Governing Multi-Agent AI: SAP & Agilent Share Deployment Insights
The landscape of artificial intelligence is rapidly evolving, moving beyond the era of single AI assistants to embrace sophisticated networks of specialized agents. These multi-agent systems are designed to collaborate, self-critique, and intelligently select the appropriate model for each task, promising a new level of competitive differentiation for enterprises. However, this advancement introduces significant challenges, particularly concerning their deployment and governance in real-world environments.
A recent discussion highlighted these complexities, bringing together insights from industry leaders. Yaad Oren, managing director of SAP Labs U.S. and global head of research & innovation at SAP, emphasized the company’s commitment to enabling customers to scale their AI agents safely. He noted that while a degree of autonomy is achievable, robust checkpoints and continuous monitoring are crucial for improvement, vulnerability mitigation, and overall system health. Oren acknowledged that the technology is still nascent, describing current efforts as merely the “tip of the iceberg” in ensuring agent scalability and security.
Agilent, an analytical and clinical laboratory technology firm, is actively integrating AI across its operations, according to Raj Jampa, their SVP and CIO. While the initial results are promising, the company is grappling with the practicalities of scaling and vulnerability management. Jampa described Agilent as being in a “second stage” of AI adoption, moving beyond exploration to address challenges like enhanced AI monitoring and cost optimization.
Within Agilent, AI is strategically deployed across three core pillars. On the product side, the focus is on embedding AI into instruments to accelerate innovation. For customer-facing operations, the aim is to identify AI capabilities that deliver maximum client value. Internally, AI is being applied to operational efficiency, exemplified by the development of self-healing networks. Jampa stressed the paramount importance of a strong governance framework for these use cases, one that establishes policy-based boundaries and guardrails to balance compliance and security with operational flexibility. He cited a recent incident where an agent’s configuration update caused immediate issues due to a missing boundary check, underscoring the necessity of robust auditing and traceability for every input and output.
For complex decisions, particularly those involving natural language processing or large-scale translations, a human layer remains indispensable. Jampa explained that for such scenarios, the AI agent is designed to flag the need for human intervention and approval before proceeding. The inherent trade-off between speed and accuracy also comes into play early in the decision-making process, as complex models operating under low-latency requirements can rapidly escalate costs. A governance layer, therefore, becomes vital for monitoring agent performance across speed, latency, and accuracy, helping organizations refine and expand their AI strategies.
Integrating these new AI agents with existing enterprise solutions presents another major hurdle. While legacy on-premise systems can connect via data APIs or event-driven architectures, the optimal approach, according to Oren, is to transition all solutions to a cloud framework first. SAP assists enterprises in migrating their on-premise installations to the cloud, simplifying connections and delivery cycles. Once within a unified cloud infrastructure, the data layer becomes critical. SAP’s Business Data Cloud, for instance, serves as a unified platform that aggregates and semantically indexes data from both SAP and non-SAP sources, enabling agents to connect and create end-to-end business processes.
Successful enterprise AI deployments hinge on three critical elements: a clean, unified data layer; a robust orchestration layer; and an unwavering commitment to privacy and security. Oren highlighted that orchestrating agents is both a science and an art, crucial for preventing failures and ensuring effective auditing. Security and privacy are non-negotiable, especially as swarms of agents interact with sensitive databases and enterprise architecture. Identity and authorization management become paramount, ensuring that only authorized personnel or agents can access specific information, such as an HR team member viewing salary data.
Looking ahead, Oren envisions a future where human enterprise teams work alongside AI agents and robotic team members. In this evolving landscape, identity management will become even more vital. He concluded that while agents are increasingly perceived as digital colleagues, they demand heightened monitoring and management. This includes meticulous onboarding, authorization, and ongoing change management, treating agents as professional personalities that require continuous maintenance and improvement, albeit with more rigorous oversight than human employees.