Many companies still struggle to trust AI agents. After all, large language models operate probabilistically and can produce false information. This is particularly problematic in financial processes. How do you build trust?

Kask: Trust begins with a stable foundation. ERP systems remain the reliable system of record. They operate deterministically, contain the business logic, and hold the relevant company data. AI agents build upon this foundation. They do not replace it.

Equally important is the human-in-the-loop principle. Employees must be able to understand what the agent is doing, verify its results, and intervene if necessary. That’s why employee training also plays a crucial role. They must understand how generative AI works and where its limitations lie.

Of course, language models can hallucinate. At the same time, we must not forget that humans are not infallible either. The key lies in the collaboration between humans and AI. This allows us to improve both the efficiency and the quality of many business processes.

Another important component is transparency. Our global AI ethics policy, for example, stipulates that users must always be able to recognise when AI is involved. In Joule, it’s possible to trace which data sources the agent used and which steps it went through in reaching its decision. This traceability is an essential prerequisite for trust.

What distinguishes an SAP agent from a general AI agent that merely accesses an ERP system?

Kask: The key difference is that Joule and the SAP agents are directly embedded in the ERP system. There, for example, we’ve built a knowledge graph that describes the semantic relationships between all tables, business objects, and data fields.

To put this into perspective: The SAP S/4HANA Knowledge Graph is based on approximately 452,000 ABAP tables, 7.3 million data fields, and thousands of analytical views. The semantic relationships between these artifacts are modelled in the Knowledge Graph and made available for AI applications.

For example, if a user wants to view all open purchase orders, the agent does not first have to laboriously search for the relevant information. It immediately knows which tables and objects are relevant and also understands the relationships between a purchase order, a purchase requisition, the responsible approvers, and other business objects. As a result, the agent not only works much more precisely but also requires significantly fewer tokens because it can greatly narrow down the search space.

If, instead, one attempts to simply overlay AI onto an existing system or extract data from a relational ERP system, many of these relationships are lost. In a sense, this destroys the semantic context that is crucial for precise answers.

That is why we view the ERP system as an enormous strategic advantage. It has been the system of record for decades and contains roughly 50 years of codified business and process knowledge. This knowledge forms the foundation for what we call the autonomous enterprise. The agents build upon this knowledge and continue to develop it.

In the future, SAP agents will also communicate bidirectionally with agents from other providers via standards such as Agent-to-Agent (A2A).

According to your study, AI currently creates the greatest added value in decision-making, customer interaction, and gaining new insights, rather than in traditional productivity gains. Will this change the way companies justify AI investments in the future?

Kask: In our study, productivity was simply rated slightly lower than, for example, gaining new insights. In the long term, however, productivity remains the ultimate goal. Europe, in particular, has been suffering from comparatively weak productivity growth for years.

At SAP, we therefore first evaluate every new AI feature based on its specific business value. For all agents and AI features that we include in our AI Feature Catalog, we first conduct a value analysis. We ask: What benefit does the feature offer the user? Does it contribute to higher revenue? Does it increase productivity? Only then is it developed further.

At the moment, the greatest added value often still lies in consolidating information from structured and unstructured data sources and making it accessible via natural language. The next step, however, is to translate these insights directly into more efficient business processes. That is precisely where the greatest productivity gains will be realised in the future.

If you could give CIOs just one or two pieces of advice for the transition from generative AI to AI agents, what would they be?

Kask: In my view, the biggest mistake would be to try to transform the entire company all at once or to attempt to perfectly prepare all the data right from the start.

Instead, you should consider what kind of agent can create significant added value, and then implement it. Of course, this agent needs access to consistent and context-rich enterprise data. That’s exactly what we’re working on at SAP with technologies like the knowledge graph, which maps the semantic relationships within enterprise data.

In addition, with data products and the SAP Business Data Cloud, we provide tools that make data from various sources usable for AI agents. Thanks to zero-copy and data fabric approaches, information from legacy systems, Snowflake, or ERP systems can be consolidated without first having to extensively replicate the data. For a procurement agent, this makes it possible to provide exactly the relevant data for the specific use case.

The key point is this: Companies do not have to wait until they have fully migrated to the cloud or consolidated their entire data landscape. With the technologies available today, data can already be made usable for specific AI agents, managed in a controlled manner, and used to quickly generate initial business value. On the other hand, those who wait for the perfect starting point run the risk of falling behind.