SAP executive: Boards must govern agentic AI

Manos Raptopoulos, Global President of Customer Success for Europe, APAC, the Middle East and Africa at SAP, urged corporate boards to impose governance controls on agentic AI at an industry conference. He called for clear accountability, machine audit trails and defined thresholds for human intervention to limit operational risk and protect profit margins.

Raptopoulos described agentic AI as systems that plan, reason, coordinate with other agents and execute workflows on enterprise data. Because these agents interact with finance, supply chain and customer systems, he said they require the same oversight applied to human staff to prevent costly errors.

He noted that evaluation for large language models has shifted from general performance to precision, governance, scalability and measurable business impact. “The distance between 90% and 100% accuracy is not incremental. In our world, it is existential,” he observed, arguing that the gap between high accuracy and certainty can determine whether a deployment helps or harms the bottom line.

The executive outlined technical and cost implications. Integrating vector databases, which capture semantic relationships in corporate language, with legacy relational systems requires substantial engineering work. Limiting an agent’s inference loop to avoid hallucinations increases latency and cloud compute costs. High-frequency database queries can raise token and compute bills and change initial profit-and-loss expectations.

Raptopoulos set out three governance questions boards should resolve before deployment: who is accountable for an agent’s error, how machine decisions will be audited, and which thresholds require human escalation. He added that geopolitical fragmentation — including sovereign cloud and data localisation rules in markets such as New York, Frankfurt, Riyadh and Singapore — complicates those answers.

He said enterprise AI depends on a reliable data foundation. Fragmented master data, siloed systems and heavily customised ERPs introduce unpredictability that can scale damage when agents make recommendations affecting cash flow, customer relations or regulatory positions. Relational foundation models trained on proprietary corporate datasets — orders, invoices, supply chain records and financial postings — were described as better suited to forecasting, anomaly detection and operational tasks than generic internet-scale models.

On user interaction, Raptopoulos predicted a shift from manual navigation to intent-based experiences in enterprise software. Employees would state the outcome they want and AI agents would assemble context, orchestrate workflows and surface actions. He tied adoption to trust: employees must see that outputs follow governance rules, reflect accurate business logic and improve productivity. Building those interfaces requires role-specific AI personas mapped to access controls and permissions.

Customer service was identified as an early area for measurable returns. Models trained on internal records and logs can handle exception-heavy processes such as disputes, claims, returns and service routing by classifying cases, retrieving documents and recommending policy-compliant resolutions. Those deployments create customer-specific intelligence that is hard to reproduce with generic tools.

Raptopoulos advised pursuing three technical layers in parallel: embedding persona-driven functionality into core applications, building agentic orchestration for multi-agent workflows, and developing industry-specific intelligence for high-value problems. He recommended that boards align ambition with technical readiness and fund core architecture and updated data pipelines to move beyond pilot projects.