SAP: Boards must govern agentic AI to protect margins

Manos Raptopoulos, Global President of Customer Success for Europe, APAC, the Middle East and Africa at SAP, told an audience ahead of the AI & Big Data Expo North America that corporate boards must tightly govern agentic AI to protect profit margins and core operations.

He described agentic AI as systems that can plan, reason, coordinate with other agents and execute workflows without human prompts. Because those agents can access financial, supply‑chain and customer records and make decisions at scale, Raptopoulos argued they should be governed in the same way organisations govern human teams.

Raptopoulos called for clear accountability, machine decision audit trails and defined thresholds for human escalation before organisations deploy autonomous agents. He said boards should resolve three baseline questions: who is accountable when an agent makes an error, how machine decisions will be audited, and what exact conditions require human intervention.

He listed governance components that companies must put in place, including agent lifecycle management, defined autonomy boundaries, policy enforcement and continuous performance monitoring. He warned that restricting an agent’s inference loop to prevent hallucinations can raise computational latency and hyperscaler compute costs and change initial profit‑and‑loss projections.

Raptopoulos noted a sharp accuracy requirement for business use cases, saying: “The distance between 90% and 100% accuracy is not incremental. In our world, it is existential.” He added that governance in these settings becomes an engineering constraint rather than a compliance checklist.

Geopolitical and regulatory fragmentation complicates governance decisions, he said. Sovereign cloud requirements, regional model limits and data localisation rules affect deployments in jurisdictions such as New York, Frankfurt, Riyadh and Singapore.

Raptopoulos said AI performance depends on the quality of enterprise data and processes. Fragmented master data, siloed systems and heavily customised ERP implementations produce unpredictable behaviour in agents. He recommended building relational foundation models trained on proprietary corporate records-orders, invoices, supply‑chain and financial postings-to improve forecasting, anomaly detection and operational optimization compared with generic large language models.

Preparing that data foundation requires engineering work: cleaning and indexing decades of planning data, converting poorly classified records into vector representations and building near‑zero‑latency data pipelines so agents can produce deterministic outputs. He noted many projects stall when organisations attempt to layer probabilistic intelligence on disjointed data estates.

On user interaction, Raptopoulos described a shift from menu‑driven interfaces to intent‑based workflows where employees state goals and agents assemble context and orchestrate actions. He said adoption depends on trust and on role‑specific AI personas for functions such as finance and supply‑chain that embed business rules, access controls and permissions into the model’s active memory.

SAP recommended a three‑layer scaling approach: embed persona‑driven features into core applications for quick returns, build agentic orchestration to coordinate cross‑system workflows, and develop industry‑specific intelligence for high‑value use cases. Raptopoulos cautioned against poor sequencing, saying organisations should align governance, data readiness and architecture funding before moving from pilots to wider production deployments.