The IBM 2025 AI Ethics Report: Values and Risks of AI Agent Systems in the Corporate Environment

AI Agents are a hot-topic for 2025 and what we are experiencing are progressive advancements of this new "paradigm" in three concurrent streams, in my opinion: - Autonomy and decision-making: A shifting rom simple task executors to systems capable of independent decision-making (in dynamic environments, able to adapt to changes, able to perform complex tasks without the human-in-the-loop); - Collaboration among agents: Multi-agent systems are becoming a standard and these agents can work together to solve complex problems; - Integration with LLMs: LLMs are powering their skiils with deep language understanding, reasoning, and decision-making (e.g. for multistep workflows and interacting seamlessly with humans / other systems). I have concerns about the acceleration we see for let them handle also the "last-mile", the decision part since now human-driven is progressively part of their workflows and that, in my opinion, should be carefully managed. Thanks to Paolo Rizza for making me aware of this paper published in March 2025 by the IBM AI Ethics Board where AI agents are explored with three perspective: the benefits (productivity, automation, augmentation), the risks (trust erosion, misalignment, security, job impact), the mitigation strategies (ethics board, governance tools, transparency, human oversight). AI agents are being published, promoted and almost recognized as powerful and use-case enabler but high-risk instruments that demand multilayered ethical guardrails and continuous monitoring. If the benefits are well promoted by the marketing and the hype, I'd like to focus on the risks: - Opaqueness due to limited visibility into how AI agents operate, including their inner workings and interactions; - Open-endedness in selecting resources/tools/other AI agents to execute actions (this may add to the possibility of executing unexpected actions); - Complexity emerges as a consequence of open-endedness and compounds with scaling open-endedness; - Non-reversibility as a consequence of taking actions thatcould impact the world. We shifted in a couple of years from a rule-based systems with minimal autonomy to a present where we have LLM-driven agents with autonomous tool execution to a future where it seems that multi-agent ecosystems with emergent, non-reversible outcomes will lead our processes... but what about this near future? I have created two infographics, witch you can find in the comments, where "side-effects" and "long-term effects" are pointed out because if they offer exponential value across enterprise functions on the other side they also introduce a new class of sociotechnical risk that demands evolved governance, ethical oversight, and rethinking of a human-AI collaboration models. What about the mitigation? IBM proposes its suite but generally speaking it's about an "end-to-end Governance". 🛜 IBM: https://shorturl.at/QlH4y #ArtificialIntelligence #AI #AIAgents #AIEthics #GovernanceAI #Risks #CyberSecurity