The battle for AI agent routes

The overwhelming amount of AI news and the unprecedented speed of updates have placed a huge cognitive burden on busy CIOs and decision makers. What are the key priorities that enterprises must consider when understanding AI and developing an adoption roadmap? Of course, there are two camps on AI agents, one represented by Microsoft and the other represented by SAP and Oracle. I personally think the following two points are the priority factors for developing a roadmap:

• The right AI agents, not the number, will drive business success

With the rapid advancement of technology, CIOs are under tremendous pressure to understand, plan, and execute. From an agent perspective, current generative AI (GenAI) experiences are on a continuum. Currently, these experiences are still relatively narrow and explicit, with limited autonomy. But over time, they will achieve higher autonomy through technological advances such as the new inference model O1, DeepSeek R1, and the Berkeley Sky-T1-32B, which costs only $450.

Technology is gradually driving agents to become more autonomous. While agents may replace traditional software in the future, the timing of this transition remains uncertain. However, enterprise applications will continue to exist for the foreseeable future, but their architecture and design will continue to evolve.

But if you think of agencies as a new workforce, the key to success is not quantity, but building the “right team” with the diversity of skills to solve complex business challenges.

For example, deploying multiple agents without clear guidance is like putting 50 experts in the same room and assigning broad tasks, which may eventually lead to directional errors. Similar to the real world, agents need to operate within the boundary conditions preset in applications such as SAP/Oracle. By combining the business insights of SAP knowledge graph, the boundary conditions of Signavio process graph and the power of large language model (LLM), enterprises can achieve the best results.

• Cloud-driven, scenario-enabled, business-ready

SAP deeply integrates AI into structured business processes, enabling agents to collaborate seamlessly across domains and ensuring efficient and controllable execution of complex tasks. Technologies such as SAP Knowledge Graph are crucial in this end-to-end process experience. Years of integration efforts have laid the foundation for this capability, giving SAP solutions a unique advantage.

Regarding the technical path, application scenarios and industry value of AI Agent, different parties have the following different views:

1. Technical architecture debate: standalone agent vs. embedded agent

1. Microsoft and other manufacturers:

It advocates separating AI agents from applications, and believes that agents should dominate processes as independent entities (such as Copilot being independent of the Office suite), and achieve cross-application collaboration by calling APIs or tools, emphasizing the "centralization" and "autonomy" of agents.

2. Enterprise application vendors such as SAP and Oracle:

It is believed that Agent should be deeply embedded in existing enterprise applications (such as ERP and HR systems) and naturally integrated with business processes and data.

SAP emphasizes "scenario-based integration" and uses knowledge graphs (Business Data Cloud) and process models (Signavio) to provide agents with structured business context to avoid "data islands" or loss of control of permissions due to being out of the system.

2. The battle of number of agents: large-scale deployment vs. precise configuration

1. Vision of some manufacturers:

The grand goal of "billions of agents" was proposed, believing that in the future every user and every task could be completed by a dedicated agent, covering all scenario needs through massive agents (similar to the expansion logic of the current ChatGPT plug-in).

2. Critical Views of SAP:

Quality over quantity:

Comparing Agent to a "new type of labor force", the key to a company's success lies in forming a "precise team" (such as specific Agents for scenarios such as financial closing and supply chain optimization) rather than blindly piling up numbers.

The boundary conditions need to be preset:

Unconstrained agent collaboration may lead to chaos (such as 50 experts discussing without a goal). It is necessary to limit the agent's behavior boundaries through business knowledge graphs and process rules to ensure controllability.

3. Business model battle: consumption-based pricing vs. value-based pricing

Traditional AI pricing model:

Billing by the number of users or call volume (such as by seat subscription or number of API calls) may cause a surge in costs for enterprises due to high-frequency use of Agents.

SAP's innovation strategy:

AI Unit model: allows customers to purchase AI computing resource pools in a unified manner, flexibly allocate them across applications (such as ERP, HR, and supply chain shared resources), and reduce fragmentation costs.

Hidden value proposition:

As the autonomy of agents increases, the traditional user-based charging model may become ineffective, and the pricing logic needs to be redefined based on the “actual business benefits created by AI” (such as improved process efficiency and reduced error rates).

4. Industry Implementation Competition: General Capabilities vs. Vertical Implementation

Limitations of General Agent:

Although LLMs have extensive knowledge, they lack understanding of industry-specific processes (e.g., drilling equipment maintenance in the oil industry, dynamic pricing in the retail industry), making it difficult to directly solve complex business problems.

SAP/Oracle's verticalization path:

Horizontal processes are prioritized:

Focus on common cross-industry scenarios such as finance, HR, and procurement, and quickly release value through standardized agents (such as Joule's accounts payable automation).

Ecological cooperation supplements long-tail demand:

Rely on partners to develop agents for specific industries (such as medical clinical trial management and predictive maintenance of manufacturing equipment), and integrate them through a unified platform to prevent customers from falling into the "customization quagmire".

5. Perception of technology maturity: radical substitution vs. gradual integration

Radical view:

It is believed that Agent will completely replace traditional software and become the new generation of interactive interface and decision-making core (such as "code-free Agent drives everything").

SAP's incrementalism:

• Long-term existence of enterprise applications:

Although we acknowledge that Agent autonomy is improving, enterprise applications (such as SAP S/4HANA) are still the core business carriers, and Agents need to evolve in tandem with them.

• The inevitability of architectural evolution:

Future application designs need to be compatible with agent interactions (such as API standardization and data semantics), but in the short term, more attention will be paid to "how to enable agents to enhance existing systems" rather than disruptive replacements.

Summary: The core logic behind the disagreement

Microsoft and other manufacturers:

Guided by technological disruption, we attempt to reconstruct the software ecosystem through Agent and seize the next generation of interactive entrance.

SAP and other enterprise application vendors:

With business value as the core, it emphasizes that Agents must respect the company's existing IT investment and process complexity, and achieve achievable efficiency improvements through "AI invisibility" (i.e. seamless embedding into business scenarios).

Customer perspective:

It is necessary to strike a balance between "technological foresight" and "implementation risk control" and choose an Agent path that matches its own digital maturity.

These differences are essentially a collision between technological idealism and business pragmatism, and also reflect the differentiated value demands of AI Agents in different industry scenarios.