Several SAP customers with AI success stories spoke at SAP’s Sapphire conference in Madrid. These case stories involved AI based on Large Language models. My favorite supply chain case story involved AI replacing Robotic Process Automation.

RPA is a software technology that uses virtual workers or "bots" to automate repetitive, rule-based digital tasks. By mimicking human interaction with graphical user interfaces (clicks, keystrokes, and data entry), RPA automates back-office work.

I had a chance to speak with Frederik Aakerlund, the CIO at Lemvigh-Müller , about this 180-year-old Danish wholesaler in the building products industry. The company is the largest wholesaler of steel and technical installations in Denmark.

Lemvigh-Müller has largely standardized on SAP applications. Their first AI case was focused on EDI or orders placed on their website. “In a perfect world, our suppliers would digitally transform their data,” Mr. Aakerlund explained, which would make purchasing more accurate and efficient.

PO/order confirmation is a key supply chain process. When done correctly, it contributes to the perfect order – the right goods delivered on time and in full.

“But the perfect world doesn't exist. About 50% of our suppliers don't use EDI. So, we send a purchase order to these suppliers, and they send back a PDF order confirmation. Then we open it in an email.” There can be deviations from the SKUs they order, the quantities ordered, the cost of those items, or the delivery and payment terms.

“We have tried to automate that via Robotic Process Automation for many years. It didn't work,” Lemvigh-Müller’s CIO explained. The wide variety of PDF formats made it impossible for their RPA solution to do the digital transformation with a high degree of accuracy.

“But then one of the guys in our procurement department said, ‘Frederik, I tried to use ChatGPT to compare our purchase order to the order confirmation from the supplier. In a few seconds, it told me the differences. Can we do this in SAP?’”

Now AI is reading these different PDFs attached to emails, finding the supplier, locating their purchase order number, and comparing the response to their data. If everything is fine, the business process is updated in their SAP ERP. If there is a deviation in price, delivery time, or the number of pieces to be delivered, the AI is drafting an email that the purchaser can send to the supplier.

I asked Mr. Aakerlund why they chose to buy a solution from SAP if ChatGPT, a free application, was working. First of all, if they were doing these transformations at scale, they might need to buy a special license. Further, “it's safer,” the CIO said. “And then we have also built it on top of this dashboard, so that they can see how many order confirmations have arrived, how many are being processed, how many are green and going through, and how many need manual intervention.”

Lemvigh-Müller built a custom model based on OpenAI’s 4.1 language. “That turned out to be the best one” for this problem. The model took 10 weeks to train and required a total of 300 man-hours of effort. “That’s not much,” said Mr. Aakerlund.

The wholesaler has been using this since March. The model is 98% accurate. The solution is being used to correct the order confirmations from 30% of their suppliers.

The payback has been “tremendous,” saving between 5,000 and 7,000 man-hours per year. In Denmark, they pay workers on average about 60,000 Euros per year. The cost of the AI solution is less than half a euro per document. Mr. Aakerlund agreed with my guess that the payback period was under 6 months.

This project has opened Lemvigh-Müller’s eyes to the possibility of using AI based on LLMs in other areas, such as inbound PDF orders, PDFs specifying delivery dates, and PDF invoices.

Robotic Process Automation is supposed to excel in stable, standardized, and well-documented processes like PO matching. However, because of the large variety of PDF formats, it failed this Danish wholesaler.

RPA can be used in conjunction with AI for labor intensive back-office work. In some cases, where RPA is reliable and cost-effective, it can serve as an agent, delivering updated information. Where RPA cannot perform reliably, an AI agent can be used instead.

EDI companies like UiPath , Automation Anywhere , and SS&C Blue Prism have rebranded themselves as AI solution providers. Depending upon how you define AI, that might be true. But they are clearly not AI based on Large Language Models. These firms may struggle to maintain sales as LLM AI eats into their use cases.