Share with your CISO
SAP is making a direct argument that agentic AI, systems that don’t just recommend but act, live or die on the governance architecture built around them before deployment. Emmanuel Raptopoulos, SAP’s global president of customer success for Europe, APAC, the Middle East and Africa, told ITWeb that enterprise context, meaning how a company actually operates and makes decisions, is the precondition for safe, predictive AI operations. A Gartner figure cited in the piece puts stakes on the table: by 2027, 40% of enterprises may decommission autonomous agents after discovering governance gaps only post-deployment.
What this means for your business
The risk profile your board is worried about right now probably looks like a data breach or a model hallucination. The risk profile that actually keeps agentic AI deployments from surviving 2027 is subtler: an agent with a broadly defined objective function, the goal it’s been told to optimize, finding unanticipated paths to fulfil it inside your enterprise systems. If your organization has moved past AI pilots into production agents that can initiate transactions, change records, or communicate with customers, this story is directly about you.
Pierre le Roux of MOYO frames the governance shift cleanly, and it’s worth sitting with. Generative AI and copilots raised accuracy questions. Agentic AI raises authorization questions. Who granted the agent permission? What systems could it touch? Where did its authority end? Those aren’t IT questions; they’re audit and compliance questions, which means the CISO who hasn’t already mapped agent permissions into the existing access-control and identity governance frameworks is running behind the risk curve, not ahead of it. SAP’s incentive here is to position its platform as the trusted context layer, so the optimistic framing around controllability is predictable, but the underlying authorization problem it describes is real regardless of who you buy from.
Narrow objective functions are the practical design principle that comes out of this, and it deserves weight. An agent scoped tightly to one task, say, flagging supplier invoices over a variance threshold, has almost no latitude to cause harm outside that task. An agent scoped to “improve procurement efficiency” has enormous latitude. The difference isn’t the AI model; it’s the governance decision made at design time. If your organization is heading into a vendor renewal or a new deployment cycle for AI tooling, the question to push on isn’t capability. It’s how the vendor constrains agent scope by default and what audit trail the platform produces when an agent acts.
Concept deep-dive: Objective function
An objective function is the precise goal an AI agent is mathematically told to optimize, the difference between “answer this question” and “maximize customer retention.” The narrower and more specific it is, the fewer unintended paths the agent can take to satisfy it. A broadly written objective is like telling a contractor to “make the house better” with no specs: technically compliant outcomes can still be expensive surprises. In agentic enterprise AI, objective function design is a governance decision, not a technical one.
Based on reporting from SAP says enterprise context is key to predictive AI operations, originally published 2026-10-09 05:55:00.
