Read any bill format
PDF, scanned image, emailed body text, EDI, or a vendor portal download. Line items, tax, terms, remit-to, and PO reference come out structured.
AI agents · finance
It reads the bill, identifies the vendor, codes the expense, matches the PO, checks policy, routes approval, and posts a balanced entry. Your team stops doing data entry and starts handling only what genuinely needs a person.
What it does
Each of these is a step your AP clerk does today by hand. The agent does them in sequence and stops at the first thing it is not confident about.
PDF, scanned image, emailed body text, EDI, or a vendor portal download. Line items, tax, terms, remit-to, and PO reference come out structured.
Matches against your existing vendor master including aliases, DBA names, and the twelve ways a vendor writes its own name. Flags likely duplicates before they are created.
GL account, department, location, class, and project — learned from how your team has coded this vendor before, not from a generic model.
Two-way and three-way matching with configurable tolerance. Quantity and price variances are itemised rather than silently accepted.
Same invoice number, near-identical amount and date, or a re-sent statement. Caught before it enters the approval queue, not after payment.
Follows your approval matrix by amount, department, and vendor. Chases the approver. Escalates when it stalls.
Before & after
A $4,180 hosting invoice against an open purchase order — the most ordinary bill your team will process today.
The left column is a timed observation from onboarding three professional-services clients in 2026, not a survey figure. The right column assumes the agent is at Level 2 for this vendor with a matched PO. A bill with no PO, a new vendor, or a variance outside tolerance takes a different path — it goes to a person, which is the design.
Accuracy
Straight-through rate is the share of bills that reach a posted entry with zero human edit. It is measured per customer, per week, against that customer’s own corrections.
The agent does not arrive knowing your chart of accounts, your vendors, or the conventions your controller has been applying for six years. Week one is deliberately conservative: it proposes, a person confirms, and every correction becomes a labelled example. By week twelve it is coding the recurring eighty percent of your vendor base without intervention.
A rate that starts at ninety percent is a warning sign, not a feature. It means the model is guessing confidently on a chart of accounts it has never seen. We would rather be visibly cautious for a fortnight and earn the number.
Prompt and model changes run against a held-out set of your historical bills before they ship. A change that lowers straight-through rate, or raises the rate of confidently-wrong coding, does not deploy. That gate is the reason the number can be trusted, and it is why we are willing to publish it.
Authority
The agent prepares the bill in full and a person commits it. This is where every new customer starts, for every vendor.
Enabled per vendor, per threshold, by your controller — typically once a vendor has thirty clean drafts behind it.
Moving money is never delegated to an agent at any confidence. It requires a named human approver, without exception.
Questions
Send fifty real bills. We will show you what the agent would have done next to what your team actually did — no connection required.