SAP data is finally allowed to leave SAP. What agents need is for the meaning to travel with it
In one week SAP, Databricks and Google opened new paths to SAP data and SAP semantics for AI agents. They solve four different problems, and only some of the pieces are generally available.
Listen to this article · 4 min
AI-generated narration of the full article.
For twenty years, getting data out of SAP meant extraction: a pipeline that copied tables into a warehouse, stripped them of the business context that made them meaningful, and left governance behind at the border. A field called NETWR arrived in the warehouse; the knowledge that it is a net order value in document currency, and which documents it applies to, stayed in SAP.
That was tolerable for dashboards built by people who knew SAP. It fails for agents, which read fields literally and at scale. In the last week of October, three vendors announced different answers to that problem, and it helps to separate them.
What was announced
SAP, at TechEd Berlin on October 29, detailed new Business Data Cloud capabilities. It completed the Dremio acquisition and now offers federated SQL across lakehouses and databases such as SQL Server, Oracle, MongoDB and Postgres without moving data, plus native Apache Iceberg and a universal catalog powered by Apache Polaris. It completed the Reltio acquisition, and SAP Reltio is generally available in BDC for multi-domain master data with AI-based entity resolution. And it introduced Ontology Manager, an AI-assisted way to build domain-specific views of the SAP Knowledge Graph, which connects business entities, relationships and rules. SAP’s own example is a collections agent view covering customers, invoices, payment terms and credit risk.
Databricks, on October 30, described the next phase of SAP Databricks inside BDC. SAP Databricks and BDC Connect are available today. Genie, Lakeflow, Unity Gateway and Lakebase inside SAP Databricks are planned, as is a two-way integration between Databricks’ Genie and SAP’s Joule.
Google, on October 29, made expanded semantic model support in Knowledge Catalog generally available. It ingests business definitions from Looker, BigQuery, dbt MetricFlow, Snowflake, Databricks and, through BDC, SAP’s Core Schema Notation. When two tools define the same metric differently, both appear with their provenance so data stewards can reconcile them. Google cites internal benchmarks of nearly four times higher agent accuracy and up to 75 percent fewer tokens; they are Google’s numbers with undisclosed methodology.
Four problems, not one
It is tempting to file all of this under “semantic layer”. The announcements address four different ways an agent fails with enterprise data.
| Agent failure it addresses | Status | |
|---|---|---|
| Federation and zero-copy (Dremio, Iceberg, BDC Connect) | Cannot reach the data, or reads a stale copy | Shipped by SAP; Databricks additions planned |
| Entity resolution (SAP Reltio) | Treats two records of one customer as two customers | Generally available |
| Business meaning (CSN in Knowledge Catalog, Ontology Manager) | Uses the wrong definition of a metric or relationship | Catalog support GA; Ontology Manager newly introduced |
| Acting in the application (Genie and Joule) | Can answer but not act where the process lives | Planned |
An agent can have perfect access and still double-count a customer. It can resolve customers correctly and still compute margin with a definition finance abandoned last year. Each row needs its own owner and its own test, which is the reason we argued that an agent’s context needs a specification with meaning, evidence and freshness as separate properties.
The question that comes next
When we wrote about agents reading business definitions where they live in October, the open question was whose definition wins. Google’s answer, show both with provenance, is honest and incomplete: it surfaces the conflict and leaves the decision to a steward.
SAP’s position is different and, for SAP customers, persuasive. The ERP is where transactions are recorded, so it should be where the meaning of those transactions is defined, and agents should read that meaning rather than reconstruct it. It is consistent with SAP’s bet that the most valuable agents stay inside the transaction system.
For architecture teams, the practical sequence follows from the table. Decide who owns each critical definition before buying a catalog. Treat entity resolution as a data quality program, not an AI feature. And plan around what is generally available today; Genie inside SAP and the Genie and Joule integration are still roadmap.