Design principle

The goal is not 100% automation. It is confident automation of the easy majority, with uncertain cases routed to a person. A system that knows what it does not know is far more valuable than one that always produces an answer.

Document processing is one of the most reliable AI use cases — high volume, structured output, and machine-verifiable results. It also has specific failure modes worth understanding before you commission anything.

What the pipeline actually contains

  1. Ingestion — email attachments, scans, uploads, API deliveries.
  2. Preprocessing — deskewing, rotation, quality assessment, page splitting.
  3. Classification — what kind of document is this? Invoice, PO, delivery note, contract.
  4. Extraction — pulling the specific fields that document type requires.
  5. Validation — do the numbers add up? Does the supplier exist? Is the date plausible?
  6. Confidence scoring — how sure are we, per field?
  7. Routing — high confidence proceeds, low confidence goes to review.
  8. Posting — writing to your ERP, accounting system or database.

Steps 5 and 6 are what separate a production system from a demo. Extraction alone gives you values with no indication of reliability. Validation catches errors the model cannot see — a total that does not match the line items is wrong regardless of how confident the extraction was.

Validation is your best accuracy tool

Deterministic checks catch errors more reliably than any model confidence score:

  • Arithmetic — do line items sum to the stated total? Does tax match the rate?
  • Reference matching — does this supplier exist in your master data? Does the PO number resolve?
  • Format checks — dates plausible, currency codes valid, tax numbers well-formed.
  • Range checks — is this amount within an expected band for this supplier?
  • Duplicate detection — has this invoice number already been processed?

A model that is 97% accurate plus arithmetic validation is more trustworthy than a model claiming 99% with no checks — because you know which 3% to look at.

Confidence routing in practice

SituationRoute
All fields high confidence, validation passesAutomatic processing
One field low confidence, rest fineReview that field only
Validation failure (totals mismatch)Full human review
Unknown document typeClassification queue
Unreadable scan qualityRequest a better copy
Unknown supplierReview plus master data update

Design the review interface as carefully as the extraction. Show the document alongside the extracted values, with uncertain fields highlighted and the relevant region of the page indicated. A reviewer who can confirm a document in ten seconds makes the whole system viable; one who must hunt across pages does not.

Where accuracy actually degrades

  • Poor scan quality — the single largest factor. Skewed, low-resolution or photographed documents.
  • Layout variation — hundreds of supplier formats, each different.
  • Multi-page documents where a table continues across pages.
  • Tables generally — line item extraction is materially harder than header fields.
  • Handwriting — variable and often unreliable.
  • Multiple languages or mixed-language documents.
  • Stamps and annotations overlapping printed text.

Piloting properly

  1. Collect 100–200 real documents, deliberately including the awkward ones — not a curated clean set.
  2. Have a person extract them manually to establish ground truth.
  3. Run the system and measure field-level accuracy, not document-level.
  4. Categorise the failures — quality, layout, field type. This tells you what to fix.
  5. Tune confidence thresholds against your tolerance for review volume versus error rate.

Field-level measurement matters. A document with nine correct fields and one wrong is not 90% useful if the wrong one is the total.

Modelling the return

InputNotes
Documents per monthFrom your current volume
Minutes to process manuallyTime a sample honestly
Expected straight-through rateStart conservative — 60–75%
Review time for flagged documentsFaster than full manual entry
Current error rate and its costOften the larger saving

Error reduction is frequently worth more than time saved, particularly where a mis-keyed invoice causes payment problems or a supplier dispute.

Where this fits with your systems

Extraction is only useful if the output lands somewhere. Plan the write path early — into your ERP, accounting system or database — including how a corrected review is written back and how duplicates are prevented. Projects that treat integration as a later phase frequently stall there.

Processing documents manually at volume? Send us a sample of what you receive — the document quality determines what is realistic. See our AI agent service, back-office automation, and human-in-the-loop design.

Frequently asked questions

On clean, consistently formatted documents, field-level accuracy in the high nineties is achievable. On poor scans, handwriting, or highly variable layouts it drops substantially. Any vendor quoting a single accuracy figure without asking to see your documents is quoting a number, not a prediction.
Yes, and designing for that is the point. Confidence scoring routes uncertain extractions to a human while high-confidence ones pass through. Aiming to remove humans entirely produces a system that fails silently on the cases it should have flagged.
Considerably harder and much less reliable. Modern models handle clear handwriting reasonably and struggle with anything else. If handwriting is a meaningful share of your volume, pilot on real samples before committing.