Key points
- Irfan Khan says enterprise AI agents cannot coordinate reliably until they have business context such as purchase order status and inventory levels.
- Khan argues leaders should measure end-to-end productivity rather than generated content, which he calls an activity measure rather than a result.
- SAP is building company memory within its process integration layer and targeting around 50 unified data products by year end, while stressing openness to avoid vendor lock-in.
SAP Says AI Agents Stall Without Business Context and Governed Data
SAP argues that enterprise AI agents stall after the pilot because they lack business context, governed data and a sound way to measure value. Its answer is a platform that contextualizes data from SAP and non-SAP sources, applies governance end to end and stays open to other data lakes and models. Khan says openness matters because frontier models change quickly.
Takeaways
- Agents fail without business context. Agents coordinating across systems need to know the state of a purchase order, orders on hand or inventory level. Without that shared context, passing instructions between agents is pointless. Closing this context gap is the first requirement for an autonomous enterprise.
- Reliable AI depends on reliable data. Problems surface when a proof of concept runs at scale. Data quality, governance and security must be in place first. Khan says the most valuable uses today are unglamorous, such as task automation and approval workflow management.
- Governance must match deterministic IT systems. An independent agent must face the same security and fine-grained access controls as conventional IT. Customers are rightly wary of ungoverned agents because of reputational risk, and they add manual checks, like verifying actuals in a spreadsheet.
- Measure end-to-end productivity, not output. Generated text or content measures activity, not results. Khan points to productivity across the whole chain, from the CFO's view of earnings and EBITDA down to developer productivity, so value is judged across an entire process.
- Token costs need guardrails early. Organizations that start without guardrails hit thresholds that force review and then drastic change. Khan recalls AWS replacing roles with AI that consumed roughly twice the cost, though he flags uncertainty. SAP is now choosing which tasks justify tokens.
- Company memory gives agents institutional knowledge. Agents do not remember business patterns such as a Black Friday or fourth-quarter peak. SAP's company memory initiative, inside its process integration layer, supplies what agents should and should not do. Trust also varies by task, since order entry and inventory need precision.
- Open data layers avoid vendor lock-in. Business Data Cloud ingests SAP and non-SAP data, adds Dremio, and shares zero-copy with Databricks and Snowflake over Iceberg and Polaris. A cost center spread across SAP, Salesforce and Snowflake becomes one data product, with about 50 targeted by year end.