Take three-way matching. Finance teams check whether an invoice matches the purchase order and the actual delivery before it moves to payment. That means recognizing the invoice, checking whether it has already been processed, finding the purchase order number and validating the supplier details. Individual order lines, quantities, prices and deliveries need to match as well.

And then you get an invoice without a PO number. Or a supplier uses a slightly different name. Or the delivery was split over two dates. Nothing unusual for someone in finance. They know where to look and what to check. However, this takes a lot of time, and AI can automatically make small decisions and bring larger decisions to the finance team. The repetitive checks and exceptions are what make three-way matching an interesting use case for agentic AI.

Adding a system of action to finance

ERP and finance systems remain the trusted source for financial administration. We use AI to handle part of the work that happens around and between these systems, without changing that foundation.

We call this the “system of action”. It interprets incoming information, retrieves data from different systems and determines what should happen next based on the rules of the process. When an exception needs human judgement, the workflow can bring it to the right person, including the information needed to make the call.

For this use case, we work with Lleverage, a low-code platform for building agentic workflows and connecting them to existing applications. The platform supports multiple LLM providers, so the right model can be selected based on what the specific task requires, instead of tying the complete workflow to one AI supplier.

Of course, having the platform doesn't mean you suddenly have a working finance process. You still need to teach the workflow what a good invoice looks like, which checks matter and what to do when reality doesn't quite follow the happy flow.

Getting into the details of the process

Before we build, we map how invoices move through the organization today. Where do people spend time? Which decisions already follow clear rules? What information lives in which system? We also sit down with the people doing the work, because quite a few process rules turn out not to be written down anywhere.

Invoice extraction shows why those details matter. An LLM can turn an invoice PDF into structured data, but the next system needs more certainty than an answer that simply looks right. Dates need the expected format, prices need to be interpreted correctly, and a purchase order number has to be an actual purchase order number. Sounds obvious, until you start processing invoices from dozens of suppliers with their own formats.

Exceptions deserve just as much attention. Finance teams deal with them every day, often without thinking much about the decision they are making. But once you want a workflow to make that decision, you have to get specific. What does a valid PO number look like? When is a price difference acceptable? What should happen if there is no PO at all? We translate those decisions into prompts, validation steps and routes through the workflow.

For the first version, we ran invoices from the first half of 2026 through the workflow and compared the results with how finance had handled those same invoices. Comparing the results of processing real invoices helps spot exceptions that you didn't think of beforehand.

There are also less visible parts of the work. Invoices contain business and supplier data, so access, security and privacy need to be sorted out. Depending on the systems, models and data involved, supplier due diligence and business impact assessments may be part of the implementation too. Not the part you put in an AI demo, perhaps, but definitely part of getting one into production.

This is where our Itilians get under the hood, together with the customer team. We need the people who know why an invoice sometimes takes a different route, as well as the IT knowledge of the systems behind it. We use that input to build the AI components into a working flow and connect it to the existing financial administration.

This ensures an agentic workflow that actually can take care of more of the repetitive checking, the manual actions, and speed up the process. So that the finance team can spend that time-saved on those invoices and decisions that actually need their attention.