Six PDFs, one job file, forty minutes of retyping.
The information already exists. Somebody is paid to move it from one screen to another without making a mistake, and occasionally they make one.
The problem, as it actually is
An order arrives as a PDF. A spec arrives as a photograph of a drawing. A delivery note arrives as a scan. Somebody opens all of them, types the details into your system, cross-checks the quantities against the order, and files the lot.
It is the most automatable work in most businesses and the least likely to be automated, because it looks like it needs judgement. Mostly it needs attention, which is exactly what a person loses at four in the afternoon on the eleventh one.
The errors are the real cost. A transposed quantity becomes a wrong delivery, which becomes a credit note, an angry phone call, and a day of someone's week.
What I build
- Documents read wherever they arrive
- By email, from a shared folder, or from a scanner.
- The fields you care about extracted
- Whatever the layout, including from scans and photographs.
- Checked against the order
- Not just captured. Quantities, prices, dates and references compared against what you were expecting, with mismatches raised.
- In your system
- As a job, order or record, in the shape your team already recognises.
- Low-confidence items queued for a person
- With the original and the extraction side by side, so checking takes seconds.
What it connects to
Email, shared drives or scanners for intake. Your job, order or ERP system for the output. Your document store for filing.
What it will not do
Guess at a field it cannot read. Accept a document that fails your checks. Discard an original: everything stays, and every extraction is traceable back to the page it came from.
What it takes
A representative stack of real documents, including the bad scans, and the rules for what makes one acceptable.
What people ask before they buy
- “Our documents are handwritten and terrible.”
- Some will be readable and some will not. The honest answer is a confidence threshold: clean ones flow through, poor ones go to a person with the work half done. The Blueprint measures which is which on your own paperwork, before you commit to anything.
- “We tried OCR and it was useless.”
- Plain character recognition is not what this is. The difference is checking against the order and routing anything doubtful to a human, rather than capturing confidently and being wrong.
- “The system we would put it into is ancient.”
- Usually fine. If it has no interface at all, the Blueprint says so plainly rather than discovering it in week three.