AI Confidence Scoring

Every document RecordEngine processes receives an AI confidence score from 0 to 100. This score tells you how certain the AI is that it correctly extracted all the important fields — so you can focus your review time where it actually matters instead of manually checking every document.

How the Score is Calculated

The confidence score is generated in the same AI pass as field extraction — there are no extra processing steps or API calls. The AI evaluates:

Whether all expected fields from the extraction profile were found
Whether the values found are plausible (e.g. a date that looks like a date, an amount that looks like an amount)
Whether line item subtotals add up to the stated total
Whether the document quality is sufficient (blur, rotation, partial scan, low contrast)
Whether the document language and format matched the selected extraction profile

Confidence scoring is automatic on every document. There is nothing to configure — just upload and the score appears.

Score Bands

Score	Label	Badge colour	What it typically means
90–100	High	🟢 Green	Clear document, all expected fields found and plausible
70–89	Good	🔵 Blue	Minor uncertainty on one or two fields — quick check recommended
50–69	Low	🟡 Yellow	Several fields uncertain or missing — careful review required
0–49	Poor	🔴 Red	Significant extraction problems — manual data entry likely needed

Where the Score Appears

In the document list

The badge appears next to every document filename so you can spot low-confidence documents at a glance without opening them.

In the document detail panel

Open any document and the full score, label, and AI reasoning text appear below the AI Summary in the centre panel. Confidence score reasoning in the document detail panel

Reading the AI Reasoning

Below the score badge you’ll see a plain-language explanation from the AI. For example:

“Clear invoice with all 8 required fields present. Line items sum correctly to the stated total of ¥12,450. High confidence.”

“Vendor name and invoice number extracted successfully, but the invoice date field appears partially cut off at the bottom of the page. Stated total does not match sum of line items — difference of ¥120. Moderate confidence.”

“Document appears to be a blurry photograph of a handwritten receipt. Several fields could not be read with certainty. Manual verification strongly recommended.”

This reasoning tells you exactly which field to check rather than making you re-read the whole document.

Using Confidence Scores in Your Workflow

Prioritise your review queue

Sort the document list by confidence score (lowest first) to tackle your most uncertain documents at the start of the day. High-confidence documents can often be approved quickly without line-by-line verification.

Filter in Advanced Search

In the Advanced Search page, use the Confidence Level filter to show only Low or Poor documents. Combine with a date range or contact filter to focus on a specific batch.

Advanced Search with confidence level filter applied

Connect to the Rules Engine

Confidence score is one of the most powerful condition fields in the Rules Engine. Some useful rules:

IF confidence_score < 60
→ Set Status = Exception
→ Send Notification to reviewer

See Rules Engine for how to create these automations.

Confidence Score in Outbound Webhooks

The confidence score and label are included automatically in every outbound webhook payload — so your downstream systems can make decisions based on extraction quality:

{
  "id": 847,
  "filename": "acme-invoice-nov-2025.pdf",
  "status": "Export",
  "confidence_score": 87,
  "confidence_label": "Good",
  "confidence_reasoning": "All 8 fields extracted. Line items sum correctly. Minor uncertainty on tax rate field.",
  "extracted_fields": {
    "vendor": "Acme Corporation",
    "invoice_number": "INV-2025-1147",
    "invoice_date": "2025-11-15",
    "total_amount": "12450.00",
    "currency": "CNY"
  }
}

Tips for Improving Confidence Scores

If you’re consistently seeing low scores on a particular document type, the most common fix is to create a custom extraction profile with field labels that match exactly what’s in your documents. The AI performs significantly better when the profile is tailored to the document format.

Issue	Fix
Low scores on scanned documents	Scan at 300 DPI minimum — higher resolution dramatically improves vision extraction
Chinese documents scoring low	Make sure you’re using the Chinese Fapiao or Chinese Business License profile, not the Default profile
Line item totals mismatch	Check if the document includes taxes, discounts, or fees that aren’t in a standard line item row
Audio transcriptions scoring low	Use a quiet recording environment — background noise reduces transcription quality

Getting Started

User Guide

Deployment

Troubleshooting

FAQ

How the Score is Calculated

Score Bands