What document processing achieves on UK business documents, how accuracy should be measured, and where exception handling dominates cost.
Intelligent document processing replaces manual data entry from invoices, forms, contracts, and correspondence with automated extraction. The technology works well. Projects fail anyway, and they fail for reasons that are predictable before starting.
The recurring pattern: a pilot achieves 94% accuracy on a clean sample, the business plans for 94%, and production delivers 78% because production documents are not the sample. Everything downstream was sized on the wrong number.
Accuracy is not one number
"95% accurate" is meaningless without specifying the unit of measurement, and vendors vary the unit to suit.
Character-level accuracy counts correctly recognised characters. Highest number, least useful. 99% character accuracy on a 12-digit account number means roughly one digit wrong in every eight documents.
Field-level accuracy counts correctly extracted fields. More meaningful, and the standard you should insist on.
Document-level accuracy counts documents where every required field is correct. Lowest number and the one that determines actual labour saved — because a document with one wrong field still requires human review, and reviewing is most of the cost of doing it manually.
The relationship is unforgiving. At 97% field accuracy across 12 fields, document-level accuracy is approximately:
0.97^12 = 0.69Roughly 31% of documents need human attention. That is a very different operating model from what "97% accurate" suggests to a finance director.
Insist on document-level accuracy, measured on your documents, including the messy ones.
Where accuracy actually degrades
Vendor demonstrations use clean, well-scanned, structurally consistent documents. Production contains:
Photographs taken on phones at an angle, with shadows and partial glare.
Scans of photocopies of faxes, still genuinely present in UK construction, legal, and healthcare workflows.
Handwritten annotations over printed forms, frequently carrying the operative information.
Multi-page documents split incorrectly, or several documents in one PDF.
Layout variation across suppliers — 400 suppliers means 400 invoice layouts.
Multi-language content, and documents mixing languages.
Structured, consistent documents extract at 90–98% document-level accuracy. Semi-structured documents with layout variation reach 70–90%. Unstructured documents requiring interpretation land at 50–80%, and the residual demands judgement rather than correction.
Test on a stratified sample of real documents, deliberately including the difficult tail. A pilot on your cleanest 200 documents predicts nothing about production.
Exception handling is the project
Extraction is the visible part. Exception handling is where the cost and the value sit.
Design decisions that determine whether the system succeeds:
Confidence thresholds per field, not globally. A misread invoice total is a financial error; a misread supplier address is a nuisance. Route by consequence, not by a single number.
A review interface built for speed. Reviewers should see the document with the extracted field highlighted in place, correct inline, and move on. If review means opening the PDF separately and retyping into another system, you have not automated anything — you have added a step.
Corrections feeding back into improvement. Without a loop from correction to model or template refinement, accuracy is static forever.
Explicit ownership of the exception queue. Unowned queues grow until someone declares bankruptcy on them.
A realistic saving is not "eliminate data entry." It is "reduce data entry by 60–75% and change the remainder from typing into verifying," which is faster and less error-prone but not free. Business cases built on total elimination fail on contact with production.
UK-specific considerations
GDPR and lawful basis. Documents contain personal data. Processing requires a lawful basis, and automated processing producing legal or similarly significant effects engages additional requirements around human involvement. A DPIA is generally appropriate before deployment.
Data residency. Many providers process in specific regions. Where documents contain personal or commercially sensitive data, verify processing location contractually rather than assuming, and confirm whether documents are retained for provider model improvement — often a default that requires opting out.
Retention. HMRC requires business records for six years; other obligations vary. Your IDP system becomes part of the retention story, including the original images, not merely the extracted data.
Making Tax Digital. Invoice and receipt processing intersects with digital record-keeping requirements. Extracted data feeding VAT returns needs an auditable link back to the source document.
Handwriting and legacy formats. UK sectors with long document histories — legal, construction, NHS supply chains — carry a higher proportion of handwritten and poor-quality material than vendor benchmarks assume.
Costing an engagement
Discovery and document analysis £8,000-25,000
Pilot (single document type) £15,000-45,000
Production implementation £40,000-150,000
Per-page processing (cloud) £0.01-0.10
Ongoing tuning and support £2,000-8,000/moCost drivers: number of distinct document types, layout variability within each, integration surface with downstream systems, and required accuracy threshold. Moving from 85% to 95% document-level accuracy frequently costs more than reaching 85% did.
The per-page figure is usually the smallest line and receives the most attention in procurement. Integration and exception-handling design dominate total cost.
Scoping a pilot that predicts production
Pick one document type with real volume. Not the hardest, not the easiest.
Sample 300–500 real documents, stratified to include the poor-quality tail in its true proportion.
Define fields and their consequence tiers before measuring anything.
Measure document-level accuracy on that sample, with per-field breakdown.
Time the exception workflow with actual reviewers, not with the project team.
Compute the business case from measured numbers — documents fully automated, average review time for the rest, versus current cost.
If the pilot cannot demonstrate this on real documents, production will not either.
Frequently asked questions
How accurate is IDP really? Structured consistent documents reach 90–98% document-level accuracy. Semi-structured documents with layout variation reach 70–90%. Always ask for document-level accuracy on your own documents — character-level figures overstate performance substantially.
Why did our pilot outperform production? Almost always because the pilot sample was cleaner than production. Stratify your sample to include poor-quality documents in their real proportion.
What does it cost in the UK? Pilots run £15,000–45,000, production implementations £40,000–150,000, and per-page processing £0.01–0.10. Integration and exception-handling design dominate; per-page cost is usually the smallest component.
What are the GDPR implications? Documents containing personal data require a lawful basis, and a DPIA is generally appropriate. Verify processing location contractually and confirm whether the provider retains documents for model improvement — opting out is frequently required.
Can it read handwriting? Modern systems handle clear handwriting reasonably and struggle with cursive, annotations in margins, and poor-quality scans. If handwritten content carries operative information, test it specifically rather than accepting a general accuracy figure.
Further reading
Related
Written by
Moueen Togarvi
Founder & CEO at Voquarn Code, focused on product engineering, search growth, and practical AI systems.