Sites that load fast and convert
Custom builds and integrations
Bounded, auditable, human-approved
Being found, staying fast, staying safe
Invoices, purchase orders, receipts, delivery notes, application forms. Somebody in your business opens each one, reads the same six fields, and types them into another system. It is dull, it is slow, and it is where a surprising number of errors originate.
Extraction reads the document, pulls out the fields that matter, and hands them to your accounting or operations system in the right shape. The part that makes it usable rather than alarming is what happens next: every field carries a confidence score, and anything the system is not sure about stops for a person instead of quietly entering a wrong number.
Move over the page ↓Reality
Every business assumes their documents are basically standard. Nobody's are. These six problems account for most of the work in any extraction project.
Forty vendors, forty layouts. The total sits bottom-right on one, top-left on another, and is labelled "Amount Due", "Balance", "Total Payable" or nothing at all.
Why it matters: template-based tools break the moment a supplier redesigns.
A phone picture of a receipt taken at an angle, in bad light, slightly creased. Perfectly readable by a person, considerably harder for software.
Mitigation: de-skew, enhance, and lower confidence rather than guess.
Line items spanning pages, merged cells, subtotals mid-table, a column that only exists on some invoices. Naive text extraction turns this into unusable soup.
Reality: line-item extraction is roughly twice the work of header fields.
Delivery notes, job sheets, signed forms. Accuracy on handwriting is far below print, and it varies wildly between one person's writing and another's.
Approach: extract, but assume a person confirms.
Emailed, then posted, then re-sent because nobody replied. Detecting that three files are one invoice is its own problem, and paying twice is expensive.
Needed: deduplication before anything reaches your ledger.
A date read as 03/04 could be March or April. A total read correctly but assigned to the wrong currency. The extraction was accurate and the record is still wrong.
Fix: validation rules, not better reading.
Interactive
Three documents of decreasing quality, run through the same system. Hover a field to see where it came from on the page, and drag the threshold to change how cautious the system is.
This control is the entire commercial trade-off in one slider. Raise it and errors approach zero while the review pile grows; lower it and almost everything posts automatically, including the occasional mistake. Where you set it depends on what a wrong number costs you — which is a business decision, not a technical one.
The stack
Extraction is a pipeline, not a product. Documents arrive somewhere, get read, get checked, and land in the system that actually runs your business — usually one of the accounting platforms below.
Where documents arrive
Reading the page — OCR & vision
Understanding & extraction
Structuring & validation
Where the data lands
Records, search & storage
Applications
Ordered roughly by how quickly they pay for themselves. The first is where most projects start, and usually for good reason.
The classic. Header fields, line items, tax treatment, and a match against the purchase order and the goods received note before anything is approved.
Highest volume, clearest payback
Photographed receipts, reconciled against card transactions, categorised and coded — with the original image attached to the entry for the auditor.
The one staff notice immediately
Incoming orders from customers who send PDFs rather than using your portal, turned into sales orders without anyone rekeying part numbers.
Wholesale, distribution, manufacturing
Packing slips, bills of lading, proof of delivery — frequently handwritten, photographed in a yard, and needed against the invoice later.
Where paper still genuinely rules
Applications, registrations and intake forms that arrive as scans or PDFs and currently get retyped into a CRM by whoever has the time.
Clinics, brokers, professional services
Key dates, notice periods, values and parties pulled out of agreements so renewals stop being discovered a month after they auto-renewed.
Extraction, not interpretation
Honesty
Extraction projects fail in predictable ways, and almost all of them are visible at the start if anyone bothers to look.
Any vendor quoting perfect accuracy is quoting it on clean, printed, single-layout documents. On a real mixed pile, the honest target is a high automation rate with a review queue you can live with.
Below roughly a few hundred documents a month, the build and supervision cost more than the typing. We will say so.
Flagged documents need a person who checks them daily. If nobody owns that, the queue grows until the system is abandoned.
Reading the invoice is the easy half. Getting it into an accounting system with a closed or awkward API is frequently where the real effort sits.
If approvals are unclear or coding rules live in one person's head, automating the typing just moves the bottleneck. Sometimes the process needs fixing first.
Print extraction is mature. Handwriting is not. If your documents are mostly handwritten, expect meaningful human review permanently, not temporarily.
Where originals are stored, for how long, and who can retrieve them. This is a records question with legal weight, and it belongs in the plan.
Without a labelled test set of your own documents, nobody can tell you what accuracy you will get — including us. Building one is part of the work.
If you only need standard invoices into a mainstream accounting package, an existing product may do it for a subscription. We would rather point you there than build it.
Scope
The document sample comes first. Everything else is shaped by what is actually in the pile.
A real sample across your suppliers and formats — including the awkward ones people apologise for. That pile determines the whole design.
Exactly which fields matter, their types, their formats, and what a valid value looks like. Written down, because "the total" is less obvious than it sounds.
Watching a mailbox, a folder, a scanner or an upload form, with the original preserved and linked to every record produced from it.
Working out what each document is before trying to read it, because an invoice and a statement need entirely different handling.
Pulling the fields with a confidence value on each, plus the region of the page each value came from so a reviewer can check it in one glance.
Arithmetic that must balance, dates that must be plausible, references that must exist. This catches the errors that confidence scores miss.
A screen showing the document beside the extracted fields, built for speed — because someone will use it every day.
Into your accounting, ERP or operations system, idempotently, so a retry never creates a duplicate entry.
Measured against a labelled set, reported per field, so you can see where corrections cluster and improve that field specifically.
Questions
On clean printed documents with a consistent field set, high — comfortably enough that most documents post without anyone looking. On photographed, handwritten or unusual layouts it drops, sometimes considerably.
The honest answer for your business needs a sample of your documents and a labelled test set. Anyone quoting a percentage before seeing your pile is quoting somebody else's.
Yes, for the flagged minority. The threshold slider above shows the shape of it: a higher bar means a bigger review queue and fewer errors reaching your ledger.
What changes is the volume. Reviewing the uncertain tenth is a different job from typing every document.
That is the point of the modern approach. Rather than a template per supplier, the system reads the document the way a person does — by understanding the labels and structure — so a new supplier usually works without configuration.
Genuinely unusual formats may still need attention, which is why the survey happens first.
Into whatever system you already run — accounting, ERP, CRM or a database — through its API. Posting is idempotent, so a retry after a network failure never creates a duplicate entry.
Where a system has no usable API, options are a scheduled file import or direct database writes, decided case by case.
Stored and linked to every record produced from them, so anyone reviewing an entry later can open the source in one click. That link matters more than the extraction for audit purposes.
Retention period and storage location are decided with you before launch.
Some. Printed handwriting on a form is often workable; cursive on a creased delivery note frequently is not.
Where handwriting dominates, we design for permanent human confirmation rather than pretending it will improve away.
For one document type into one destination system, usually weeks — with the document survey and the labelled test set taking a meaningful share of that.
Additional document types afterwards are considerably faster, because the pipeline already exists.
Usually some of it, if you use a cloud model or OCR service — configured so the provider excludes it from training. We will tell you exactly what leaves your systems and where it goes.
For sensitive documents, an on-premise or self-hosted model is possible. It costs more and is occasionally the right call.
Call back
Leave a number and a good time. We will call you back to talk about what you are trying to build — no charge for the conversation.
Or call us directly +1 (647) 385-5532