Somebody still retypes the supplier invoice. That is the job to remove.

What actually happens to an invoice today

A supplier sends an invoice — usually as a PDF attached to an email, sometimes as a scan, and increasingly as a structured e-invoice. Somebody in finance opens it, reads it, and types the supplier, the date, the lines and the totals into the ERP. Then they find the purchase order it belongs to and check that what was ordered is what was billed.

That last step is the one people underestimate. Reading the document is mechanical; matching it is not. The price on the invoice rarely equals the price on the order — there is a discount, a partial delivery, a freight line, a rounding difference — and each of those is a small judgement call that has to be made the same way every time, by whoever happens to be at their desk.

So the work does not disappear when a new person joins. It gets retaught, informally, until the person who learned it last leaves.

Why the standard tools stop halfway

OCR and template-based capture solve the first half. They turn a PDF into text, and with a template per supplier they can pull out the fields that matter. The half they do not solve is everything after: which purchase order does this belong to, does the amount agree, who approves the difference, and what happens to the ones that do not match.

The second half needs your records — the purchase order, the goods receipt, the price list, the cost centre that every line has to land in — and it needs one consistent rule for every exception. That is not a document problem any more. It is an operations problem, and it is why a capture tool that scored well in a demo can sit unused in production while the team goes back to typing.

What the agent does, step by step

The system we build treats one document as one unit of work, end to end:

• Read the document — PDF, email attachment, scan, or a structured e-invoice — with no template per supplier to maintain.
• Match it against the purchase order and the goods receipt, line by line, using your own records rather than a generic assumption.
• Post it into your ERP in the shape that ERP expects, with the cost centre and account already resolved.
• Escalate only what it could not resolve, with the reason attached, into one queue a person can clear in minutes.
• Leave an audit trail of every step: what was read, what was matched, what was decided, and by which rule.

Where it has to stop and ask

A price variance against the purchase order, a missing order, a duplicate invoice, an unfamiliar supplier — those go to a person, not through the system. This is deliberate. The value of the agent is that it removes the volume and keeps the judgement visible, so your people stay in control of the decisions that matter instead of keying the ones that do not.

Everything that reaches that queue arrives with the evidence already assembled: the document, the order, the difference and what was already tried. The question a person answers is small and specific, which is the only way an exception queue stays short instead of becoming a second inbox nobody wants to open.

AI handles the volume. Your people stay in control of the decisions that matter.

What working looks like in production

One deployment, live since 2026 and reached through an ERP integrator, processes about 320 documents a day against a contracted 33 — roughly ten times the volume it was agreed for — at about 98% field-level accuracy, with a reconciliation report nobody has to run by hand.

Since it went live, no invoice has been re-typed by hand. That is the number worth measuring, because it is the one the finance team feels: the work did not get faster, it left the building.

How the first engagement usually starts

We map how documents actually move through your operation — where they arrive, where they get stuck, what is done by hand and why — then build the smallest flow that proves the point against real volume. Stock and orders usually follow once the invoice path is trusted. The system is shaped around your process and your ERP, not the other way around, and your data stays isolated, encrypted and under your control with a full audit trail.