Updated
What document processing achieves on UK business documents, how accuracy should be measured, and where exception handling dominates cost.
Intelligent document processing uses automated extraction to reduce manual data entry from invoices, forms, contracts, and correspondence. Its usefulness depends on the documents, validation rules and review workflow. Evaluate all three before treating an extraction demonstration as a production business case.
Consider a hypothetical failure: a pilot achieves 94% accuracy on a clean sample, but a different production document mix yields 78%. These numbers illustrate sampling risk; they are not a measured client result or an industry benchmark. Staffing a review queue from the cleaner sample would underestimate its workload.
Accuracy is not one number
"95% accurate" is incomplete without specifying the unit of measurement, evaluation sample and error definition.
Character-level accuracy counts correctly recognised characters. It does not directly measure whether a critical identifier is correct. In an illustrative independence model, 99% accuracy for each of twelve digits implies a probability of at least one error of 1 − 0.99^12, or about 11.4%. Real OCR errors may be correlated, so measure complete identifiers directly.
Field-level accuracy counts correctly extracted fields. Report it separately for consequential fields such as account numbers and totals, so many correct low-risk fields do not hide a serious error pattern.
Document-level accuracy counts documents where every required field is correct. It helps estimate which documents can proceed without corrections, but does not establish labour savings on its own. Some workflows require additional approvals even when extraction is correct; others can isolate review to a particular field.
The relationship is unforgiving. For an illustrative calculation only, if 12 fields each have 97% accuracy and errors are independent, the all-fields-correct probability is:
0.97^12 = 0.69Under those assumptions, roughly 31% of documents contain at least one field error. If every such error requires review, that is a very different operating model from what an unqualified "97% accurate" suggests.
Insist on document-level accuracy, measured on your documents, including the messy ones.
Where accuracy actually degrades
A demonstration may use clean, well-scanned, structurally consistent documents. Check whether your production mix also contains:
Photographs taken on phones at an angle, with shadows and partial glare.
Scans of photocopies or faxes, where repeated copying may obscure characters.
Handwritten annotations over printed forms, which may carry information the workflow needs.
Multi-page documents split incorrectly, or several documents in one PDF.
Layout variation across suppliers, including multiple versions from one supplier and shared formats across others.
Multi-language content, and documents mixing languages.
Accuracy varies with document mix, label quality, extraction rules, and the chosen model. This guide has no representative benchmark dataset, so establish performance on a held-out sample of your own documents rather than using generic percentages.
Test on a stratified sample of real documents, deliberately including the difficult tail. A pilot containing only the cleanest documents cannot establish performance across the rest of the production mix.
Exception handling is the project
Extraction is the visible part. Exception handling can determine whether the resulting workflow saves time.
Design decisions that determine whether the system succeeds:
Confidence thresholds per field, not globally. A misread total, supplier identity or address can have different consequences depending on the workflow. Define review rules from those consequences and measured errors rather than treating one global confidence score as approval.
A review interface built for speed. Reviewers should see the document with the extracted field highlighted in place, correct inline, and move on. Opening a separate PDF and retyping into another system can erode the benefit; measure the complete task rather than extraction latency alone.
Corrections feeding back into improvement. Record reviewed corrections and test proposed extraction or validation changes on held-out examples. A correction should not become a training label automatically when its accuracy is uncertain.
Explicit ownership of the exception queue. Assign a team, response target and escalation route, and monitor unresolved work by age and consequence.
Measure whether reduced typing outweighs the remaining verification and exception work. Review may be faster for some documents and slower for others; time it with the people who will operate the process instead of assuming total elimination of manual effort.
UK-specific considerations
GDPR and lawful basis. Documents contain personal data. Identify the applicable lawful basis and assess whether a DPIA is required for the planned processing. The ICO explains changes introduced by the Data (Use and Access) Act; some automated-decision guidance is under review. Have the current safeguards and requirements checked for the actual decision rather than assuming an older rule applies unchanged.
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 the provider's applicable retention and model-improvement terms rather than assuming a default.
Retention. Retention depends on entity type, tax regime, and the record concerned. The self-employed record guidance is not a universal six-year rule for every business. Your IDP system becomes part of the retention story, including the original images, not merely the extracted data.
Making Tax Digital. Review the applicable digital-record and linking requirements with the business's tax adviser. HMRC VAT Notice 700/21 explains VAT record keeping; preserve source references and review how the extraction workflow connects to the accounting records.
Handwriting and legacy formats. Inspect the organization's actual archive for annotations, photocopies and mixed document formats. This guide has no comparative sector dataset establishing how common those problems are.
Costing an engagement
Price discovery, labelled evaluation documents, extraction, review tooling, integrations, and support separately. Use measured page volume and reviewer time to model recurring cost. A per-page API price is not an implementation quote.
Cost drivers include document types, layout variability, downstream integrations and the required error limits. Measure the effort needed for each additional difficult case rather than assuming accuracy improves at a constant cost.
Separate per-page processing from integration and exception handling. Their relative cost depends on volume, review time and the existing systems; none should disappear from the business case.
Scoping a pilot that predicts production
Pick one document type with real volume. Not the hardest, not the easiest.
Choose a representative sample size, including important document types and rare consequential errors. Size it for the precision required; one universal document count is not sufficient for every workflow.
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 an acceptable result on representative documents, revise the design or scope before assuming production will improve it.
Keep extraction confidence separate from financial approval
A document model can propose fields; the accounting workflow should decide whether they are acceptable. Match supplier identity against an approved record, validate totals and currency, and route a changed bank account for independent verification. A high extraction confidence score is not payment authorisation.
Keep the original source reference and correction history so reviewers can reconstruct where a value came from. For international delivery, agree which samples may leave the client environment and who can label them. Privacy and retention obligations need case-specific review against the ICO’s AI guidance.
Test duplicated and missing pages.
Include currency and decimal-separator variation.
Check handwriting independently of printed text.
Measure review time as well as extraction quality.
Frequently asked questions
How accurate is IDP really? Measure required-field and document-level performance on representative documents, with a separate held-out test set. Generic accuracy percentages do not predict your exception workload.
Why did our pilot outperform production? Investigate differences in document mix, labels, model versions, configuration and downstream validation. A cleaner pilot sample is one possible cause, not a diagnosis established without evidence.
What does it cost in the UK? Ask for a scoped quote covering documents, review tooling, downstream integrations, and support. Calculate recurring processing and review costs from measured pilot data; this article does not provide verified UK market rates.
What are the GDPR implications? Establish the appropriate basis, processing agreement, risk assessment and any required DPIA for the particular workflow. Check current ICO guidance after the 2025 Act reforms, plus processing location, retention, and model-improvement terms.
Can it read handwriting? Handwriting performance depends on the model, script and document quality. If handwritten content carries operative information, evaluate it separately, including cursive, margin annotations and poor-quality scans, 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.