AI Agents Replace RPA at Danish Wholesaler, Cut 7,000 Hours
Lemvigh-Müller deployed an LLM-based system to process supplier order confirmations after years of failed RPA attempts, achieving 98% accuracy in 10 weeks.
AI succeeds where RPA failed
A 180-year-old Danish wholesaler has replaced robotic process automation with large language model-based AI to handle a critical supply chain task that RPA could never reliably complete. The project saves the company between 5,000 and 7,000 labor hours annually with a payback period under six months.
Lemvigh-Müller, Denmark's largest wholesaler of steel and technical building products, struggled for years to automate the processing of supplier order confirmations. About half the company's suppliers do not use electronic data interchange, instead sending PDF confirmations by email that must be manually checked against purchase orders for discrepancies in SKUs, quantities, pricing, or delivery terms.
Frederik Aakerlund, CIO at Lemvigh-Müller, told Forbes that RPA technology failed to handle the wide variety of PDF formats with acceptable accuracy. The breakthrough came when a procurement team member experimented with ChatGPT and discovered it could compare purchase orders to supplier confirmations in seconds.
Why it matters
This case demonstrates a fundamental shift in enterprise automation. RPA was designed for exactly this kind of repetitive, rule-based task, yet failed because of format variability. LLM-based AI handles unstructured data far more effectively, suggesting that many existing RPA deployments may be candidates for replacement. Companies that invested heavily in RPA infrastructure now face a strategic decision about when and how to migrate to AI agents.
Building a production system
Lemvigh-Müller chose to build a production solution on SAP's platform rather than continue using the free ChatGPT interface. Aakerlund cited security, scalability, and the need for a management dashboard showing processing status and exceptions requiring human review.
The company trained a custom model based on OpenAI's GPT-4.1, which proved most effective for the task. Training required 10 weeks and approximately 300 person-hours of effort. The system has operated since March with 98% accuracy, processing order confirmations from 30% of the wholesaler's supplier base.
When the AI identifies discrepancies in price, delivery time, or quantities, it drafts an email for the purchaser to send to the supplier. When confirmations match purchase orders exactly, the system updates the SAP ERP automatically.
Economics and expansion plans
The solution costs less than half a euro per document processed. With Danish labor averaging around 60,000 euros annually, the thousands of hours saved translate to substantial cost reduction with rapid payback.
Lemvigh-Müller now plans to apply similar AI approaches to other document-intensive processes including inbound PDF orders, delivery date specifications, and invoice processing.
Implications for the RPA industry
Aakerlund's experience suggests that LLM-based AI may displace RPA in scenarios involving unstructured or variable-format data. Traditional RPA vendors including UiPath, Automation Anywhere, and SS&C Blue Prism have rebranded as AI solution providers, but their technologies are not based on large language models.
In some enterprise workflows, RPA and AI agents may work together, with RPA handling stable, standardized processes while AI agents tackle variable-format documents and unstructured data. However, as LLM capabilities expand and costs decline, the use cases where traditional RPA remains the optimal choice may narrow significantly.
These details were first reported by Steve Banker at Forbes, who spoke with Aakerlund at SAP's Sapphire conference in Madrid.
This is an original analysis by the Omega editorial team. Source reporting: AI Watch.
Want systems like this working for your business?
Book a Call
