SAP puts AI agents to work on invoices, forecasts and maintenance
A senior executive at the German software giant says AI agents are already replacing manual back-office tasks at scale
Enterprise artificial intelligence (AI) is generating a new division of labor in the back office. Accountants no longer check contractor invoices line by line. Business analysts no longer manually pull data from disparate systems to build dashboards. AI agents are doing it instead.
The shift is being driven not by better models, but by better data architecture. When an agent knows exactly where to find the data it needs, results move from impressive demos to measurable business outcomes.
"All the accountants who did that in the past manually can just focus on what the cost and benefit of the project is, instead of working on checking the various invoices," said Andreas Krause, Vice President of Customer Advisory at SAP EMEA Data and AI. "The agent assigns project costs exactly to the different metrics, and measures what was agreed from a budget perspective when the project was agreed with the contractor."
Krause gave two examples of AI agents already operating in enterprise environments.
The first is a customized agent built for a utility company that automates contractor invoice validation for grid maintenance projects. The agent checks whether each invoice meets the required detail levels, compares actual costs against agreed budgets, flags overruns and incomplete entries, and automatically sends an email to the contractor requesting missing information.
The second is an embedded analytical agent that generates complete data visualizations from a natural language prompt. In an HR use case, the agent analyzed staff terminations by pulling salary grades, headcount and demographic data. It displayed its reasoning step by step, showing which data sources it was connecting and which key performance indicators (KPIs) it would use, before producing a finished dashboard.
SAP provides an agent builder that allows customers to define agent tasks and connect whichever large language model (LLM) they prefer. The company has partnerships with all major LLM vendors, leaving the model choice entirely to the customer.
Data gaps, dead pilots
Krause was speaking at the AI and Big Data Expo, part of the TechEx event series, held in London. His session, titled “Winning the AI Race: Turning Enterprise Data into a Strategic Asset,” examined why so many AI initiatives stall before reaching production.
He said a study found that poor data readiness is the leading cause of AI pilot failures, accounting for roughly 30% of cases in which projects never reach production.
He illustrated the problem with a common business query requiring five years of historical financial data, forward-looking forecasts and addressable market figures from external systems. An agent that cannot locate and connect all three data types will either return incomplete answers or produce a hallucination.
“A lot of AI prototypes never leave the stage of prototype into production. One of the main reasons is poor data readiness. Do we really have access to all the data we need, and is the data at the necessary quality so that it gives back real results?” he said. “When the agent doesn’t exactly know where to access the data, the result will simply not be what you expected, and in the worst case it brings back a hallucination.”
SAP’s own AI assistant, Joule, faces the same constraint. Without clear data access, even an enterprise-grade tool will fail.
SAP’s solution is a three-layer architecture: SAP and non-SAP applications at the base, Business Data Cloud in the middle and AI agents at the top.
“A customer needs to have a unique definition within all SAP applications, but also the non-SAP applications. Business Data Cloud harmonizes this together, so that there is one unique definition of a customer, a cost center or material information,” Krause said. “We have an open data partnership with various vendors. We started with Snowflake, in addition to Databricks, but we opened it up to Microsoft Fabric, Google BigQuery, and more to come during this year.”
Business Data Cloud incorporates SAP Datasphere for semantic modeling and KPI calculations, SAP Analytics Cloud for front-end visualization, and SAP Databricks for AI and machine learning workloads involving non-SAP data. Customers using SAP’s legacy Business Warehouse system can also connect it to the platform.
Krause demonstrated two intelligent applications that combine SAP and external data. A working capital application pulls enterprise resource planning (ERP) financial data alongside external credit risk ratings. Finance teams can model scenarios, estimating the cash flow impact if customers pay within 30 or 60 days, and receive best-case, realistic and worst-case projections.
A second application, built by a manufacturing customer, compares preventive versus corrective maintenance costs. It draws on SAP material and cost data, supplier pricing and sensor data from machinery to identify the optimal maintenance cycle and estimate potential savings from each approach.
Security layer built in
Data security emerged as a significant concern in the Q&A, with participants noting that companies are racing ahead with AI while cybersecurity readiness lags, a pattern linked to the incoming EU Cyber Resilience Act. Krause acknowledged that after poor data readiness, security is the second most common reason AI projects fail to reach production.
For embedded AI capabilities, security is inherited from SAP’s existing application-level controls. For customized AI built on Business Data Cloud, a separate framework applies.
“Business Data Cloud also has the security framework behind there, so that you know who is going to access the information and what type of data they are able to see. We have a knowledge graph where you can apply security down to the data level,” Krause said.
The knowledge graph also governs what each LLM is permitted to query, ensuring models cannot access data beyond their defined scope.
An audience member raised the question of whether banks in countries with strict data regulations should build AI exclusively using locally deployed LLMs rather than cloud-based models.
Krause said SAP is progressively building sovereign cloud environments for countries with data residency requirements, designed for public sector entities and regulated industries. These ensure data does not cross national borders. Some countries are also developing their own LLMs for exclusive domestic use, and SAP’s open LLM framework allows these to be connected to its platform.
On implementation efficiency, pre-built data products cover cost centers, materials, customers and other standard business objects. They include validity periods, currency information, and full descriptive metadata, eliminating the manual data modeling that previously required reconciling multiple SAP ERP tables. The reduction in implementation effort can reach up to 81%, growing in proportion to the complexity of the data environment.
SAP plans to expand its library of intelligent applications into industry-specific use cases, with consumer products and retail already in development. Further verticals and additional open data partnerships are expected before the end of 2026.