Browse the knowledge base

Understanding confidence scores

Document scanning produces ONE overall confidence score per document (0–100). At 70 or above, the scan auto-creates the invoice. Below 70, the document is automatically re-read with a stronger model; if confidence stays low, the document is still stored and searchable. You just create the invoice yourself. Bank-transaction matching has its own separate 0–100 score (90+ links silently). The two scales never mix.

Two different scores you shouldn't mix up

TaxItEasy uses two independent 0–100 scores, and the number 70 means something different in each:

  • Document confidence (scanning). One overall score per scanned document, rating how sure the AI is about what it read. At 70 or above, an invoice is auto-created from the scan. This score is about reading the document.
  • Match score (bank reconciliation). A separate score for how well a bank transaction fits an invoice. At 90 or above the match links silently; at 70–89 it links with a review flag, never silently. This score is about linking a payment. See how the matching pipeline works.

So "70" in scanning means "confident enough to create the invoice", while "70" in matching means "linked, but a human should glance at it".

The document confidence score

Every scan produces one overall score for the whole document. There are no per-field scores. What happens at each level:

  • 70 or above: the invoice record is created automatically from the extracted data: vendor, amounts, VAT, dates, line items.
  • Below 70: the document is automatically re-processed with a stronger model before anything is decided. Most borderline scans clear the bar on the second pass.
  • Still below 70: the document is stored, readable, and searchable like any other; it just doesn't auto-create an invoice. On mobile you'll see an honest "OCR confidence too low" hint with a Create anyway button; on web you can create or extract the invoice manually.

A low score is not a failure. It's the system declining to book numbers it isn't sure about.

Consistency checks instead of per-field scores

Rather than scoring each field separately, the pipeline runs hard consistency checks on the numbers that matter:

  • Amount reconciliation. The total paid (gross) is the anchor. If the extracted net + VAT doesn't add up to the total, the net is re-derived from the total and the invoice is flagged for human review, because contradictory numbers are never booked silently.
  • Country-aware VAT. The default VAT rate comes from your company's country; the extraction prompt knows the valid rates for the detected country of the document.

If an invoice shows up flagged for review, that's usually the amount check having caught something. Open it and confirm or correct the totals.

Improving low scores

  • Retake the photo. The pipeline already straightens, crops, and normalizes contrast on phone photos, but a photo you can't read yourself won't score well. Flat angle, good light, no glare.
  • Rescan. The document detail page has a re-scan action, up to 3 re-runs per document.
  • Correct once, benefit later. When you correct a scanned invoice, the system learns vendor-specific patterns (category, currency, VAT rate, net-vs-gross layout). After 3 confirmed corrections, the hint becomes active and is applied to future scans of that vendor. If a vendor's values genuinely vary, the system learns not to assume instead of guessing.

The match score, briefly

Matching scores four signals: invoice number (40), amount (25), date (20), counterparty (15). The total then maps to tiers, from strong (green, 90+, silent) through likely (amber, 70–89, review flag) and possible (blue, 50–69, suggestion) down to weak (gray, 30–49, hidden). Every suggestion and auto-match carries a "Why this match?" breakdown showing exactly which signals scored, and every match is undoable. See why this matched, and how to undo.

Related

Didn't answer your question? Write to [email protected] · the AI chat in the bottom-right corner answers most common questions.