Agentic AI refers to a class of AI systems that autonomously execute multi-step tasks within a defined environment, calling tools, making decisions, and adapting to changes in context. Unlike traditional generative AI systems that provide a single response to a single prompt, agents operate in a sequence of perception, planning, tool invocation, observation, and adaptation. This terminology emerged in 2024 and 2025, with Microsoft, Anthropic, OpenAI, Google, and SAP as key drivers.
In the context of SAP change management, Agentic AI means that software agents perform activities previously handled by human developers, consultants, or change managers. Examples: An agent generates ABAP code from a requirements description. An agent creates a change record, classifies it, and routes it to the appropriate approval level. An agent checks a transport against security and quality rules and blocks it if deficiencies are found. An agent analyzes the impact of a planned change on the production landscape.
The operational shift in the SAP world in 2026: SAP-native agents are emerging from the Joule Studio Agent Builder. Agents that access SAP systems using non-SAP models like Claude, GPT, or Gemini are emerging from the MCP server landscape. Code-generating agents are emerging from AI coding tools like Claude Code, Cursor, and GitHub Copilot. These three worlds overlap, compete, and complement each other simultaneously.
In 2026, SAP customers are faced with three different agent ecosystems, each with distinct governance requirements:
World 1: SAP-native agents (Joule Studio). These agents are developed and deployed in the Joule Studio Agent Builder. They run within the SAP ecosystem, use SAP APIs, and are integrated into SAP processes. The governance chain is clear: SAP is the provider, the customer is the deployer. Changes to these agents go through Joule Studio and must be incorporated into the change process.
World 2: MCP-based external agents. The Model Context Protocol (MCP) makes it possible to connect external AI models (Claude, GPT, Gemini) to SAP systems via specialized MCP servers. These agents run outside of SAP but access SAP data and processes via APIs. The governance question is complex: Who is the provider, and who is the deployer? How are changes to these agents documented?
World 3: AI coding agents. Tools like Claude Code, GitHub Copilot, and Cursor generate code, including ABAP code, which is deployed into SAP systems. These agents are not visible in the SAP change process unless the generated code is explicitly recorded as a transport request. The AI-generated code runs through the standard transport process, but its AI origin is not documented.
A robust Agentic AI governance architecture for SAP environments consists of three layers:
Inventory layer. All agents that access SAP systems or influence SAP processes are recorded: name, type (SAP-native, MCP-based, coding agent), purpose, affected SAP systems, provider, deployer, risk classification according to the EU AI Act, and change manager.
Pipeline layer. Each agent has a defined deployment pipeline: How is the agent developed or configured? How is a new version tested? How is it moved into production? Who approves the production release? The pipeline is the technical implementation of the change process for agents.
Control point layer. Before any agent update goes live, the change must pass defined checkpoints: risk classification verified (has the class changed?), impact analysis performed (which SAP processes are affected?), test coverage ensured (has the new behavior been tested?), human-in-the-loop activated (has a human confirmed the result?), and audit trail maintained (is the change documented in the change record?).
The speed-governance paradox. Agents are deployed because they complete tasks faster than humans. However, rapid execution and robust governance are in tension: an agent that makes decisions in seconds requires a governance framework that still makes those decisions traceable. Without a human-in-the-loop design, there is no point at which a human can review the agent's actions.
The MCP governance vacuum. MCP server-based agents run outside the SAP change process. They can read SAP data, trigger processes, and generate ABAP code without any of these activities being automatically recorded in the change system. The MCP ecosystem is evolving rapidly, but governance standards are still emerging.
The provider-deployer vacuum for custom agents. Anyone building their own agent in Joule Studio becomes a provider under the EU AI Act (Article 25: substantial modification). The provider obligations (Article 16, conformity assessment, technical documentation) then lie with the customer, not SAP. Many customers are unaware of this shift in roles.
Human-in-the-loop as a design principle. No productive agent should act completely autonomously without defined points where a human reviews and approves. These points are determined in advance: during initial deployment, for every change, for unexpected results, and when thresholds are exceeded. Human-in-the-loop is not a convenience feature, but a regulatory requirement under Article 26 (2) of the EU AI Act.
Agent inventory as a mandatory foundation. Before the first production agent runs, an inventory must be established: which agents exist, what is their risk class, and who is responsible. The inventory is the foundation for EU AI Act compliance, as well as operational control: who knows which agents have access to which SAP systems?
Change process for AI modifications. Every change to a productive agent must go through the change process. This applies to model updates (SAP provides a new version of the Joule agent), configuration changes (in Joule Studio), prompt updates (modified instructions for the agent), and tool changes (new APIs the agent is permitted to use). The change record includes the agent name, the type of change, the impact on the risk class, and the human-in-the-loop verification.
By 2026, Agentic AI in SAP change management will no longer be a futuristic concept, but an operational reality across three domains: SAP-native agents, MCP-based external agents, and AI coding agents. Each domain has its own governance requirements. Human-in-the-loop as a design principle, agent inventory as a mandatory foundation, and a change process for every AI modification are the three operational consequences.