Number7AI — Docs

Extraction failure modes taxonomy

The critical question is not whether an IDP ever fails — it always does on some document. The question is whether failures are silent (wrong output posted as correct) or visible (flagged, routed, corrected). This taxonomy maps the major classes.

Last updated: April 2026

TL;DR

AP failures split into four classes: structural, contextual/numeric, identity, and boundary.
The highest-risk class is silent wrong output — numbers in the right position but wrong semantic field.
Vanity accuracy metrics (e.g. "99% accuracy") mask silent error risk. ERP-safe pass rate is the real number.
Good systems route every failure to auto-correct, human review queue, or explicit rejection — never silent posting.

Why silent failures matter more than obvious ones

When a system clearly fails — blank output, parse error, obvious garbage — a reviewer catches it immediately. The dangerous failures are confident wrong answers: unit price and quantity transposed, tax line dropped, vendor identity mismatched to a ghost record. These pass through automated queues and surface only at period close, audit, or when a vendor chases payment.

Designing for failure visibility is more important than headline accuracy. A system that flags its own uncertainty is safer than one that is confidently wrong.

Structural failures

Multi-row line fragmentation
A single product line spans 2–3 rows; the parser treats each row as a separate item, producing phantom duplicates and wrong totals.
Multi-page table continuity
Table headers appear on page 1 only; continuation rows on page 2 lose column context and are mapped incorrectly or dropped.
Nested table flattening
Sub-tables (e.g. GST breakdown inside a line cell) are flattened into the wrong parent column, corrupting both the line item and the tax field.

Contextual / numeric failures

Indian lakh notation misparse
1,00,000 is parsed as 100,000 or 1,00,000.00 depending on locale assumption. Silent wrong amount reaches ERP.
Format artifact in price fields
Scanned rupee symbols produce 4200/- or values like 00:80 where OCR misreads the decimal. Passes as a valid number.
Tax rate without base amount
Invoice shows CGST 9% but not the taxable base. System either infers the base incorrectly or silently drops the tax line.
Mixed date formats
01/02/26 could be DD/MM/YY or MM/DD/YY depending on vendor origin. Wrong parse leads to wrong posting period.

Identity failures

Trade name vs legal entity
Vendor invoice says 'ABC Traders' but the accounting master has 'ABC Traders Pvt Ltd'. Unmatched record blocks posting or creates a ghost vendor.
GSTIN format variations
Spaces, dashes, lowercase — GSTIN on invoice does not match master due to formatting, not actual mismatch.
Remit-to vs bill-from divergence
Multi-entity vendors invoice from entity A but payment should go to entity B. No structural signal in the document.

Boundary failures (bulk PDFs)

Cover page treated as invoice
Email print-to-PDF includes a cover page before the invoice. System treats it as the first page of a document, shifting all fields.
Same-vendor invoices merged
Two invoices from the same vendor land in adjacent pages. Weak boundary detection merges them into one record with combined totals.
Multi-invoice PDFs out of order
40-page PDF with 12 invoices in arbitrary order. Boundary detection must identify start/end without relying on consistent headers.

Observed residual failure rates

Rates from production AP workflows on Indian documents. "Residual" means after IDP processing — these are the failures that reach the exception queue or (worst case) posting.

Failure class	Observed residual range	Primary trigger
Multi-row table continuity	~1.5–2%	Long descriptions, uneven scan quality
Nested table semantics	~3–4%	3+ level nested structures
Locale / format numeric	<0.5% → ~5%	Missing GSTIN or legacy format artifacts
Tax amount inference	~8%	Rate present but base amount unclear
Vendor identity mismatch	~2–5%	Trade name vs legal entity, GSTIN formatting
PDF boundary confusion	~1–3%	Cover pages, same-vendor adjacency

How AIdaptIQ routes exceptions

Every document outcome falls into one of three paths — nothing is silently posted if a failure is detected:

Auto-corrected
Known patterns (lakh notation, rupee symbols, GSTIN formatting) normalized and logged in the audit trail with correction reason.
Flagged for review
Low-confidence fields highlighted in the review UI with context. Human corrects and approves before any ERP push.
Rejected with context
Document returned with a specific error explanation — not a generic failure. Reviewer knows exactly what to fix on resubmit.