Services / AI / Document intelligence

Documents that read themselves, and admit when they're unsure

Every invoice, delivery note, claim form and signed contract that arrives as a PDF becomes somebody's afternoon. We build the layer that reads them, pulls out the fields your systems need, checks the result against your own records, and sends only the doubtful ones to a person.

Start the conversation →See the technology

Why it matters

The cost is the wait

The cost of paperwork is hidden because it arrives thinly sliced, ten minutes typing an invoice into the ledger, twenty checking a delivery note against a purchase order, an hour on Friday chasing the one form somebody completed by hand. Nobody's job title is "retyping", so it never appears as a line in a budget. Added up across a year it is usually the largest pile of recoverable time in the business, and the one nobody has ever measured.

What changed recently is worth understanding, because a lot of organisations tried this a decade ago and were disappointed. Older systems could turn a scan into text but could not tell you which number on the page was the total; they needed a template per supplier and broke the week somebody moved a logo. Modern document models read layout, tables, stamps and handwriting, and cope with a document they have never seen before, which matters enormously, because your suppliers are never going to agree on a format.

What we build is not a model, it is a pipeline. Documents arrive by email, upload, scanner or portal. They are classified and split. Fields are extracted with a confidence score attached to each one. Those values are validated against your own data (does this purchase order exist, does the arithmetic work, have we seen this invoice number before) and then either flow straight into your finance or operations system, or land in a review queue where a person corrects one field instead of typing thirty.

What you actually get

Built to be trusted

Reading the page is the easy part now. Everything below is what makes the output safe to post into a system your team actually trusts.

Reads the formats you really receive

Digital PDFs, flatbed scans, phone photographs taken at an angle, email bodies, spreadsheets, handwriting. Multi-document files are split, rotated pages are straightened, and each document is classified before anything is extracted from it.

A confidence score on every field

Not one score for the document. One per field. The supplier name can be certain while the bank details are not, and that distinction is exactly what lets the routine majority post automatically while the risky fields get ten seconds of human attention.

Checked against your own records

Extracted values are validated before they move: does the purchase order exist, do the lines sum to the total, is this supplier known, has this invoice number already been paid. This catches the model's mistakes and your suppliers' mistakes in the same pass: duplicates and overcharges included.

A review queue built for speed

Doubtful documents land in a screen showing the original alongside the extracted fields, each one highlighted where the value was found on the page. Confirming or correcting takes seconds, and the reviewer never retypes a document from scratch.

Straight into the systems you already run

Posted into your accounting package, ERP, claims system or database through their own interfaces, with the original file archived and linked to the record so anyone can pull up the source of a figure without asking who filed it.

It improves from your corrections

Every correction a reviewer makes is captured as a labelled example, so the fields your team keeps fixing are the fields that get better. Accuracy is re-scored against a held-back set of documents rather than assumed to be holding steady.

Where it earns its keep

Same pattern, different desks

The document changes from one industry to the next; the shape of the problem does not, a queue of paper in front of people who were hired to do something else.

01 · Accountancy & bookkeeping

Purchase invoices, without the typing

The problem: A month-end pile of supplier invoices in twenty different layouts, arriving by email, by post and through three separate supplier portals. Each is opened, read, matched to a purchase order, coded to a nominal account and keyed into the ledger, and the ones that arrive late are the ones that cause the arguments.
What we build: Invoices are collected automatically from a shared mailbox and folder, classified, and extracted down to line level: supplier, date, invoice number, net, VAT, total, and each line with its own code. Everything is matched against open purchase orders and screened for duplicates before a single entry is posted.
What changes: Clean invoices post themselves into the accounting system, and only genuine exceptions reach a person. A price that does not match the order, an unrecognised supplier, an invoice number that has been seen before.

02 · Insurance & claims

First notification of loss, read on arrival

The problem: Claims arrive as a mixture of forms, photographs, repair quotations, medical letters and email threads. Somebody opens each one, works out what it is, keys the details into the claims system and decides where it goes next, while the customer waits and the service-level clock runs.
What we build: Every attachment is classified and read the moment it arrives (policy number, incident date, amount claimed, third parties involved) and cross-checked against the policy record. Photographs are described and attached to the right claim. Anything ambiguous, or anything above a value threshold, is routed to a handler rather than progressed automatically.
What changes: Claims are triaged within minutes instead of days, handlers open a file that is already populated, and the cases that genuinely need human judgement are identified early rather than after a week in a queue.

03 · Logistics & freight

Proof of delivery that files itself

The problem: Delivery notes come back signed and photographed on a phone in a lorry cab: creased, half in shadow, occasionally upside down. Each has to be matched to a consignment before the customer can be invoiced, and the ones that go missing become a credit-note argument weeks later.
What we build: Drivers' photographs are read as they upload: consignment number, date, whether a signature is present, any handwritten note about damage or a short delivery. Then matched automatically to the job. Unreadable images are rejected back to the driver while they are still at the drop, rather than discovered in the office three days later.
What changes: Invoicing happens on the day of delivery, disputes are settled by pulling up the original image in seconds, and the paperwork the office used to chase largely stops needing to be chased.

The technology

The tools behind it, named

Extraction is three things working together, a model that reads layout, rules that check the result, and a queue for the cases nobody should guess at. We choose each per document type, weighted by what a mistake on that document would actually cost you.

6 layers · 30 technologies

Models that read documents

The reading itself. Which model suits depends on the document: a printed invoice, a handwritten intake form and a forty-page contract are three genuinely different problems.

Anthropic Claude
OpenAI
Google Gemini
Mistral
Hugging Face
Ollama

Specialist extraction services

Where a document type is high volume and well understood, a dedicated extraction service is often cheaper and more accurate than a general model, and we will tell you when that is the case rather than reaching for the interesting option.

Amazon Textract
Google Document AI
Azure Document Intelligence
PaddleOCR
Tesseract OCR

Checking the answer

An extracted value is worthless until it is valid. Every field is forced into a strict schema and validated against your real records before anything is allowed to move downstream.

Python
Pydantic
TypeScript
Zod
PostgreSQL

Where documents come from, and go

Most of the engineering is plumbing: collecting from mailboxes, folders and portals, then posting the result into the system your finance or operations team already lives in.

Gmail
Google Drive
Xero
QuickBooks
Sage
n8n

The review screen

Built for one job: confirm or correct a handful of highlighted fields as fast as a person can read them, with the original document alongside.

Next.js
React
Supabase
Redis

Measuring accuracy

A held-back set of your documents with known correct answers, re-scored on every change, so a model upgrade can never quietly make things worse without anyone noticing.

Langfuse
pandas
Grafana
Sentry

Product names and logos are the property of their respective owners and are shown to describe the technologies we work with. Their use does not imply any partnership, sponsorship or endorsement.

How we deliver it

Live behind a human first

Four to eight weeks to a first document type running in production. Every additional type after that is faster, because the collection, review and posting plumbing already exists.

We take a hundred of your real documents

Not samples and not tidy examples: the actual mess, including the bad scans, the awkward supplier and the form everybody complains about. That set defines what "working" means on this project and becomes the test we score against later.

We agree what has to be right

Field by field: which values must be perfect, which can tolerate an occasional correction, and what a mistake would genuinely cost. An invoice total and a delivery reference deserve very different thresholds, and treating them the same makes the system either too slow or too risky.

We build the pipeline

Collection, classification, splitting, extraction, validation against your records, and the route into your finance or operations system. Every document keeps a link back to its original file, so no figure in your system is ever unverifiable.

We score it before it touches anything

The pipeline is run against those hundred documents with known answers and we report per-field accuracy as numbers, not adjectives. With that in front of you, you decide which fields are allowed through without a human and which are not.

Humans first, automation second

It launches with everything routed through the review queue. As the per-field scores prove themselves on live documents, thresholds are raised and the queue empties, with your agreement at each step, and the ability to pull any field back to manual review at any time.

It keeps being measured

Corrections are logged, accuracy is reported monthly, and new supplier layouts or form revisions are added as they appear. A document pipeline nobody monitors decays quietly, so we make the decay visible.

Before you commit

The questions worth asking

How accurate is it, honestly?

It depends far more on your documents than on the model. Clean digital PDFs in a familiar layout are routinely read more accurately than a person typing them at four o'clock on a Friday. A creased fax of a handwritten form is not, and anyone who has actually done this work will tell you the same. That is why we score your real documents before proposing a design, and why every field carries a confidence value. The goal is not perfection, it is knowing reliably which fields you can trust and which need ten seconds of a human's attention.

What about badly scanned or handwritten documents?

They are handled, but with a lower automation rate and more of them landing in the review queue, and we would rather quote that honestly than promise a number we cannot hit. We also fix causes rather than only effects: rejecting an unreadable photograph back to the person taking it, while they are still standing in front of the thing, is worth more than any model upgrade.

Do our suppliers have to change how they send things?

No, and coping with the formats you already receive is most of the point. That said, if a large supplier can send structured data instead of a PDF, that is always cheaper and more reliable than reading it, and we will point that out rather than build extraction you did not need.

Is it safe to post straight into our accounting system?

Only for the fields and thresholds you have agreed, and only after validation against your own records: a purchase-order match, a duplicate check, arithmetic that adds up. Everything posted keeps a link to the source document and a record of what was extracted, by which version of the pipeline and with what confidence, so a disputed entry can be traced rather than argued about.

Where do our documents go?

That is decided with you before anything is built. Options run from commercial APIs with contractual zero-retention terms, through processing pinned to a particular region for residency requirements, to models running entirely inside your own infrastructure so nothing leaves the building. You get a written data-flow map suitable for handing to a client's security reviewer, and an honest account of what each option costs in accuracy and money.

We only receive a few hundred documents a month. Is it worth it?

Sometimes not, and we would rather say so in week one than in month six. At low volumes the payback is slower, and occasionally a much cheaper change wins outright: a better form, a supplier portal, a structured template for your five biggest senders. We will model the hours against the build and running cost and tell you if the numbers do not work.

Send us a hundred of your worst documents

The bad scans, the handwriting, the supplier whose invoices nobody can read. We will run them, show you the per-field accuracy, and tell you honestly which ones are worth automating.

Start the conversation →

Explore more

The rest of AI & Automation

Assistants & chatbots Knowledge search Workflow automation Team copilots Custom AI features ← All of AI