Three hours a day of order entry, removed.
0101 / 09Order entry, per day 3hrs <1hr Order entry went from three hours a day to under one.
Supplier invoices arrive as PDFs, in whatever layout the supplier happens to use. AI reads them and writes them into the ERP. Four weeks from kickoff to live.
One person was the only route into the system of record.
Around 200 supplier invoices arrive each month as PDFs attached to emails. Every supplier uses a different layout, and creating each one in the ERP took several manual steps.
All of it sat with one person, who had other jobs to do as well. Order entry took two to three hours of their day, every day. Everything else they were responsible for waited until it was done.
EDI — the standard electronic format large suppliers use to send invoices machine-to-machine — solves this for suppliers big enough to implement it. Most of this client’s suppliers never will, and that is where the volume sits. So the invoices stayed manual, indefinitely.
No templates, and nothing re-keyed.
AI reads each PDF as it arrives and pulls out the fields the ERP needs. It works on a layout it has never seen before, so there are no templates to maintain and no supplier has to change anything on their side.
The output is a file the ERP loads. Nothing is re-keyed at any point between the supplier’s email and the system of record.
We did not build a document model from scratch, and nobody should.
The work is choosing the right components and making them hold together in a business that runs on them every day.
Does agentic document extraction, or ADE. It works out what each part of a document means — the invoice number, the line items, the totals — wherever they happen to sit on the page. A layout it has never seen before still works.
That is why we chose it. Older capture tools read the characters and then need a template for every supplier, which is exactly where invoice automation usually dies.
Structures what comes back and checks it against what the ERP will accept, so bad data is caught before it reaches the system of record.
Runs the pipeline end to end — the email trigger, the extraction call, the validation step, the handoff.
Puts the finished upload-ready document into a secure Microsoft environment, configured to load it into the ERP. It runs inside the tenancy the client already had, so there was nothing new for IT to approve or administer.
Extraction is not perfect, and a system that pretends otherwise fails quietly.
Which is worse. A portal shows every document the system could not read, and every one that failed on import. Someone corrects it and re-runs it from the same screen.
Nothing sits unnoticed in an inbox waiting to be chased at month end.
The point is not the hours. It is that a single person was the only route by which supplier invoices reached the system of record, and everything else they owned queued behind that.
Put your own invoice volume in.
This is the same work the ROI Lab models — document extraction replacing manual keying. It will give you a range.
At 200 invoices a month, the ROI Lab says a pilot takes over a year to pay back. This client was below that threshold and the work was still worth doing, because the constraint was one person’s time rather than the cost of the invoices.
A call with a principal.
To find what’s worth building, and whether we’re a fit.
I answer from what’s published on this site — nothing else. If I don’t know, I’ll say so.