Getting your organization ready to scale responsible agentic AI

AI governance requires visibility and understanding of risks within AI models and systems, especially for emerging technologies like agentic AI. While AI agents can boost efficiency and productivity, their risks are amplified compared to traditional AI systems, according to IBM experts Manish Bhide, Heather Gentile, and Jordan Byrd.

Key Characteristics of AI Agents Introducing Risk

• Opaqueness: Limited visibility into an AI agent’s inner workings hinders understanding of its actions.
• Open-endedness: AI agents can self-select resources, tools, and other AI agents, increasing unexpected actions.
• Complexity: As AI agents learn and adapt, their inner workings become harder to analyze.
• Non-reversibility: Acting autonomously, AI agents may take irreversible actions with real-world consequences.

New and Amplified Risks

Agentic AI introduces new risks, such as data bias, where an AI agent might modify datasets in ways that introduce undetected bias. It also amplifies known risks, like system evaluation challenges and untraceable actions. For example, an AI agent with unrestricted access might inappropriately share sensitive or confidential information.

Mitigation Strategies

IBM recommends an end-to-end approach to risk mitigation, including:

• Human-in-the-loop: Ensuring human validation and feedback to maintain accuracy and alignment with organizational values.
• Holistic AI governance: Adapting best practices from other AI systems, such as generative AI and machine learning.

For deeper insights, IBM’s white paper, AI agents: Opportunities, risks, and mitigations, provides a thorough exploration of these challenges. Learn more about IBM’s approach to responsible AI.

Related Content

• How to choose the right foundation model
• DeepSeek’s reasoning AI shows power of small models, efficiently trained