Job DescriptionTeach AI how business actually worksAI agents are starting to take on real business work: releasing blocked invoices, re-planning supply, preparing the financial close. To do it safely, they need more than access to data. They need to know what an order, a supplier and a cost center are, how they connect, which rules and authorizations apply, and what happened the last time someone made a similar decision. When executing mission-critical business processes, an answer or action that is right 90% of the time is not good enough.
SAP systems hold the context for many of the world's largest companies. SAP's knowledge graph for SAP S/4HANA alone is based on 452,000 tables, 80,000 CDS views (semantic data models) and 7.3 million fields. Yet much of the meaning, especially in each company's own extensions, configuration and decision history, is still implicit. A 'customer' in sales, a 'business partner' in finance and an 'account' in CRM can be the same company described three ways, and an agent deciding whether to raise a credit limit has to know that. This is not a data-cleaning problem, it is a knowledge-representation problem. Today that context, the semantic layer agents rely on, is mostly modeled by hand, one company at a time. We think it can be learned, so AI agents can act on business data, not just read it. SAP's Autonomous Enterprise depends on that, and it is what makes the agents people work with in Joule Work, SAP's AI workspace, accurate enough to trust.
As Principal Data Scientist, you will define and lead our research toward that goal and work with product teams to ship the results. This is a research and modeling role, not product analytics.
SAP Research & Innovation North AmericaWe are a team of scientists and engineers at SAP Labs in Palo Alto, minutes from Stanford. We look a few years ahead and scout emerging technologies, take the early, riskier bets that product teams can't always take, and prove or drop them fast. SAP's product teams build and run SAP Knowledge Graph, we look beyond and find and prove methods that demonstrate improvement or paradigm shifts. Our job is to spot the opportunities that will matter most to SAP and our customers, and to get there early. We prototype with dedicated GPU compute on realistic business data and, in co-innovation projects, on real data for training and evaluating models. We publish research, work with startups, universities and institutes including Stanford HAI and UC Irvine, and hand what works to SAP's product teams. We move like a small lab, with the reach of a company whose software runs finance, supply chain and HR for many of the world's largest companies.
What you will work on- A research agenda with a point of view. Pick the two or three bets where SAP can lead in AI on business data, agree up front how we will know each one is working, and stop the ones that aren't. Closely collaborate with top universities and institutes such as Stanford HAI.
- What comes next. Track emerging research, such as relational and graph foundation models, neuro-symbolic agents and context graphs, test it on SAP data before the market settles, and make the case for the next bet.
- Business semantics that can be learned. New methods to learn, align and evolve ontologies and knowledge graphs of business objects, rules and actions from SAP metadata, master data, documents and process data, combining LLM-assisted ontology learning, schema matching and entity resolution with targeted expert review.
- Agents grounded in business semantics. Graph-grounded retrieval; semantic parsing into SQL, SPARQL and OData calls, including finding the right API, across very large schemas; and governed actions: checks that an agent's plan respects business rules and the user's authorizations before it runs.
- Models that combine structure and statistics. Graph learning and relational foundation models (such as SAP-RPT-1.6), using knowledge-graph structure to predict, explain and simulate business outcomes.
- The yardstick. Benchmarks and simulation environments for agents on ERP data, where today's public benchmarks fall short, so SAP knows when AI is accurate and reliable enough to act on.
- Results that land. Work directly with data teams and domain experts in co-innovation projects, turning messy enterprise data problems into tractable ontology problems, and work with the teams behind SAP Knowledge Graph, SAP Business Data Cloud and Joule, SAP's AI assistant, to ship what works.
- A stronger research community. Publish at top venues, release code and benchmarks where we can, co-lead university collaborations, and mentor scientists and PhD interns.
In your first year, success could look like:- Two or three research directions you defined, agreed with leadership, at least one with a product partner and one longer-horizon
- A result running in an SAP product or pilot
- Papers submitted, and a benchmark or open-source release others use
- Scientists doing better work because of you
This is a principal-level individual-contributor role, comparable to Staff or Principal at other Bay Area technology companies. You will drive research and set standards other teams adopt, mentor scientists who don't report to you and represent SAP's research externally, while still writing code and running experiments yourself.
QualificationsRequired- Typically 8+ years in applied research, data science or machine learning, or a PhD and 5+ years, including research you led that changed a product or platform. An MS or PhD in computer science, machine learning, statistics, mathematics or a related field, or equivalent practical experience.
- Deep expertise in one of these areas and hands-on experience in another, with at least one of the two in knowledge graphs, ontologies or semantic data integration:
- Knowledge graphs, ontologies and semantic data integration: RDF, OWL, SHACL, SPARQL or property graphs; ontology and schema matching; entity resolution; semantic layers; you have made the hard modeling calls yourself (entity boundaries, inheritance, relationship semantics), not just queried existing graphs
- Machine learning on relational, graph or tabular data: GNNs, relational deep learning, relational or tabular foundation models
- LLM systems over structured data: retrieval, tool and API discovery, text-to-SQL and schema linking, tool-using agents, fine-tuning or post-training
- Evaluation and benchmark design for AI systems
- Strong foundations in machine learning and statistics; fluent in Python and SQL, with PyTorch or JAX; at ease with schemas spanning thousands of related tables.
- Strong problem solving skills, shown by concrete examples: you have turned an ambiguous business question into a crisp, testable research problem, started with the simplest approach that could work, and stopped or redirected work, including your own, when the evidence said so.
- Strong communication skills, shown by concrete examples: you can explain a trade-off to an executive concisely and clearly. A paper, design doc or memo of yours led a team outside your own to adopt, fund or drop something. You are able to work through a design with skeptical engineers.
Preferred- Publications at venues such as NeurIPS, ICML, ICLR, KDD, ACL/EMNLP, SIGMOD/VLDB, ISWC, WWW, or widely used open-source software, datasets or semantic standards work.
- LLM-assisted ontology learning, neuro-symbolic methods, agent memory, context graphs and decision traces, or reinforcement learning for tool-using agents.
- Process mining (including object-centric event data), causal inference, forecasting, or privacy-preserving and synthetic data.
- Familiarity with how enterprises run (finance, supply chain, procurement, HR), and experience mentoring scientists or leading university collaborations.
- SAP experience is a plus, not a requirement.
If you recognize yourself here but do not match every line, apply.
