Extraction is the easy half. What decides whether a document pipeline is safe to run is the threshold: which fields are confident enough to post automatically, and which ones a person still checks. We make that a number you set, not a number we hide.
One document, two destinations
Intelligent document processing extracts structured data from invoices, contracts, KYC packs and forms using OCR and language models. Every field carries a confidence score, and anything below the agreed threshold routes to a human review queue instead of entering your systems unchecked.
The demo used five clean invoices. Production has eleven supplier formats, a fax and something photographed at an angle.
Ninety-something per cent, on a sample chosen by whoever was selling. Accuracy varies enormously by document type, field and scan quality, and an average conceals exactly the cases you need to plan for. Ask instead what the review rate is at a threshold you would accept.
The review interface shows a document and an empty form, so the person reads and keys exactly as before, with an extra login. A correct review queue shows the page image and the extracted values side by side, with low-confidence fields highlighted, so the work is correcting rather than transcribing.
Fields come out cleanly and then sit in a CSV because the accounting system has no interface. Getting verified data into the system of record is frequently the larger half of the project, and it needs integration work or a robot to key it.
Reviewers fix the same field on the same supplier's invoice every week and nothing improves, because corrections were never captured as labelled data. That feedback loop is the difference between a review queue that shrinks and one that is permanent staffing.
Twelve fields from one supplier invoice, with the confidence the extraction returned. Three of them are actually wrong. Lower the threshold to automate more and you will start approving them.
Fields scoring below this go to a person.
Posted automatically
4
of 12 fields
To a person
8
review queue
Wrong values approved without review
0
Nothing wrong is getting through
Every misread field is caught. The cost is a large review queue, which is the correct starting position for anything touching money or identity. Lower it deliberately once you have data.
Illustrative sample, not a benchmark. On a real engagement these numbers come from a labelled set of your own documents, which is why step two of the build is labelling rather than modelling.
Run this on your documentsThe extraction engine is a component you can swap. These are the parts that make it operable.
The page image and the extracted values side by side, low-confidence fields highlighted and focused first, keyboard navigation between them, and the source region on the page highlighted as each field is selected. A reviewer should be able to clear an item in seconds rather than minutes.
Every correction is captured as labelled data, which is what makes the queue shrink over time rather than becoming permanent headcount.
A mailbox receives invoices, credit notes, statements and things that are none of those. Classifying first means each document type gets the extraction schema and threshold that suits it, instead of one model attempting everything.
Does the supplier exist, does the total match the lines, does the PO reference resolve? Cross-checks catch errors that confidence scores miss, and they are cheap to add.
Built once, hand-checked, and used to measure every change afterwards. It is the only thing that makes an accuracy claim falsifiable, and it belongs to you at the end.
Invoices and identity packs carry personal data. Where residency rules bite, extraction runs entirely inside your network. Reviewed under AI governance and by our security team.
Five steps, and the first two are the ones that get skipped. Same delivery shape as the rest of the AI and automation practice.
Two to three hundred documents, deliberately including the bad scans, the unusual suppliers and the ones somebody photographed on a phone. A sample of clean examples produces an accuracy figure that does not survive its first week in production, and everyone knows it except the person who chose the sample.
A subset labelled by hand, field by field. This is tedious and it is the foundation of everything after it: without it, accuracy is an assertion and the threshold conversation has nothing to stand on. It also becomes an asset you keep.
AWS Textract, Azure Document Intelligence, Tesseract and layout-aware language models all win on different material. We measure rather than standardise, and the winner for invoices is frequently not the winner for identity documents.
We model the trade-off explicitly, per field, using the labelled set: what auto-approves, what queues, and how many errors slip through at each level. Finance signs the number, not engineering, and it is recorded as a decision with a date.
Side-by-side correction, captured labels, and monitoring on review rate and correction rate so a drop in quality is visible before somebody complains. Then handover, with the labelled data and the configuration assigned to you.
Each needs a different schema, a different threshold and a different definition of a bad day.
The most common starting point and the best understood. Header fields extract reliably; line items are where the difficulty lives, because tables wrap across pages, merge cells and use supplier-specific column orders.
The strongest control here is not the model, it is the three-way match against the purchase order and the goods receipt, which catches errors no confidence score would flag. Usually orchestrated as a workflow.
Hard part
Line item tables across pages
Best control
Three-way match, not confidence alone
Six stages between a file arriving and a record existing. Select one to see what it does and how it fails.
Files arrive from a mailbox, an SFTP drop, a portal or a phone. Everything is converted to a consistent image and page format, deskewed, and given a stable identifier so the same document arriving twice is recognised as a duplicate.
Fails when
Duplicate detection is missing and the same invoice enters twice, which is a finance problem long before it is a data problem.
The same document handled manually and through a pipeline. Notice that a person still appears in the second version, and that this is the design rather than a shortfall.
Around 20 minutes of work, spread across two days
Invoice sits in a shared mailbox
Until someone on AP duty opens it.
Header and line items keyed by hand
Reading from a PDF on one screen and typing into another.
Purchase order looked up separately
A second system, a second search, from memory of the reference.
Quantity mismatch noticed, or not
Depending on how busy the day is. This is the expensive branch.
Approver decides
The one step that genuinely required judgement.
What disappeared was the transcription and the two-day wait. What did not disappear was the judgement, and any proposal that claims otherwise is describing a different process. The wider version of this argument sits under business process automation.
Four sectors where extraction pays back fastest. Profiles are anonymised.
Identity documents, address proofs and application forms arriving in one bundle. Extraction plus cross-document consistency checking removes the assembly work, and the analyst spends their time on the judgement rather than the collation.
Banking and insuranceDozens of supplier formats into one accounting system, with three-way matching against purchase orders and goods receipts. The highest-volume and best-understood application in this list, and usually the first one built.
Manufacturing engagementsPhotographed on a phone at a loading bay, often with a signature and a handwritten annotation. Capture quality dominates results here, and improving how drivers take the photograph beats any model change.
Logistics engagementsArriving by fax, email and portal in equal measure. Extraction plus routing gets a referral to the right team the same day rather than whenever somebody opens the shared inbox. Personal data throughout, so residency is settled first.
Healthcare engagementsThree approaches frequently sold under the same name. The row that decides it is usually the third one.
| Criterion | Plain OCR | Template extraction | Intelligent document processing |
|---|---|---|---|
| Output | Raw text | Fields, per known layout | Named fields with confidence per field |
| New supplier layout | Works, but output is unusable | Needs a new template built | Usually works without configuration |
| Tells you when it is unsure | No | Rarely, and only per document | Yes, per field, which is what makes review possible |
| Handles tables | No structure preserved | If the template covers it | Yes, including continuation across pages |
| Ongoing maintenance | None | A template per layout, forever | Threshold tuning and periodic re-evaluation |
| Best when | You only need searchable text | Few layouts, never changing | Many layouts, and errors have consequences |
If you genuinely have three supplier layouts that never change, template extraction is cheaper and more predictable, and we will say so. The case for IDP is variety plus consequence, and the confidence score is what buys you the second half of that.
Extraction is one stage in a longer process. These are the neighbouring parts of the AI and automation practice.
The orchestration that carries a document from arrival through approval to posting.
For keying verified fields into the system that has no interface to call.
Where extracted data feeds a system that reconciles and decides rather than just records.
The retrieval and evaluation engineering behind clause extraction and document question answering.
Once documents are readable, they become a knowledge base worth asking questions of.
Structured document data is frequently the missing history a forecast needed.
For support queues where customers attach documents somebody currently opens by hand.
Where extraction has to run inside your own network, our cloud team provisions it. To start, send us a sample of your documents including the bad ones, or read the help centre.
The ten that come up in almost every first conversation.
Especially the bad ones. We will come back with a measured review rate at a threshold you would accept, rather than an accuracy percentage. Reply in under 3 minutes.
