Documents in,
checked data out
We build document processing that reads your invoices, forms and contracts, ties each value to its place on the page and checks it against your records, so what reaches your systems is checked, and what’s unsure goes to a person.
Thanks, we’ll be in touch soon.
{{ heroStatus }}
The drawing: a new supplier’s form, each answer written by hand. Each answer is read in turn and ticked. One box was left blank by the sender, and a note beside it says: “Nothing written here. Left empty, not guessed.” One tap sends it to a person, and the note then reads: “With Purchasing, with the page attached.”
{{ heroSay }}
You’re in good company
A few of the companies we’ve worked with.
{{ wallSay }}
Read by a person, typed into a system
Much of what a business receives still arrives as documents: invoices, forms, contracts, claims. Someone opens each one, finds the values that matter, each one piece of information such as a date, a name or a total, and types them into a system, where they’re saved as a record. Tools that read by fixed rules need each layout to stay the same, and stop working when a supplier changes theirs. An AI model doesn’t need the layout to stay the same. But its answers read just as well when they’re wrong, and it can return a value that isn’t on the page at all. So the question isn’t whether AI can read your documents. It’s how you’d know when it hasn’t.
What we build
The reading, the checks each value has to pass before it’s saved, and the screen where your people settle what’s unsure.
Sorted by kind
Each file recognized as an invoice, a form, a contract or something else, and split where one file carries several documents.
Read as it looks
Print, handwriting, stamps, checkboxes and tables read from the page as a person sees it, on scans and photos as well as PDFs.
Each value, with its place
Every value saved with the page and the place it was read from, so checking it takes one look.
Tables and line items
Rows read as rows, each line kept together, across a page break where a table runs on.
Checked against itself
Lines that should add up to a total, dates that should come in order, a name that should match its number: each document is checked against what it says elsewhere.
Checked against your records
A supplier you already know, an order that exists, a document you haven’t received before, checked in your own systems before anything is saved.
A screen for what’s unsure
What doesn’t pass waits for a person, with the page on one side, the value on the other, and the place it was read from marked. A value that can’t be read, or isn’t on the page, is left empty, never guessed.
Sent on, with its record
What passes goes to your systems, with which values were read by AI, which a person checked or changed, and who.
Measured on your documents
A set of your own documents with agreed answers, run before launch and before every change of model, so you see what changed, value by value.
Moving the work on once the data is in? Workflow Automation →
AI inside your own product’s screens? AI Product Engineering →
Answers from your documents, for customers and staff? AI Chatbots & Assistants →
What the page shows, and what the file says
Most tools read the text stored in the file, because it’s quick. We read each page the way a person sees it, then compare that with the stored text. Where they agree, the value passes that check. Where they differ, a person sees both.
| What the page shows | What the file says |
|---|---|
| Delivery note | Delivery note, the same as the page |
| Deliver to: | Deliver to: Front deskDiffers from the page |
| A stamp: Received | Nothing |
Read as it looks: Back gate. Sent to a person, with both.
A scan is a picture of a page. Text you can select on a scanned page can be an earlier program’s reading of that picture, stored out of sight, and nothing in the file says whether anyone checked it.
Notes, stamps and ink added to a page later can be kept apart from its own text, or be only in the picture. A program that reads the stored text alone can miss them.
Stored text has no set reading order unless the file gives one, so columns and tables can come out in the wrong order.
Text stored out of sight can still reach a model, and a model can’t reliably tell the text it’s reading from an instruction hidden inside it. AI Security & Governance →
Sure isn’t the same as right
Many models give each value a score for how sure they are. A high score isn’t proof: a model can be sure and wrong. How far its scores can be trusted has to be measured on your own documents, not assumed. So a score alone never decides what’s saved. Each value also has to be found on the page, agree with the rest of the document and match your records. What passes is saved. What doesn’t waits for a person, with the reason beside it.
Which values always wait for a person, whatever the checks say, is yours to set.
An example
- On the page
- Barn Lane
- Read as
- Bam Lane
- The model was sure.
Your records have no Bam Lane for this customer. So it waits for a person.
Start with one kind of document
We begin with one kind of document, the one your team handles most. You give us a set of real ones, and your people agree the right answers for each. We build the reading and the checks for that kind, run them against the set, and go through the results with you, value by value: what was right, what was wrong, and what was left for a person. You get that one kind of document running beside your team’s usual way of working, a written note of each check and why it’s there, and the set itself, to run again before every change.
Your documents, and the data read from them, stay in your accounts and systems. If the model doesn’t answer, documents wait their turn, none is dropped, and your team can still enter one by hand. Where our engineers open your documents, it’s for a task you’ve agreed, and it’s recorded.
Work with us as a dedicated team, on a fixed price project or a mix of both. Compare engagement models →
How each is paid for: Pricing →
Questions worth asking
How accurate is it?
We don’t give a figure, for accuracy, speed or how much goes through without a person, before it’s been measured on your documents. After the first run you see the results value by value, for each kind of document. Then you decide which values can be saved once they pass their checks, and which always go to a person.
Is this a product, or something you build?
Something we build, in your systems. We don’t sell or resell a model or a document processing product. The reading is done by the model that reads your documents best, measured on your own set, and we build so that model can be changed: the checks, the screen for your people and your set stay the same, and the set is run again before any change goes live.
Can it read handwriting and poor scans?
It depends on the writing and the scan, so we measure it on yours. What can’t be read is left empty and sent to a person, never guessed.
What if a wrong value gets through?
The checks are there to stop it before it’s saved. No set of checks stops everything, and neither does a person typing. So every saved value keeps its place on the page and who checked it, and a correction takes one look, not a search.
Do our people still check everything?
No, and how much they check is yours to decide. Values that pass their checks can be saved without a person. Which values always wait for one is agreed in writing before launch.
Isn’t this just reading text from a scan?
That part is optical character recognition (OCR), and it’s one part of this. The rest is knowing what kind of document it is, which text is which value, whether each value passes its checks, and what happens when one doesn’t. Together it’s often called intelligent document processing (IDP).
Who sees our documents?
Your documents, and the data read from them, stay in your accounts and systems. What’s sent to a model’s provider, what the provider may do with it, and what’s kept and for how long, is agreed with you before launch, and written down. Where documents mustn’t leave your systems, the reading can be done by a model that runs in your own accounts, measured on your set like any other. Our engineers open your documents only for a task you agree, and each time is recorded.
What about personal data in the documents?
We read only the values the task needs, and leave the rest. What’s kept, where and for how long follows your rules.
Does it connect to the systems we use?
Yes. What passes its checks is sent to the systems you already have, through their APIs. Systems & API Integration →
What happens once the data is in?
Approvals, reminders and hand-offs are a workflow: Workflow Automation → Where each document needs a different next step, an agent may fit: AI Agents →
What does it cost?
It depends on the kinds of document, the values each one carries and the systems they go to. What each document costs to read is measured in the first run, and set as a limit before launch. The first call is free. Your proposal sets out the scope, the price or rate, and how you pay, before any work starts. Pricing →
Let’s ship it better
Tell us which document your team types in most. We’ll suggest where to start.
Thanks, we’ll be in touch soon.
{{ ctaStatus }}