Danish wholesaler, three AI agents: a ten-week finance transformation

Ten weeks. That is how long it took a Danish wholesaler to go from a project manager's ChatGPT experiment to three AI agents in production inside SAP. Lemvigh-Müller, one of Denmark's largest distributors of steel, plumbing, heating and electrical products, now automates more than 100,000 supplier order confirmations a year. Around 60 percent of those confirmations arrive as unstructured PDFs, and the agents read, compare and process them without a person opening a single file. If you run finance or procurement in an SAP landscape, this is the case study to read. Not because the AI is exotic; because the delivery discipline behind it is ordinary, repeatable, and rare.

The process nobody glamorizes

Order confirmations are the unglamorous engine room of procurement. Every year, Lemvigh-Müller sends about 175,000 purchase orders to more than 2,000 suppliers. Part of the traffic moves over EDI in structured form. Around 60 percent comes back as supplier confirmations in unstructured documents: PDFs with prices, quantities and delivery dates buried in layouts that differ from supplier to supplier. A human opens the file, reads the numbers, compares them against the purchase order in SAP, and decides whether the difference matters.

That manual step is where the cost hides. It is not catastrophic on any single document; it is relentless at 100,000 documents a year, and it is exactly the kind of work that makes a procurement officer feel like a data-entry clerk instead of a negotiator.

The RPA failure that matters

Here is the part I want finance leaders to sit with. Lemvigh-Müller tried RPA and traditional automation before this. Frederik Aakerlund, its IT director, put it plainly: previous attempts did not really achieve the desired effect. The difference this time was not a better tool; it was a different decomposition of the work.

RPA tends to model the process as it exists: a robot clicks the screens, mimics the keystrokes, replicates the human workflow. That works when documents are regular. Supplier confirmations are not regular. Layouts drift, suppliers merge, terms change. Every exception breaks the happy path, and the automation either stops or, worse, processes the exception incorrectly.

The Lemvigh-Müller team inverted the problem. Instead of one broad automation trying to do everything, they built three agents, each with one narrow job. The e-mail agent receives and sorts incoming supplier e-mails, identifies the relevant confirmations and attachments, and routes them forward. The data extraction agent pulls prices, quantities and delivery dates out of the PDFs and structures them. The matching agent compares the extracted data against the purchase orders in SAP and returns a verdict: match, or deviation.

Three narrow agents beat one broad automation, and the reason is practical: each piece is small enough to be tested, trusted and replaced. When one supplier changes its PDF format, you touch the extraction agent, not the whole process.

The decision moment

Every profile has a moment where the direction changes. Here it is: a project manager named Jess Frederiksen, working in Lemvigh-Müller's Market and Procurement organization, matched an order confirmation against a purchase order using ChatGPT. Just as an experiment. He took the result to the IT director and asked whether it could become a real system. From that conversation to production took ten weeks.

What matters in that story is not ChatGPT; it is that the experiment existed, that it produced a concrete artifact, and that someone with no formal mandate carried it to the person who could say yes. In most SAP organizations, that chain breaks in one of three places: nobody runs the experiment, the result stays in a notebook, or the idea never meets the person with the budget.

Aakerlund said the short timeline was critical, because it let the solution prove its business value quickly and build internal support. That is the part most transformation programs get backwards. They spend a year on architecture and governance, then ask for patience while value accumulates. Lemvigh-Müller spent ten weeks getting something real in front of procurement officers, and let the value argue for itself.

The orchestration is the product

Klaus Heinemann, head of SAP ERP at Lemvigh-Müller, led development with the project team, and he described what changed on the ground. When confirmations were handled manually, changes could take hours or even days to flow through the organization. Now the agents update the data almost immediately, customers get a more accurate picture of deliveries far sooner, and price discrepancies are caught before the final invoice is issued. That last point is quietly the most valuable. A price discrepancy found before invoicing saves money for the wholesaler and its suppliers, and it removes a reconciliation headache from the finance team's month-end.

David Pontoppidan, head of AI at SAP for the Nordics and Baltics, made the observation that gives this case its real lesson. Many companies talk about AI agents in terms of automation, he said. The real challenge, and the real opportunity, is coordination. Multi-agent initiatives fail not because individual agents are weak, but because coordination is insufficient. Lemvigh-Müller succeeded by anchoring the solution in its SAP landscape, where the data, business rules and governance frameworks already exist.

Read that again, because it contradicts the common picture of AI projects as greenfield experiments. The advantage here was not starting from a blank page; it was starting inside a system that already knew the business rules. The agents did not have to invent governance. They inherited it.

What happened to the people

The FTE math is the part every CFO will ask about first. Lemvigh-Müller expects the solution to free resources equivalent to three to four full-time employees. But Aakerlund was explicit about the intent: the objective is not to reduce headcount, but to use expertise more effectively. The agents take the routine work, and procurement officers move to the complex, exception-driven orders where their experience genuinely matters.

This is the shift I would bet on across SAP finance and procurement over the next few years: from processor to exception-handler. The person is not removed from the process; the person is moved to the parts of the process that require judgment. Heinemann noted that procurement officers clearly recognize the value of being relieved from the most tedious routine work, and that it sparked a constructive conversation about how technology should support their day-to-day work. That is what healthy automation looks like: nobody is defending the right to open PDFs by hand.

The lessons for finance leaders

If you are a finance or procurement leader in an SAP environment, here is what I would extract from this case.

Start with a process where the exception is the value. Order confirmations work as a candidate because the routine is tedious and the exceptions matter. If the process is already simple and regular, RPA may be fine. If it is document-heavy and exception-prone, that is where agents earn their keep.

Break the work into narrow agents, not one broad one. The decomposition into e-mail handling, extraction and matching is the design pattern to copy. Each agent has one job, a clear handoff, and a defined failure mode. That is what makes the workflow testable and the system explainable, both of which matter in an audit context.

Anchor it in the system that already has the rules. The agents live in the SAP landscape, next to the purchase orders they compare against. No parallel data lake, no shadow database, no governance model invented on the side. The existing controls become the agent's controls.

Plan for master data. Heinemann said the project made the importance of master data quality increasingly clear, including areas like Incoterms. Treat that as a finding, not a failure. AI has a way of exposing data you thought was clean. Budget for the cleanup, because it is part of the ROI, not an obstacle to it.

Get to production before you try to be perfect. Ten weeks is not a normal timeline for SAP projects, and Pontoppidan's remark is worth quoting: ten weeks from idea to production is far from the norm, but perhaps it should be. The company measured its return in quarters, not years, and that was only possible because the solution shipped fast enough to start paying back.

The honest caveats

None of this means AI agents are trivial to deploy, or that every SAP shop can copy the timeline. Lemvigh-Müller is a Danish company with short decision paths and a pragmatic approach to technology, and it worked with NTT DATA Business Solutions to make the solution production-ready and fully integrated. Kristian Dahl, SAP UX manager at NTT DATA, credited the modular, agent-based architecture for the fast transition from pilot to production, and for a solution that can be extended as new requirements appear.

The company is also clear that this is a first step. The agents run on selected supplier inboxes and business areas for now. Aakerlund said the experience gave them the confidence to consider the same approach across invoice processing and order management. The same orchestration pattern, applied to the next process, with the master data lessons already paid for.

The image to keep

Here is the picture I will keep from this story. A procurement officer opens SAP in the morning, and the day is not a pile of PDFs to read. It is a short list of exceptions the agents could not resolve, each one a real commercial decision. A price discrepancy that used to surface weeks later, at invoice time, is now flagged before the invoice exists. That is what working AI looks like in a finance operation: the system handles the routine, and the human handles the judgment.

The Danish wholesaler did not wait for permission to be pragmatic. Ten weeks, three agents, more than 100,000 documents a year. The blueprint is public, the architecture is modest, and the discipline is available to anyone who stops trying to automate the process as it is, and starts decomposing it into jobs that agents can do well.

Sources

• Lemvigh-Müller: AI Automates Order Confirmations