SAP says enterprise agents need business context, not just better data

Context is more than a knowledge graph

A company task sits inside a process: a payment, an order, a shipment or a new hire. That process tells an agent where the task stands, what came before and what should happen next. But some decisions depend on people’s judgment, which may not be recorded in corporate systems. As Oren puts it, an agent needs to recognize when a situation does not fit the formal steps, then explain the exception in its response or action.

The context can also widen as the task unfolds. A question about a customer and their marketing campaigns may lead to discounts, delivery timing and then payments—spanning supply chain management and finance.

A knowledge graph can connect company records, but connections alone do not settle what those records mean. “Invoice,” for instance, might mean a customer payment to one team, an expense reimbursement to another, and a supplier bill to a third. An ontology gives the term a shared meaning across systems. For a global company with hundreds of thousands of employees, Oren says, records must be linked and “speak the same language.”

SAP Business AI Platform (BAIP) is designed to bring data and business context together in a governed environment. Its knowledge graph maps relationships among customers, suppliers, assets, products, organizational structures and business processes, with context that can be reused across agents.

That context also includes operating rules: service-level agreements, payment terms, tax and security requirements, regional restrictions, and data-anonymization rules. A shared definition of the data is only part of what an agent needs to act appropriately.

The data itself is a different problem

Much enterprise data is relational and transactional: it lives in tables, not natural language. General-purpose language models were not built specifically for that structure.

SAP is developing tabular foundation models, including RPT-1.6 (Relational Pre-trained Transformer) and TabPFN-3.5 Plus from Prior Labs, SAP’s AI lab. They are designed to work directly with structured business data, including missing values, mixed data types and fields with many unique values. AI Core’s tabular-model orchestration selects context and data at runtime so predictions can be incorporated into workflows.

Oren’s case is that business data is inherently relational and transactional, so models should be built to handle it. That still leaves a harder question: even if an agent can interpret the data and retrieve the relevant rules, can the company tell when its judgment is wrong?

Acting raises the cost of error

A bad answer can be checked or ignored. A payment blocked in error, an order rejected or a credit-risk flag assigned can immediately affect a supplier, customer or business process. Oren says agent responses must always be checked. Business context gives AI more information, but it also gives it boundaries.

On BAIP, agents operate under separate accounts with limited permissions. Audit systems record who took an action, what authority they had and what they did. AI capabilities undergo ethics and risk reviews, and agents are assessed function by function against regional rules, privacy laws and regulations.

Those controls also need to account for exceptions. Oren describes a “parked document,” SAP’s term for a document temporarily held from moving forward. After a company event, a controller may need to split catering invoices across several budgets. If the controller is away, an agent monitoring the process might treat the delay as a serious deadline breach. With more context, it could recognize that the responsible employee is on vacation or parental leave and treat the delay as an exception instead.

Oren describes enterprise AI as four layers:

The platform is the layer where shared definitions and process context can be reused. A finance agent and a supply chain agent should not each have to work out independently what “customer,” “supplier” or “transaction” means. Governance belongs there too, so actions remain traceable as agents spread across systems and departments.

SAP maps its products onto these layers, from tabular foundation models and BAIP to Joule Agents for departments including finance and supply chain management. Oren argues that models and agents cannot understand a business on their own: they need business context, knowledge graphs, semantics and governance. SAP’s claimed advantage is that its context and semantics cover both SAP data and other sources.

My read is that the architecture makes a stronger case for shared context than for autonomy. The announcement describes how agents can be given rules, identities and audit trails, but not how reliably they recognize the human exceptions that those systems fail to record. That distinction matters most when an agent moves from explaining a process to changing it.