How ExtractBee's Human Review UI Works: Catch Errors Before They Reach Your Books

No AI extraction system is perfect. A poorly scanned document, an unusual vendor layout, a handwritten annotation — any of these can reduce extraction confidence. The question isn't whether errors will occur. It's whether you catch them before or after they enter your accounting system.

Most document extraction tools answer this question the same way: they push the data downstream and leave error detection to you. You find the mistake during bank reconciliation, or worse, when a supplier disputes a payment.

ExtractBee takes a different approach. When extraction confidence is low, the document is held in a Human Review queue — not sent anywhere — until a person reviews it. This is the Human Review UI.

What Triggers Human Review

Every field ExtractBee extracts has a confidence score — a measure of how certain the AI model is about that specific value. (For a deeper look at how scoring works, see how ExtractBee calculates confidence scores.)

When a field's confidence falls below your threshold (configurable in workspace settings), that extraction is flagged for review. The document is not exported, not sent to Xero, not pushed to your Google Sheets destination — it waits.

Common triggers:

Poor quality scan (low DPI, skewed, overexposed)
Handwritten values on printed templates (e.g. handwritten PO number)
Ambiguous date formats (01/02/2024 — is that January 2nd or February 1st?)
Obscured text (stamped "PAID" overlapping the total)
Unusual currency formats or non-standard number separators
Multilingual documents where context switches language mid-page

Critically: a document can be fully extracted with high confidence on 14 of 15 fields, and still enter the Human Review queue because one field — say, the due date — is ambiguous. You review only that one field, not the whole document.

The Review Interface

The Human Review UI shows two panes side by side:

Left pane: original document The full PDF or image, rendered at full quality. Flagged fields are highlighted with a coloured overlay on the document itself — you can see exactly where on the page ExtractBee found each value.

Right pane: extracted fields All extracted fields listed, with their values. Fields that triggered Human Review are visually distinguished — highlighted, with the confidence score shown. Fields with high confidence are greyed out — they're shown for context but don't require attention.

Workflow:

You see the flagged field and its extracted value
You look at the highlighted region on the document
You confirm the value is correct (one click) or type the correct value
Repeat for any other flagged fields
Click Approve — the extraction completes and data flows to your configured destinations

Average review time for a flagged invoice: 30–90 seconds. You're not re-entering the document from scratch — you're correcting one or two uncertain fields.

Confidence Threshold Settings

You control the threshold. In workspace Settings → Extraction, you set the minimum confidence score required for automatic approval.

Higher threshold (e.g. 95%) More documents enter Human Review. More human time spent reviewing. Fewer errors reach your accounting system.

Lower threshold (e.g. 75%) Fewer documents enter Human Review. Less human time. Slightly higher risk of errors passing through automatically.

The right threshold depends on your documents. A team processing clean digital PDFs from established vendors can set a lower threshold — extractions are consistently high-confidence and Human Review is rarely triggered. A team processing scanned paper invoices from small suppliers should set a higher threshold.

Most teams find a threshold of 85–90% works well in practice: Human Review is triggered for genuinely uncertain cases, not for routine extractions.

Why This Approach Matters

The alternative — pushing all extractions downstream and relying on downstream error detection — has a real cost.

Cost of a data entry error in accounts payable:

Wrong invoice number: supplier disputes payment → manual investigation → 30–60 minutes
Wrong amount: overpayment or underpayment → reconciliation issue → potential late fee
Wrong date: missed due date → late payment → damaged supplier relationship
Wrong vendor: payment applied to wrong account → audit finding

Human Review catches these at the point of extraction — when the document is in front of you and the correct value is one glance away — rather than weeks later during reconciliation when the audit trail is cold.

Human Review in the Workflow

Human Review integrates naturally into the broader automated extraction workflow:

Document received (email, upload, Drive)
        ↓
Extraction runs automatically
        ↓
High confidence → auto-approved → sent to destinations
Low confidence → Human Review queue → notification sent
        ↓
Reviewer opens queue → corrects flagged fields → approves
        ↓
Data sent to destinations

Notifications for Human Review items can be delivered by email or Slack — your team is alerted when something needs attention without having to check the dashboard manually.

Who Reviews?

Human Review access is role-based. In ExtractBee's permission system:

Owner and Admin can review all extractions
Member can review extractions they submitted
Viewer can see extractions but cannot approve

For accounting firms, this typically means the reviewing accountant has Member or Admin access, while clients who submit documents have Viewer access to track status.

Human Review vs Manual Processing

Human Review is not manual processing — it's supervised automation.

Manual processing: Every document is reviewed by a human. The human reads the PDF, types the values, and moves on. 100% human time per document.

Full automation (no review): Every document is processed automatically. No human time. But errors pass through silently.

Human Review (ExtractBee's approach): Most documents are processed automatically. Only documents with uncertain fields are reviewed. Human time is concentrated where it's actually needed — on the cases where uncertainty exists.

The practical difference: if 90% of your invoices extract with high confidence and 10% trigger Human Review, you're spending human time on 10% of your document volume rather than 100%. And the 90% that skip review are processed instantly, around the clock.

Integrating Human Review into Your Team Workflow

Human Review works best when it's part of a defined team workflow rather than an ad-hoc step.

Daily review sessions: Some teams check the Human Review queue once per day — typically in the morning before starting other work. This works well when extraction volume is moderate and timing of data entry isn't critical.

Real-time notifications: For teams where invoice data needs to be available quickly (e.g. for same-day approval workflows), Slack notifications for each flagged document work better than batch review sessions.

Role assignment: In larger teams, assign Human Review responsibility to a specific role. In ExtractBee, Admin and Member roles can approve reviews — Viewer roles can see but not approve. This prevents documents from sitting in the queue unreviewed because everyone assumes someone else will handle it.

Review SLA: Consider setting an informal SLA — "all flagged documents reviewed within 4 hours" for example. This ensures Human Review doesn't become a bottleneck that slows down your AP workflow.

Frequently Asked Questions

What happens if no one reviews a flagged extraction? It stays in the queue indefinitely. No data is sent downstream until a human approves it. You can configure reminder notifications — daily digests of pending reviews sent to your email or Slack.

Can I turn off Human Review entirely? Yes. Set your confidence threshold to 0% and all extractions are auto-approved regardless of confidence. This is not recommended for most workflows, but it's available for teams who have validated that their document mix consistently produces high-confidence extractions.

Is the confidence score per document or per field? Per field. A document with 15 fields might have 14 at 98% confidence and one at 62% — only the one uncertain field triggers Human Review. The other 14 are shown in the review interface but are pre-approved.

Can I see historical Human Review decisions? Yes. The audit log (Business plan and above) records every Human Review action: who reviewed it, when, what value was originally extracted, and what value was confirmed or corrected. This is useful for identifying systematic extraction issues with specific vendors.

Does reviewing a flagged field improve future extractions? Not automatically in the current version. AI Training (Business plan and above) allows you to submit corrected extractions as training examples, which improves extraction accuracy on similar documents going forward.

Try ExtractBee free — 20 pages/month →

ExtractBee is operated by MB Dokigo, Vilnius, Lithuania.