By use case: Document extraction
AI document extraction
PDFs, forms and emails read into structured records, checked, and filed where your team already works.
- First workflow live
- Intake, usually inside the first sprint
- Usually owned by
- Finance and accounts teams keying invoices and statements
- Built on
- n8n, Make, Claude or OpenAI
- Judged on
- Straight-through rate and Field accuracy
In short
AI document extraction reads incoming documents such as invoices, purchase orders, statements, forms and emailed requests, pulls out the fields you need, checks them against your own records, and writes them into your ERP, accounting tool, CRM or database. SBD builds it with a confidence score on every field. Clean documents go straight through, and anything uncertain or inconsistent waits for a person with the source page beside the extracted values.
Who this is for
The people who usually own document extraction.
- Finance and accounts teams keying invoices and statements
- Sales operations reading emailed enquiries and orders
- Back-office teams processing forms and applications
- Operations managers who want a record of every document received
Is this you?
If three of these sound like your week, this page is about your team.
- People copy numbers from PDFs into another screen every day
- An OCR tool was bought and the templates went stale
- Typing errors are found at month end
- Nobody can say how many documents arrived last week
- One person is the only one who understands a supplier format
What it costs today
Where document extraction loses the hours.
Retyping what is already written down
Someone opens a PDF, reads it, and types the same numbers into another system, one field at a time.
Every sender uses a different layout
Template-based OCR breaks the day a supplier changes their form, so the team goes back to typing.
Errors surface weeks later
A wrong quantity or a swapped digit is caught at reconciliation, month end or by an annoyed customer.
Documents live in inboxes
Attachments sit across personal mailboxes and shared drives, with no record of what arrived or what was done with it.
For back-office teams
Stop typing what is already written.
Send us twenty real documents of the type you key most. We will show you what reads cleanly, what needs a person, and what should stay manual.
Send us a sample
By company size
Document extraction at 10 people, and at 1,000.
- 1 to 10
The founder or one administrator types documents in at night.
- 11 to 50
A shared inbox of attachments and one person who knows how each sender lays things out.
- 51 to 200
Several teams retyping different document types, an OCR tool that needs templates nobody maintains.
- 201 to 1,000
High volumes across entities, audit requirements on every record, and a backlog at every month end.
What we build
What we build for document extraction teams.
- 01
One intake for every document
A watched inbox, upload form or folder. Each document gets a record the moment it arrives, so nothing is lost in a mailbox.
- 02
Extraction with field-level confidence
An AI step reads layout and text together, returns the fields you define, and scores each one. No per-sender templates to maintain.
- 03
Checks against your own data
Totals re-added, dates checked, part numbers and customers matched to your master data. A mismatch is an exception, not a silent write.
- 04
A review screen for exceptions
An internal tool showing the document beside the extracted values, with low-confidence fields highlighted. A person corrects, approves and the record is written.
The workflows
The document extraction workflows, and what sets each one off.
- 01 Intake
When a document arrives by email, upload or shared folder
ThenLogged with sender, type and a link to the file
- 02 Extract
When a document is logged
ThenFields read and scored, line items split out
- 03 Check
When extraction completes
ThenTotals, dates and codes checked against your master data
- 04 Review
When a field is below the line or a check fails
ThenSent to the review screen with the page beside the values
- 05 Write
When all checks pass or a person approves
ThenRecord created in the ERP, accounting tool or database
The numbers we move
How a document extraction build is judged.
Every build has a number written into the blueprint. These are the ones document extraction builds are measured by, and how we measure them.
| Metric | Before | After | How it is measured |
|---|---|---|---|
| Straight-through rate | Every document typed | Most clean documents need no edits | Share of documents written without any field corrected, per week |
| Field accuracy | Unknown until errors appear | Measured per field | Fields corrected in review divided by fields extracted, sampled weekly |
| Time from arrival to record | Days in a busy week | Same day | Median time from document received to record written |
| Exceptions waiting | Invisible in inboxes | One counted queue | Open documents in the review queue and their age, daily |
Documents and records
The paperwork that generates itself.
Names and tax rules differ by country. We build for the ones you file under, whether that is VAT, sales tax or GST.
Systems we connect in document extraction
What is usually already there, and what we add.
Anything with an API or a database we can read works the same way. Each tool links to its page in the integrations directory.
- Gmail
- O Outlook
- Google Drive
- S SAP
-
Xero - QuickBooks
-
Supabase -
Airtable
Proof
Proof from builds like this.
- Customer story
Enquiry and quotation platform on SAP
Emailed RFQs read by AI and parts matched against the SAP item master, from enquiry to order.
Read the story - Customer story
Property operations automations and agent fleet
Mortgage statements read, matched to the property and reconciled, where they were reconciled by hand before.
Read the story
How it starts
An audit of the document extraction work your team still does by hand.
One to two weeks. We map the workflows, put an hours-a-week figure on each, and give you a costed blueprint in order of payback. The audit is credited in full to the build.
- Week 1 to 2Audit. We sit with the people who do the work and map it.
- Week 3Blueprint. Fixed scope, fixed price, the first build chosen by payback.
- Week 4 onBuild in two-week sprints, working software every second Friday.
- After go-liveSixty days of hypercare, then a care plan or run it yourself.
What this does not do
So the scope is clear before the call.
- A handful of documents a month. Typing them is cheaper.
- Documents where a wrong value carries legal weight and no person will review. We keep a person on those.
- Unreadable sources. If a person struggles to read it, the model will too, and we will say so.
Terms, defined
Document extraction terms, in plain English.
- Intelligent document processing
- Reading documents of varying layout into structured data using models rather than fixed templates.
- Straight-through processing
- A document that goes from arrival to record with no person touching it.
- Field-level confidence
- A score for each extracted value, so one doubtful field can be checked without redoing the rest.
Questions
What people ask about document extraction automation.
How is this different from OCR?
OCR turns an image into text. Extraction decides which text is the invoice number, which is the total, and which line items belong together, across layouts it has not seen before. We use OCR underneath where scans need it.
How accurate is it?
It depends on your documents, so we measure it on a sample of yours during the audit and report field by field. Nothing writes to your systems without passing the checks or a person.
What about handwritten or poor scans?
Clear handwriting and decent scans usually read well. Faint, skewed or partial pages get low scores and go to review rather than being guessed.
Where does our data go?
Documents are stored in your own accounts, and we use model providers whose business API terms say inputs are not used for training by default. Where data must stay in a region, we check which providers and hosting options meet that during the audit.
How much does it cost?
The $500 Automation Audit comes first and is credited to the build. One document type into one system is often around $2,500. Several types with a review tool usually sits in the $5,000 to $15,000 range.
Reading
- Systems
Invoice intake with an agent: the first thirty days
The first month of an invoice intake agent from the inside: how the queue behaves, which invoices escalate, and the three settings that decide whether it lasts.
- Strategy
Automation vs AI systems: what moves the number
Stacking task automations stalls at around twenty. The difference between a task and a system, the five layers that hold it up, and what to rebuild first.
- Systems
How much should an AI agent be allowed to do?
An agent with an unscoped token dropped a production database and its backups in nine seconds. Our rule: read widely, write narrowly, cap spend, escalate.
Related
- Use case
Invoice processing
Supplier invoices read as they arrive, matched to the order and the goods received, approved by the right person and posted to your accounting system. People only see the ones that do not match.
- Use case
Order to cash
Orders entered once, fulfilment visible to everyone, invoices raised the moment goods leave, payments matched as they land and reminders sent before anyone has to ask.
- Function
Back office
The admin layer that quietly consumes your team’s week, handled by systems instead.
- Function
Finance and accounting
Close faster, chase less, and see the numbers without waiting for month end.
By use case: Invoice processing, Lead routing, Customer onboarding, Employee onboarding, Quote generation, Order to cash, Ticket triage, Approval workflows, Inventory sync, Weekly reporting, Client portals
Document extraction
Get the document extraction hours back.
Bring the one document extraction workflow your team still runs by hand. You leave the call with a yes or no on fit and the first thing we would build.
Book the call