Beyond PoC: How sovanta Detects Purchasing Anomalies with a Custom Joule Agent built in Joule Studio

[DanielJckl]

Agents are the future of intelligent enterprise automation. They can understand context, enhance workflows, and interact autonomously within business systems. With the new agent builder capability in Joule Studio, SAP is providing the tools to get closer to that future.

As an early adopter of Joule Studio, we at sovanta explored how it can be used to solve real-world problems with custom AI agents. For more than 15 years, sovanta has been supporting SAP customers with Extensions, Integrations, and AI solutions in the cloud, forming into a specialized SAP BTP partner.  Using this experience in SAP BTP and AI, we have successfully implemented a real-world use case with the agent builder that goes beyond a proof-of-concept.

In this blog post, we will explore this use case that we have adopted together with one of our most innovative customers Endress+Hauser: Detecting anomalies in customers’ purchasing behavior.

Why we chose Agent Builder in Joule Studio

Before exploring the agent builder, we already gained hands-on experience creating Joule Skills in Joule Studio. This gave us an early glimpse into the potential of integrating AI more deeply into SAP business processes, and we were excited to take it to the next level.

When SAP introduced the agent builder, we immediately saw its potential. It provides a low-code environment that allows users to design, configure, and deploy AI agents with minimal technical effort. By combining skills, data connections, and prompt logic, it enables teams to create powerful agents that can understand business context directly within the SAP ecosystem and scale in a heterogeneous landscape efficiently.

For us, it was the perfect opportunity to explore how AI agents could become a natural part of everyday business processes. Our goal was clear: to validate whether the agent builder could support real business-critical scenarios beyond just prototypes and evaluate its potential for large-scale customer adoption. 

The Business Challenge: Identifying Anomalies in Purchasing Behavior

Together with our long-time customer Endress+Hauser, a global process and laboratory instrumentation and automation supplier, we decided to focus on a challenge faced by sales representatives: identifying anomalies in customers’ purchasing behavior to be better prepared for sales meetings.

Currently, it’s almost impossible for sales representatives to quickly recognize if a customer’s purchasing patterns have changed significantly across Endress+Hauser’s extensive product portfolio. Such anomalies could include:

• Sudden increases or decreases in a specific product area
• Shifts towards unexpected measuring principles
• Recent orders of products that are soon to be discontinued or put on order stop

Detecting these changes manually would require searching through multiple systems, analyzing past purchases, checking order stop lists – a process that is extremely time-consuming and not really feasible in practice. As a result, these insights could be overlooked, leading to slower reactions in the sales process, missed opportunities, and lower customer satisfaction. This made it a perfect challenge to be solved by an agent and the ideal case for the agent builder.

Our Solution: Detecting Anomalies and More

Our idea was simple: let sales representatives interact with their data conversationally. To identify anomalies, they should be able to simply ask:

• “Are there any anomalies in purchasing behavior for my customer?”
• “Did my customer purchase any products recently that are running in order stop over the next year?”.

The agent would extract the needed data, detect anomalies, and provide a clear explanation.

We did not want to be limited to anomaly detection, which is why we decided to connect the agent directly to the database. This would allow users to go beyond anomaly detection and explore additional insights about customers and products. For example, instead of just identifying products running in order stop, the agent would also be able to identify successor products that should be recommended instead.

And it doesn’t stop there: by additionally connecting the agent to the product documentation, we were able to further extend its capabilities. This allows sales representatives to instantly compare the ordered product with its successor and understand what has changed.

How our Agent can detect Anomalies

To be able to answer questions to prepare for sales meetings, we identified several data sources we needed to connect:

• BDIH – Business Data and Integration Hub: SAP HANA Cloud containing all types of operational data, including technical details of millions of sold devices.
• Product Knowledge Base: Contains Product Documentation that is reachable via an internal RAG System.

The exact technical setup can be seen in the architecture below:

Since we wanted the agent to handle a wide range of questions from the BDIH, we decided to implement a Text-to-SQL approach with Joule Studio. This enabled the agent to transform natural language questions into structured queries, allowing it to directly extract data from the database.

To achieve this, we created a Text-to-SQL pipeline in Joule Studio. This involved grounding the agent with the following documents:

• The database schema of all relevant tables
• Example SQL queries
• General documentation about the database structure and business context

These documents were added via document grounding, allowing the agent to better understand how to map natural language questions to actual database queries.

For the actual data extraction from the BDIH, we included a custom Joule Skill that takes a SQL query as input and retrieves the corresponding data through a proxy endpoint to the database. Because SAP Build Actions currently do not fully support dynamic outputs, we applied a workaround to handle the dynamically changing query results coming from the database.

In addition to the Text-to-SQL pipeline, we also connected the Product Knowledge Base with a separate Joule Skill. With these two integrated data sources connected to the agent through separate Joule Skills, the agent can flexibly retrieve and combine relevant information depending on the type of question.

The Agent Flow and its Business Impact

The interaction flow is simple and powerful:

1. The sales representative asks a question in natural language.
2. The agent interprets the intent and decides which data source to use.
3. It will extract information from the DB schema and example queries before formulating an appropriate SQL query.
4. The query result is retrieved and enriched with additional knowledge from the Product Knowledge Base.
5. The agent summarizes the findings in plain language and is available for follow-up questions.

The following two examples demonstrate how this flow can help find anomalies in purchasing behavior for certain measuring principles, find products in order stop and their successor products:

This conversational and data-driven approach significantly improves efficiency and empowers sales teams to prepare more effectively for customer meetings. It also allows sales representatives to identify opportunities and gaps earlier and provide recommendations for the customer, demonstrating how Joule Studio can create measurable business value.

Our Final Thoughts on Agent Builder in Joule Studio

Our experience as an early adopter has shown that the agent builder is more than just a proof-of-concept. It’s a promising addition to the SAP ecosystem, fully native to SAP Build, which allows us to leverage existing components such as actions, automations, and workflows directly within Joule Studio. It also proved to be highly open and interoperable by seamlessly connecting agents and skills with third-party systems. Combined with its built-in business understanding and grounding, it enables enterprise-ready AI Agents to reliably interact with real business data.

Looking ahead, we plan to further explore capabilities such as Agent-to-Agent collaboration and additional agent tools coming to the agent builder. For example, a native database tool could eliminate the need for our current Text-to-SQL pipeline and even further simplify the agent development process. At sovanta, we are excited to continue exploring how Joule Studio can empower our customers with AI Agents that can seamlessly and autonomously interact within their business systems.