Building AI agents for ERP process optimization: One ISV's experiences addressing new supply chain challenges

Companies of all sizes are experimenting with AI agents, and Microsoft partners there are unique incentives to accelerate the introduction agentic experiences. MSDW spoke with Daniel Wessel, Head of Products at Yaveon, about his company's exploration of agents in industry-specific scenarios. 

Yaveon is a German software firm founded in 2008. Over its history, it has served industries like process manufacturing, pharmaceuticals, chemicals, food, and medical technology, including a growing portfolio of apps.

Wessel told MSDW that Yaveon hears from customers that one of the key opportunities of AI is in its ability to better understand complex business information from an ERP.

I think in general, we see two different asks from our customers. On the one hand, it’s everything about information and intelligence. Can we make the complex data in the ERP accessible more easily through AI? Sometimes we see users struggling with the complexity of the product and how to connect. One example is if one of our vendors isn’t able to deliver a product in time, how does it impact our production.

Yaveon customers are looking to AI to help them find information hidden in their ERP systems or to assist with item setup, procurement, and assigning items to sales orders. Wessel says customers express a desire for new insights, like how to optimize recipes or predict payment outages. 

What we already delivered is exactly this. We call it a Supply Chain Agent that sits in Copilot Studio and allows user to ask all the different questions around delay of the whole supply chain, starting from purchase through manufacturing to sales. Maybe [there is] like the delay of an internal production order or missed shipment date: how does it impact the supply chain? Can we deliver substitute items to still meet our shipping date?

Yaveon has also created a native Business Central agent to create incidents based on incoming emails, allowing users to capture anything that impacts product quality. The agent scans an inbox based on incoming emails, finds out if there is a sales order, matches the shipment and what is impacted, finds out the lots from the items that were sent out, then proposes corrective action.

So far, AI agent design and development is a bit more exploratory and involves more prototyping than traditional development. Yaveon starts with a customer requirement, then decides what platform would best deliver on that requirement. In some cases, the end result might involve a program working within Business Central itself, or it might involve a program outside of Business Central interacting with it. Because there is a lot of prompt engineering, there is significant A-B testing involved. 

The Yaveon team runs through questions like whether are there deviations in the results, changes to prompts, or adjustment needed. Typically, they will keep prototyping as part of an iterative approach. 

When it comes to development, Yaveon turns to Copilot and Microsoft’s own built-in tooling.

One of the best tools to use is Copilot itself to draft the first prompt and iterate on it, which works quite well, especially for Copilot Studio agents. You can keep track of all versions using the Visual Studio Code extension for Copilot Studio. From there, the agents can be synced and published, so you don't have to manually copy prompt variations into the user interface.

Wessel explained that this agent code is the main driver for working with Yaveon’s codebase. The company maintains its Copilot agents within Copilot Studio, giving its customers the option to connect with Business Central using the MCP server provided out of the box, via APIs, or with BC’s internal agent model.