on the document, the field on the left highlights back. Reviewers stop scanning blindly - they know exactly which scribble in the corner became "PO Ref: PO-08823".

Boxes are colour-coded by extraction confidence so the eye is drawn to exactly the fields that need a second look.

"It cut review time in half - not because we're faster at reading, but because we stop reading the green ones."
95%+
High Confidence
75-95%
Worth a check
< 75%
Needs review
Confidence calibrated per field type Live data
02 / Drag-to-select redraw

Wrong region? Just redraw it.

If the AI grabbed the wrong PO number - or missed a field entirely - drag a box around the right region. Words snap to character boundaries automatically. The AI re-extracts text from the new region and the field updates without a page reload.

Word-aware redraw means you don't have to be precise. Click roughly around a value and it locks to the actual word boundaries.

PURCHASE ORDER REFERENCE
PO-08823 · supplier ref BF-2026-Q2-0481
Re-extracting…
03 / Anomaly hover labels

Warning triangles that actually explain themselves.

When the credit tracker, 3-way matcher or anomaly engine flags a value, the warning sits inline next to that value - with a hover label that explains why. No bouncing to a separate alerts panel to figure out what the system was complaining about.

The same hover pattern surfaces match warnings from PO/GRN reconciliation, duplicate invoice flags and supplier verification status.

Subtotal
£4,039.33
VAT (20%)
£807.87
VAT calc mismatch · expected £807.87, line items sum to £812.40
Total Due
£4,847.20
Hover the warning triangle for context
Built for reviewers

Nine features shipped to make manual review faster.

Every detail in the review interface earned its place by reducing time-on-task or improving extraction accuracy. The full set:

Multi-region bboxesSingle field, multiple page regions
Word-aware redrawSelections snap to word boundaries
Optimistic saveUI updates before round-trip completes
Keyboard navigationTab through fields, ↵ to confirm
Auto-fit documentPage scales to container width
Zoom in/outPlus page navigation controls
Review state trackingReviewed · Edited · Alert per field
Anomaly hover labelsNative tooltips explain flagged values
Credit tracker integrationmatch_warnings shown on flagged amounts
Frequently asked

The technical details.

How is "AI Invoice Review" different from your IDP/extraction page?
IDP (Intelligent Document Processing) is the extraction engine that runs on every document automatically. AI Invoice Review is the human-in-the-loop interface that surfaces what the engine extracted, with the bounding-box and confidence overlay, for the cases where someone needs to verify or correct the extraction. Most invoices never need it - but when one does, the review experience is where reviewer time is saved or wasted.
Are confidence scores calibrated, or just decorative?
Calibrated, per field type. A 95% confidence on the supplier name is not the same model output as 95% on a line-item quantity - different fields use different validation logic (vendor lookup, regex, sum-checks against headers) so the threshold for "needs review" differs. Below 75% always shows red and routes for manual confirmation; above 95% routes straight through unless an anomaly is flagged.
What happens when I redraw a region - does the AI actually re-process?
Yes. The new region's pixels are sent to the extraction model for that field type, the result is returned with a new confidence score and the field updates. Word-aware snapping happens client-side first so the visual feedback is instant; the actual re-extraction completes in the background and the field repaints when it lands. Average round-trip is under a second.
Do edits made in review feed back into the model?
Yes - every edit is captured as supervised training data, scoped to your organisation. The model learns supplier-specific quirks (FS Foodservice's quirky PO format, that one supplier who puts the date in the footer) without sharing your data with other tenants. Improvements compound over time and are most visible in the first 30 days of use.
What anomaly types trigger the warning triangles?
VAT calc mismatch, line-item sum vs header total, duplicate invoice number, supplier on credit hold, PO/GRN quantity mismatch, unit-price drift vs trailing 90-day average and several more. Each anomaly type has its own hover label explaining the rule that fired and the value the rule expected. New anomaly checks ship roughly monthly.
Can we configure which fields show in the review panel?
The default review fieldset covers the standard invoice spec - supplier, invoice number, dates, PO ref, financials, line items. Custom fields configured at the org level (cost centre, project code, custom approval routing tag) automatically appear in the panel and get their own bounding-box treatment if the AI is set up to extract them.
Ready when you are

Stop guessing what the AI did.

Bring a real invoice to the demo - we'll process it live and walk through the review interface field by field.

Book a Demo See IDP
30-min demo · Bring your gnarliest invoice