AI that reads your purchase orders and invoices
We build systems that watch a mailbox, work out which messages are genuine business documents, read them into structured data, and check that reading with arithmetic before a person or a downstream system ever trusts it.
Not a demo. Running daily.
The system this is drawn from has been live since April 2026, handling 100 to 150 purchase orders a month for a medical equipment distributor, used daily by five staff and two owners.
- 100–150
- Orders a month
- ~1 in 7
- Need a human check
- 3
- AI agents
- 4
- Roles
- 44
- API endpoints
- 2
- Applications, one API
Reading, then checking the reading
- A worker watches the inbox and picks out the emails that are genuinely purchase orders, turning away quotations, acknowledgements and supplier mail with a one-sentence reason for each.
- The document, PDF, scan or photograph, gets read: order number, hospital or customer, dates, and every line item with its quantity, rate and tax rate.
- Arithmetic checks the reading independently of the model: does quantity times rate match each line, and do the lines sum to the stated total?
- A clean record is confirmed automatically. A record with a problem is held and routed to a person, with the reason already written out in plain English.
We can also build
Reading documents accurately is one job. We can also build assistants that answer from your own documents and records, and first-line support bots with a clean handoff to your team, on the same principle: the model's answer stays traceable to the source it came from, and nothing gets treated as true just because a model said it.
About this capability
What kinds of documents can the AI read?
PDFs, scanned pages and photographs, not just clean digital files. In production, it reads purchase orders and invoices out of a mailbox and lifts out the order number, the line items, quantities, rates and tax rates, whatever format the document arrives in.
What stops the AI from just making something up?
Arithmetic, applied independently of the model. Does quantity times rate match each line? Do the lines sum to the stated total? Anything that doesn't check out is flagged for a person with the reason already stated in plain English, rather than accepted on the model's word.
Can this read invoices too, or just purchase orders?
Both. In the system we've built, invoice reading uses the tax authority's own signed QR code first where one exists, falls back to a layout-aware reader, and only uses AI as the last resort, with duplicates matched and skipped automatically.
More on how we price, host and support everything we build: the full FAQ.
Retyping documents by hand, or chasing a shorthand system?
A 30-minute call is enough for us to tell you whether this is a fit for what you actually receive.