How accurate is AI document processing?

01Field level or document level — what an accuracy figure actually refers to

When you see "99 % accuracy" quoted somewhere, it almost always means the field level: how often is a single field — invoice number, date, amount — read correctly? What matters for day-to-day work, though, is the document level: how many documents are completely error-free and pass through without rework? The difference is pure arithmetic: at 98 % field accuracy and 20 fields per document, only about 67 % of documents are fully correct (0.98²⁰ ≈ 0.67). Even at 99 % per field, roughly one in five documents still contains at least one error. Anyone who conflates the two levels overestimates every system.

02What current studies and benchmarks show

Reliable numbers come from published studies and research benchmarks — three examples:

• Invoice extraction with language models: An October 2025 evaluation study (arXiv) measures 94 % overall accuracy on invoices for an LLM-based extractor — a classic OCR layout pipeline reached 63 % in the same test.
• Difficult fields: In the same study, the LLM approach reached 91 % on line items versus 58 % for the classic pipeline; tax and net amounts were among the most error-prone fields for the classic pipeline, while for the LLM approach it was mainly varying number formats that caused errors.
• The DocILE research benchmark: The ICDAR benchmark, with around 6,700 real business documents and 55 field classes, shows how demanding real-world document variety is: the baseline models reported in 2023 reached F1 scores of 0.66 to 0.72. Newer models score higher — but even research is far from "every field always right."

03Print, scan, handwriting — how big the differences are

• Clean print: A digitally generated PDF or a clean printout is the ideal case — here current systems read almost error-free, and this is where the high field accuracies come from.
• Scans and photos: Skewed scans, low resolution, stamps and creases noticeably reduce recognition. The quality of the source feeds directly into the rate.
• Handwriting: On the academic IAM benchmark for English handwriting, the best models now reach character error rates of around 1.2 to 1.5 % (as of April 2026); classic OCR software sits above 12 % there. On real-world forms with hurried handwriting, however, you should plan for considerably more manual checking than with print.

04What accuracy depends on in your project

• Field type: Invoice number and date are easy; totals, tax rates and line-item tables are the typical sources of error. Arithmetic validation helps here — net plus tax must equal gross.
• Layout variety: 20 suppliers with fixed forms are easier than 2,000 suppliers with changing layouts. The broader the intake, the more important the review loop.
• Validation against master data: Cross-checking against supplier, purchase-order and account data catches many extraction errors before they reach the ERP.
• Measuring on real material: A system tuned and measured on your own documents tells you more than any blanket figure from a benchmark.

05Why 100 % is not a realistic promise — and what counts instead

Language models and OCR work statistically: every extraction carries residual uncertainty, and the source itself may be illegible or ambiguous. The right response to this is not a higher percentage, but an architecture that expects errors:

• Confidence threshold: The system outputs a confidence value for each field. Values above the threshold pass through automatically; anything below goes to manual review.
• Four-eyes fallback: Uncertain documents land in a short review screen. An employee confirms or corrects the flagged fields — logged and auditable.
• The right metric: What matters is not accuracy alone, but the automation rate at an accepted error rate: how many documents pass through untouched without wrong values reaching the system?
• Measuring before going live: The rate is measured on a test set of real documents; the threshold is set so that almost no errors remain in the automatically processed portion. That is how we handle it in the fixed-price pilot too.