“Securing Profits Through AI Governance in the Enterprise”

SAP explains that enterprise AI governance protects profit margins by shifting from statistical estimates to precise, deterministic control.

If you ask a consumer-level AI model to count the words in a document, it will frequently be off by around ten percent. Manos Raptopoulos, Global President of Customer Success for Europe, APAC, Middle East & Africa at SAP, points out that in enterprise operations, the gap between nearly perfect and fully accurate is not a small step—it is a critical divide.

“The jump from 90% to 100% accuracy is not gradual. In our context, it is a matter of survival,” Raptopoulos states.

As companies deploy large language models into live production systems, Raptopoulos highlights that the benchmarks for success have officially shifted to focus on accuracy, governance, scalability, and measurable business outcomes.

The urgent issue for corporate leadership involves moving from passive tools to active digital agents—a shift Raptopoulos describes as the defining governance challenge. This topic will be a key focus for SAP at this year’s AI & Big Data Expo North America.

Today’s agentic AI systems can plan, reason, coordinate with other agents, and carry out workflows independently. Because these systems directly handle sensitive data and shape decisions at scale, Raptopoulos contends that failing to manage them with the same rigor applied to human employees puts the organization at serious operational risk. He cautions that unchecked agent proliferation will resemble the shadow IT problems of the last decade, but with far greater consequences.

His framework insists on implementing agent lifecycle management, setting clear limits on autonomy, enforcing policies, and maintaining continuous performance oversight as essential practices.

Merging modern vector databases—which capture the meaning behind enterprise language—with older relational systems requires significant engineering resources. Teams must carefully limit the agent’s reasoning process to prevent errors from disrupting financial or supply chain operations. Enforcing these tight controls increases computational delays and raises cloud computing costs, which can alter initial financial forecasts.

When an autonomous model depends on frequent, high-speed database queries to ensure reliable results, token expenses accumulate rapidly. Governance thus becomes a core engineering challenge, not just a compliance formality.

Raptopoulos maintains that corporate boards must address three fundamental questions before launching agentic models: determining who is responsible for an agent’s mistakes, creating clear audit trails for automated decisions, and setting precise criteria for when human intervention is required. Geopolitical complexities make resolving these questions even more difficult.

Sovereign cloud platforms, localized AI models, and data residency requirements are now regulatory realities in key markets including New York, Frankfurt, Riyadh, and Singapore. Businesses must embed deterministic control directly into probabilistic AI systems. Raptopoulos sees this as a strategic leadership priority, not merely an IT initiative.

Organizing relational intelligence for business operations

AI systems are only as reliable as the data and processes they rely on—a point Raptopoulos refers to as the data foundation moment.

Disorganized master data, isolated business systems, and heavily customized ERP setups create dangerous unpredictability at critical times. Raptopoulos notes that if an autonomous agent uses flawed data to make recommendations affecting cash flow, customer relationships, or compliance status, the resulting harm can escalate immediately.

Unlocking real business value requires moving beyond general-purpose large language models trained on public internet content. According to Raptopoulos, true enterprise intelligence must be rooted in proprietary corporate data—such as orders, invoices, supply chain logs, and financial entries—woven directly into business workflows. He asserts that relational foundation models designed specifically for structured business data will consistently outperform generic models in forecasting, detecting anomalies, and optimizing operations.

The difficulty of making a heavily customized ERP environment understandable to a foundation model stalls many AI projects. Data engineering teams often spend excessive time cleaning up fragmented master data just to establish a usable baseline for the AI.

When a relational model must accurately analyze complex, proprietary supply chain records alongside raw invoice data, the underlying data pipelines must operate without delay. If data ingestion fails, the model’s predictive accuracy drops instantly, making the agent a potential risk to the business.

Connecting legacy systems with modern relational AI demands a major overhaul of long-established data pipelines. Engineering teams must index years of poorly categorized planning data so that embedding models can produce accurate vector representations. Following Raptopoulos’s reasoning, leadership must honestly assess whether their current data infrastructure is truly ready, rather than simply adding probabilistic AI on top of disconnected systems.

Creating intent-driven interfaces

How users interact with enterprise software is evolving from fixed interfaces to dynamic, generative experiences—a shift Raptopoulos identifies as the employee interaction moment.

Instead of manually navigating complex software, employees will simply state their intent to the system. For example, Raptopoulos describes a user asking the software to prepare a briefing for their highest-revenue customer meeting that week. The AI agents then coordinate the necessary workflows, gather relevant context, and suggest recommended actions.

However, Raptopoulos emphasizes that employee adoption depends entirely on trust. Workers will only accept these digital collaborators when they are confident the system’s outputs follow governance rules, align with real business logic, and deliver clear productivity improvements.

Building these systems requires role-specific AI personas designed for positions like CFO, CHRO, or head of supply chain. Raptopoulos observes that these personas must be grounded in trusted data and integrated into familiar business processes to bridge the adoption gap.

This level of integration is a high-stakes design choice. Companies that invest in AI-native architecture see faster returns, while those trying to attach probabilistic models to outdated interfaces face major challenges with trust, usability, and scalability.

Technology leaders attempting to layer modern AI orchestration onto rigid, monolithic applications often encounter serious integration bottlenecks. Routing probabilistic API calls through legacy middleware slows down user interfaces, breaking the seamless intent-based workflow. Developing role-specific personas goes beyond prompt engineering—it requires embedding complex access controls, permissions, and business rules into the model’s active memory.

Building a competitive moat

The financial payoff from AI becomes most visible in customer-facing interactions. Raptopoulos explains that training models on proprietary records, internal policies, and historical data creates a layer of customer-specific insight that competitors cannot easily replicate. This approach