Getting your organization ready to scale responsible agentic AI

In AI governance, you can’t govern what you can’t see. However, visibility alone isn’t useful unless you also understand the risks within your artificial intelligence (AI) models and systems.

This is especially true for emerging technologies such as agentic AI. AI agents can improve efficiency and enhance productivity, but understanding the full scope of risk they introduce is more challenging. “The risks for generative AI and machine learning can be significant to begin with, especially for certain use cases,” writes IBM’s Manish Bhide, Heather Gentile and Jordan Byrd. “Add in AI agents, and the risks are further amplified.”  

Our white paper, “AI agents: Opportunities, risks and mitigations,” provides a thorough investigation into agentic AI risks, exploring both the amplification of previously known AI risks and the emergence of new, unique challenges.

Building upon our previous work identifying risks and mitigations for foundation models, this paper equips practitioners with the foundational knowledge needed to understand, identify and mitigate risks. This is an important first step toward responsibly scaling agentic AI.

AI agents are highly autonomous, completing various tasks without continuous human oversight. They also possess 4 characteristics that can introduce risk:

·      Opaqueness: Limited visibility into an AI agent’s inner workings and interactions can hinder understanding of actions.

·      Open-endedness: AI agents can self-select resources, tools and even other AI agents to complete tasks, increasing the likelihood of unexpected actions.  

·      Complexity: As AI agents learn and adapt, their inner workings become more complex, making analysis increasingly difficult.

·      Non-reversibility: Acting without continuous human oversight, AI agents have an increased chance of taking irreversible actions with tangible consequences in both the digital and physical realms.

The autonomy and characteristics of AI agents present potential risks, challenges and societal impacts that practitioners must understand to responsibly scale agentic AI.

Agentic AI introduces new risks and challenges to the AI risk landscape, ones that practitioners are less likely to have considered in the design, development, adoption or governance of previous AI systems.

For example, one new emerging risk involves data bias: an AI agent might modify a dataset or database in a way that introduces bias. Here, the AI agent takes an action that potentially impacts the world and could be irreversible if the introduced bias scales undetected.

Agentic AI also amplifies several known risk areas, including system evaluation and the potential for unexplainable or untraceable actions. Practitioners should reevaluate these areas when working with AI agents.

For instance, an AI agent with unrestricted access to resources, databases or tools amplifies the risk of sharing sensitive or confidential information with users. Without proper guardrails, such an agent might store and inappropriately share personal information, intellectual property or other confidential data with system users. The white paper details these risks and challenges, explaining their origins and potential impacts.

Addressing the unique risks and challenges of agentic AI requires an end-to-end approach to risk mitigation, enacted through holistic AI governance. However, as IBM’s Phaedra Boinodiris and Jon Parker recently explained, “Agentic AI is advancing so quickly that organizations might have difficulty finding precedents or best practices for minimizing harms.”

Fortunately, many strategies that can help mitigate risks for other types of AI, such as generative AI and machine learning, can also help mitigate risks for agentic AI. For example, incorporating a human in the loop is a best practice for responsible AI of all types. Enabling human validation and feedback on the actions taken by AI agents can help ensure accuracy and relevance and maintain alignment with organizational values.  

Understanding the unique risks of agentic AI is a critical first step toward scaling it responsibly across the enterprise and realizing the return on investment (ROI) of responsible AI. “AI agents: Opportunities, risks, and mitigations” can help you conceptualize the agentic AI risk landscape more clearly and consider how your organization can responsibly capitalize on the immense opportunities presented by AI agents.

Read AI agents: Opportunities, risks, and mitigations

Learn more about IBM’s approach to responsible AI