An AI agent cancels a shipment, approves a refund, or sends a message to a client, and it does so without a human pressing the final button. The work is done. The question that surfaces the moment anything goes wrong is deceptively simple: who is accountable? In most organizations rolling out agents today, nobody has written the answer down, and the reason is that the old machinery of accountability assumed a person was always the last actor.
Accountability was built for humans
Every functioning organization runs on a chain of responsibility. A decision traces to a role, a role to a person, a person to a signature or an approval. That chain is what lets a firm answer to a customer, a regulator, or a court. Autonomous agents break the chain at its most important link, the point where a specific human chose to act, because the agent chose instead, often faster than any human could review.
Chan and colleagues (2024) frame the problem directly. As organizations delegate more activity to agents that pursue complex goals with limited supervision, the social risks grow and so does the need for what they call visibility: the ability to know who is operating a given agent and what it has done. Without it, an agent's actions float free of any human owner. The firm still bears the consequences, but internally there is no one to hold to account, which is another way of saying the decision was never really governed at all.
Three mechanisms, and their costs
Chan and colleagues (2024) propose three concrete instruments. The first is agent identifiers, so that any given agent's actions can be attributed to it and, through it, to the human or organization that deployed it. The second is real-time monitoring, so that an agent operating outside its intended bounds can be caught while it is acting rather than during a post-mortem. The third is activity logs, a durable record of what the agent did, which is the raw material for any after-the-fact review, dispute, or audit.
None of this is free, and the authors are careful to say so. Monitoring and identification sit in tension with privacy, and concentrating the ability to watch every agent can concentrate power in whoever holds the logs (Chan and colleagues, 2024). These are governance trade-offs, not engineering details, which means they belong to leadership rather than to the vendor's default settings. The choice of how much to log, who can read it, and how long it is kept is itself an accountability decision.
Onboarding is the missing discipline
The organizational fix looks less like a new technology and more like an old habit applied to a new kind of worker. The World Economic Forum (2025) recommends that AI agents be onboarded with the same rigor as a new employee, with a defined role, explicit permissions, and named supervision, and that governance scale in proportion to the risk the agent carries. Framed that way, accountability stops being a philosophical puzzle and becomes an HR-shaped process: no agent goes live without an owner, a scope, and a record.
Microsoft's 2025 survey gives the emerging role a name, the "agent boss," a human responsible for directing and answering for the agents under them. That is the human end of the accountability chain rebuilt for autonomous systems. It only works if it is real, meaning the agent boss has both the authority to stop the agent and the visibility, through the logs and monitoring Chan and colleagues describe, to know when to. An owner who cannot see what the agent did and cannot halt it is a name on an org chart, not an accountable party.
The practical read
The temptation is to treat accountability as a problem for later, after the productivity gains are booked. That order is exactly backward. McKinsey (2025) describes clients running "agent factories" where a small human team supervises dozens of specialized agents executing end-to-end processes; the more of the process the agents own, the more consequential the question of who answers for the outcome becomes, not the less. The gains and the exposure grow together.
So before an agent takes an action that reaches a customer, a payment, or a regulator, three questions should have written answers. Which specific agent is authorized to do this? Which named human owns it? And where is the record that lets someone reconstruct what happened? An organization that cannot answer those has not delegated a decision. It has abandoned one, and it will find that out at the worst possible moment.
Sources
- Chan et al. (2024). Visibility into AI Agents. Argues that delegating to autonomous agents demands visibility, and proposes agent IDs, real-time monitoring, and activity logs, while flagging privacy and power trade-offs. https://doi.org/10.1145/3630106.3658948
- World Economic Forum (2025). AI Agents in Action. Recommends onboarding agents like employees with defined role, permissions, and supervision proportional to risk. https://www.weforum.org/publications/ai-agents-in-action-foundations-for-evaluation-and-governance/
- Microsoft WorkLab (2025). Work Trend Index 2025. Introduces the "agent boss" role, a human accountable for directing the agents under them. https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born
- McKinsey (2025). The agentic organization. Describes small human teams supervising agent factories running end-to-end processes, raising the stakes of who answers for outcomes. https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/the-agentic-organization-contours-of-the-next-paradigm-for-the-ai-era