A UK CFO we spoke to last month had an accounts payable team of 4 people processing 400 supplier invoices per week from PDFs and scanned documents. Fully-loaded team cost was £180k annually, error rate on manual keying was 3-5 percent, triggering supplier reconciliation calls and month-end delays. His CIO wanted to evaluate IDP but was uncertain whether the accuracy claims from vendors would hold up in production against their specific supplier mix.
That is the AI document processing conversation across UK and US operations desks in 2026. IDP has matured materially. LLM integration in 2025-2026 pushed accuracy on unstructured documents from 50-70 percent to 70-90 percent. Cloud IDP options (Azure DI, AWS Textract, Google Document AI) made entry cost minimal. But vendor accuracy claims are often measured on ideal documents, and human-in-the-loop workflow is essential rather than optional.
This article is a candid guide for COOs, CIOs, and finance leads scoping IDP. IDP vs OCR. Accuracy rates. Vendors. Costs. Human-in-the-loop reality.
IDP vs OCR: What Actually Changed in 2026
Two technologies often conflated but distinct.
OCR (Optical Character Recognition). Converts image or scan to text. Recognises characters but not structure or meaning. Output is a text blob; downstream systems must parse it. Mature at 95-99 percent character accuracy on clean documents. Fails on handwriting, unusual layouts, and low-quality scans.
IDP (Intelligent Document Processing). Combines OCR with AI (traditionally ML, increasingly LLMs in 2025-2026) to classify documents, extract structured fields, understand context, and handle variable formats. Output is structured data ready for downstream systems.
The 2026 shift. LLM integration transformed handling of unstructured documents. Traditional template-based extraction required documents matching known layouts. LLM-powered IDP handles freeform documents (contracts, emails, unusual invoices) that previously required manual keying. Unstructured accuracy moved from the 50-70 percent range to 70-90 percent.
When you need which. Pure OCR for text extraction where downstream parsing is trivial. IDP for anything requiring structured output. In 2026, most enterprise document workflow needs IDP.
What AI Document Processing Actually Does
Eight capabilities that separate real IDP from OCR-plus-scripting.
Document classification. Categorise incoming document (invoice, receipt, contract, purchase order, form) and route to appropriate pipeline.
Field extraction. Extract structured data (invoice number, date, supplier, line items, totals, tax) despite variations in field naming and placement across suppliers.
Table extraction. Extract structured tabular data (line items, positions) even when layout varies significantly.
Signature detection. Detect signatures on contracts and forms; validate against expected signatories where relevant.
Handwriting recognition. Extract handwritten text from forms and annotations. Accuracy improved materially through 2025-2026 with vision-language models.
Confidence scoring per field. Each extracted field returns with confidence score. High-confidence auto-approves; low-confidence routes to human review.
Human review routing. Documents with low-confidence extractions route to human reviewer with pre-populated form for verification. Corrections feed model improvement.
Downstream integration. Structured output pushes to accounting system (Sage, Xero, QuickBooks, NetSuite), ERP (SAP, Microsoft Dynamics), CRM, or custom database.
The Gartner intelligent document processing research publishes ongoing analysis showing IDP as one of the fastest-growing enterprise automation categories through 2026.
Rossum vs Hyperscience vs Azure vs Custom
Six viable options in the 2026 landscape.
Rossum. Cloud IDP focused on invoices and financial documents. £2k-£25k/month by volume. Best for mid-market with invoice-heavy workflow.
Hyperscience. Enterprise IDP with high accuracy claims. £5k-£50k/month plus setup. Best for enterprise with high volume and strict accuracy requirements.
ABBYY. Established OCR-plus-IDP vendor. £2k-£30k/month by tier. Best for teams already using ABBYY OCR wanting to add IDP.
Azure Document Intelligence, AWS Textract, Google Document AI. Cloud IDP with per-page pricing (£0.50-£10 per 1,000 pages). Best for teams already on those clouds with development capability.
Custom builds. LLM (Claude, GPT-4, Gemini) plus vision models plus classification logic. Defensible for enterprise with unusual documents, high volume where per-page pricing exceeds custom TCO, or specific compliance requirements.
The Azure Document Intelligence documentation covers the cloud IDP model in detail; most teams evaluating cloud IDP start there or with AWS Textract for comparison.
Real Accuracy Rates (and What They Mean)
Accuracy varies dramatically by document type.
Structured documents (standard forms, template invoices). 95-99 percent field-level accuracy. Straight-through processing rates 80-90 percent.
Semi-structured (variable invoice formats across suppliers). 85-95 percent field-level accuracy. STP rates 60-80 percent.
Unstructured (contracts, emails, freeform). 70-90 percent accuracy with LLM assistance. STP rates 30-60 percent.
Handwritten documents. 75-90 percent depending on quality. Always requires human review for financial or legal outputs.
Poor scan quality drops accuracy 5-15 percent regardless of document type. Scan quality is often the biggest driver of production accuracy.
What this means in practice. 95 percent accuracy on 400 invoices per week means 20 errors per week. If each error costs 30 minutes of reconciliation, that is 10 hours weekly. Auto-approve without review only makes sense above 98 percent accuracy on non-critical fields. Financial and legal outputs always require human review regardless of accuracy claim.
Real 2026 Cost Bands
Cloud IDP annual cost. Rossum £24k-£300k, Azure Document Intelligence £5k-£120k, AWS Textract £5k-£100k, Google Document AI £5k-£100k, ABBYY £24k-£360k. All depend heavily on volume.
Enterprise IDP annual cost. Hyperscience £60k-£600k plus setup, ABBYY Enterprise £60k-£360k.
Custom build cost. Proof of concept £40k-£90k over 8 weeks, pilot £100k-£250k over 3-5 months, production £250k-£800k over 6-12 months. Add £30k-£120k annual run cost.
Break-even. Cloud IDP wins for most mid-market volumes (up to 100k documents monthly). Enterprise IDP or custom becomes defensible above 500k documents monthly with unusual requirements.
The Human-in-the-Loop Reality
Human-in-the-loop is essential, not optional.
Straight-through processing (STP) rates. Percentage of documents that process end-to-end without human review. Typical: 60-85 percent for standard invoice workflow, 40-70 percent for contract processing, 30-60 percent for unstructured handling.
Human review time. For documents routing to review, time drops from 3-8 minutes fully manual to 30-90 seconds (reviewer confirms or corrects pre-populated fields). This is where the productivity gain comes from.
Total productivity gain. 60-85 percent reduction in manual document processing time. Accuracy improves alongside productivity because IDP catches errors humans miss (transposed digits, missed fields).
Continuous improvement loop. Human corrections feed back into model. Accuracy improves 5-15 percent over the first 6 months of production. Design for this from day one.
When STP goal fails. Teams setting 95 percent STP target for unstructured documents fail. Realistic target is 60-85 percent for structured, lower for unstructured. Set expectations correctly.
What We Learned Building with IDP
We built IELTSArena, a production AI platform serving students daily with confidence-aware output design (each AI assessment includes a confidence score with routing rules). Two lessons transfer to IDP deployments.
Confidence-aware routing beats accuracy-alone optimisation. Teams focused on maximum accuracy for all documents miss the power of routing by confidence. Auto-approve high-confidence extractions (60-70 percent of volume), route medium-confidence to human review (25-30 percent), reject low-confidence for re-scan (5-10 percent). Total workflow productivity beats pure accuracy optimisation.
Continuous improvement loop is the compounding advantage. IDP with human review feedback improves accuracy 5-15 percent over the first 6 months of production. Design for reviewer corrections to feed model retraining or LLM prompt refinement from day one.
You can see our shipped work at our portfolio. If you want a candid conversation about your IDP strategy, book an IDP scoping call with WhiteStone.
Common Failure Modes
Three failure modes we see repeatedly.
Setting unrealistic accuracy targets. Team promised 99 percent STP on unstructured documents. Reality is 60-80 percent. Project labelled a failure despite significant productivity gain over manual.
Skipping human review workflow. Team auto-approves all IDP output. Errors reach downstream systems. Reconciliation cost exceeds IDP savings. Always add human review for financial and legal outputs.
Choosing enterprise IDP before validating. Team commits to £300k Hyperscience deployment before validating concept. Cloud IDP (Azure DI, Rossum) would have validated at 10 percent of cost. Enterprise becomes appropriate after validation, not before.
Frequently Asked Questions
How is IDP different from OCR in 2026?
OCR converts image to text without understanding structure. IDP combines OCR with AI (increasingly LLMs) to classify documents, extract structured fields, and handle variable formats. LLM integration in 2025-2026 pushed IDP accuracy on unstructured documents from 50-70 percent to 70-90 percent, enabling processing of freeform documents traditional template-based extraction could not handle.
How accurate is AI document extraction today?
Structured documents (standard invoices): 95-99 percent field-level accuracy. Semi-structured: 85-95 percent. Unstructured (contracts, emails): 70-90 percent with LLM assistance. Handwritten: 75-90 percent. All rates assume good scan quality; poor scans drop accuracy 5-15 percent. Financial and legal outputs require human review regardless of accuracy claim.
Rossum, Hyperscience, or custom IDP?
Rossum (£2k-£25k/month) for mid-market with invoice-heavy workflow. Hyperscience (£5k-£50k/month) for enterprise with high volume and strict accuracy requirements. Custom (£250k-£800k build) for enterprise with unusual document types or above 500k documents monthly. Most teams benefit from cloud IDP validation (Azure DI, AWS Textract) before committing to enterprise or custom.
What does human-in-the-loop actually mean for IDP?
Straight-through processing rates run 60-85 percent for typical invoice workflow. Low-confidence extractions route to human review with pre-populated form. Reviewer confirms or corrects in 30-90 seconds (vs 3-8 minutes fully manual). Corrections feed back into model for continuous improvement. Not optional for production quality; essential design pattern.
How much does AI document processing cost in 2026?
Cloud IDP annual: Rossum £24k-£300k, Azure DI £5k-£120k, AWS Textract £5k-£100k, ABBYY £24k-£360k. Enterprise IDP annual: Hyperscience £60k-£600k, ABBYY Enterprise £60k-£360k. Custom: POC £40k-£90k, pilot £100k-£250k, production £250k-£800k. Cloud IDP wins for most mid-market volumes; enterprise or custom defensible above 500k documents monthly.
The One Thing to Remember
IDP has matured materially in 2026 with LLM integration pushing unstructured document accuracy from 50-70 percent to 70-90 percent range. Real accuracy varies by document type; do not accept vendor claims measured on ideal documents. Cloud IDP (Rossum, Azure DI, AWS Textract) covers most mid-market needs cost-effectively. Enterprise IDP or custom becomes defensible at high volume with unusual documents. Human-in-the-loop workflow is essential, not optional; design for the reviewer feedback loop to capture 5-15 percent accuracy improvement over the first 6 months of production.
If you want a candid conversation about your specific IDP decision, browse our AI development services or come to the call.
.webp)
.webp)