The First Invoice Is the Worst One We'll Ever Process
If you evaluate an invoice capture tool, the most important moment is the first time you use it. That is when it knows the least about your suppliers, and it is the worst experience you will ever have with ours.
I say that on purpose. Every invoice we process teaches the system something. On day one we may never have seen a supplier's format before. By the hundredth invoice from that supplier, accuracy has already improved. The floor is the first use, and it only goes up from there.
How we measure it
We do not rely on impressions. We built a golden data set: more than a hundred real invoices, with thousands of individual data points labeled by humans. We score the system's extraction against that ground truth, and we run the same test against other extraction approaches, both open source and commercial.
One of the measures we track is how precisely the system locates each field on the page. After adding learned rules, that score moved from 73% to 83%. Capture gets better with use, and we can show it.
Features and agents are not the same thing
A feature is a tool that does one job when you ask. Capture reads an invoice. Auto-coding assigns the right GL account. Both are valuable.
An agent acts on information across several steps. It can look at what happened after capture ran, notice that a person edited the result, and decide that the system should learn a new rule from that correction. That decision is what makes our capture agentic, and it is why a human's judgment call makes the next invoice easier rather than just fixing this one.
One invoice, start to finish
Follow a single invoice through the system. It arrives by email. Capture reads every field, including the header, line items, and taxes. Match finds the purchase order and runs a two-way or three-way match when there is a goods receipt. Auto-coding assigns a GL account based on patterns learned from your business. Then your rules check things like amount thresholds, duplicate detection, and supplier verification. Throughout, the supplier can see where their invoice stands and what, if anything, is holding it up.
Each step feeds the next. Better capture improves matching. Better matching improves coding. All of it trains the rules that govern your workflow, so you touch fewer invoices and approve them faster.
What happens when it is wrong
No AI is right every time, so the design question is what happens next.
Every field gets a confidence score. Below a set threshold, the system sends the invoice to a person instead of pushing it forward. Matching works as a waterfall that runs from exact match to fuzzy match to agentic match, and then to manual review. The system never proceeds silently when it is unsure.
When a person does step in, the work is small. In capture, you drag the box to the right spot on the invoice and the system learns from it. In matching, even a weak match comes with a short list of likely purchase order lines, so choosing the right one is quick, and the next similar line item is matched with more confidence.
The knowledge that used to walk out the door
Before this, capture and matching lived in people's heads. A new hire might need dozens or hundreds of invoices before feeling confident. Someone tenured knew that one supplier always misspells its PO numbers and another changes SKUs constantly. When that person left, so did the knowledge.
Our goal is to turn that tribal knowledge into something your organization keeps.
What it means for finance leaders
AP teams often spend a large share of their time on data entry and matching: reading invoices, typing fields, looking up POs, and coding to GL accounts. Our target is to move 70 to 80 percent of invoices straight through with no human touch, while keeping people focused on the exceptions that deserve their judgment.
The same lifecycle view surfaces patterns a person would never catch. At one large supplier, duplicate payment disputes were taking more than 100 days to resolve. Checking each claim against the original record automatically returns an answer in minutes. That is a better experience for the buyer and the supplier.
Looking ahead, I am most excited about the layer that reasons across the whole lifecycle. A pattern in disputes might point to a fix in how purchase orders get created. That kind of insight gets problems solved before they ever reach AP.
If you want to see capture and match working on your own invoices, we would be happy to set that up.