AI agents and human judgement

The arrival of AI agents is not another productivity upgrade. It is a change in who does the work. A chatbot answers a question and hands the result back to a person; an agent takes a goal, plans the steps, uses tools and acts, often across systems, often without anyone watching each move. When the machine acts rather than answers, the human's job moves up a level: from doing the task to deciding what may be delegated, setting the boundaries, and owning the outcome. The danger in the agentic shift is not that the agents are unsafe. It is that accountability and capability both drift away quietly, the work still gets done, but no one made the decision and no one is building the judgement to make the next one. The organisations that handle agents well will be the ones that decide, deliberately, where the human belongs.

What an agent is, and why it changes the question

The useful distinction is simple. A chatbot responds. An agent acts: it decomposes a goal into steps, calls tools and other systems, adapts as it goes, and completes a sequence with limited human involvement. The moment this became visible to everyone, as Rahim Hirji wrote in The Agents Are Here, was less about any single product and more about a shift in posture: software that does things on your behalf while you watch from the sidelines. The right response, he argued, is not the usual pair of lenses, excitement or fear, but a third: what does this do to human capability, and what should we keep firmly human? That is the question this page answers.

The human's new job

With agents, the human moves from operator to something closer to a principal with four roles. Intent-setter: defining what a good outcome is before the agent runs. Delegator: deciding what the agent may and may not do. Verifier: checking not just the output but whether the agent stayed inside its boundaries. Accountable owner: the named person who can explain the result and had the standing to stop it. This is Human at the Start applied to agents. The consequential judgement is made at the framing, before the agent acts, and owned at the end. Everything the agent does in between is execution, however autonomous it looks.

The delegation boundary map

The practical instrument is a map of what the human keeps and what the agent may take, decision by decision rather than task by task. For any consequential piece of work:

• What problem are we solving? Human keeps the problem definition and the framing of what success means. The agent may research and generate options.
• What constraints matter? Human keeps the ethical, legal, cultural and stakeholder judgement, and what must never happen. The agent may check work against those constraints once they are set.
• Which action do we take? Human keeps accountability and the final decision on consequential moves. The agent may compare scenarios, prepare and, within agreed limits, execute reversible steps.
• Did it achieve the intended outcome? Human keeps interpretation and ownership of the result. The agent may monitor and report.

The map is not a rule that a human must touch everything, which would defeat the point of agents. It is a decision about where the human touch has to be, made in advance, so that autonomy is granted deliberately rather than by default.

Human in the loop, or human on the hook

Boards reach for "we keep a human in the loop", and with agents the phrase quietly breaks. When an agent acts quickly and at scale, a person placed in the middle to review each step is either overwhelmed or reduced to a rubber stamp, which is exactly where oversight is weakest. As Rahim Hirji sets out in the accountability work, drawn from a landmark AI-denial case and European law, meaningful oversight requires someone with the authority and competence to change the decision, and producing an automated recommendation can itself be the decision. With agents, the loop is not the answer. A named human at the start, and a named human on the hook at the end, is. If no name can be attached to an agent's work, the work has not been delegated; it has been abandoned.

Automation complacency, multiplied

The risk the loop is meant to catch gets worse with agents, not better. Parasuraman and Manzey, reviewing decades of research across aviation, medicine and the military, found that people under-question confident automated output, that this affects experts as much as novices, that it cannot be trained away, and that it worsens under load and when attention is split across tasks. An agent running many actions at once is precisely a split-attention, high-load situation. The more the agent does, the less any human scrutinises, and the Harvard and BCG jagged-frontier experiment showed the cost: people who trusted AI beyond its competence performed worse than those with no AI at all. Complacency is not a character flaw here; it is the predictable result of designing humans into the weakest position.

What agents do to early-career development

There is a slower, more serious effect. If junior staff move straight to supervising agents rather than doing the underlying work, they never accumulate the repetitions that build judgement. The output looks senior; the capability is not. This is synthetic seniority and the missing rungs, accelerated, and across an organisation it compounds into capability debt: an operation that runs smoothly on agents until a decision arrives that the agent cannot make and no human in the room has been trained to. Agents make this cheaper to ignore and more expensive when the bill lands.

What leaders should do

Decide the delegation boundaries before deploying an agent, not after an incident: what it may do, what it may never do, and what would trigger a human override. Name an accountable owner for every consequential agent workflow, in writing, who can explain the outcome without reference to the tool. Treat verification of agents as real, skilled work, checking boundaries and reasoning, not just outputs, because in an agentic operation the checking is the judgement. Protect the reps: keep deliberate practice in the system so people still build the capability the agents are now exercising. And watch the development curve alongside the throughput curve, because an organisation can look more productive and grow less capable at the same time. The point of agents is to raise what people can get done. The job of leadership is to make sure it does not quietly lower what they can decide.

Key sources

• Hirji, R. (2026). The Agents Are Here. You're Just Not Paying Attention. Box of Amazing.
• Hirji, R. (2026). Why the Real AI Risk is Not Automation, but Accountability Gaps in Leadership Decisions. The European Business Review.
• Parasuraman, R. and Manzey, D. H. (2010). Complacency and Bias in Human Use of Automation. Human Factors, 52(3).
• Dell'Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier. Harvard Business School and BCG.

Related SuperSkills research

This extends the account of judgement to agents: Human at the Start, AI and human judgement, capability debt, synthetic seniority and drift versus design. On what cannot be delegated to an agent at all, see what stays human.

About this research

Rahim Hirji is the author of SuperSkills: The Seven Human Skills for the Age of AI (Kogan Page, 2026) and the founder of The SuperSkills Intelligence Company. This work draws on research across more than 200 organisations in 30 countries over seven years, and on his ongoing writing on the agentic shift. Findings are attributed to their sources and kept separate from the interpretation and frameworks, which are the author's. This is a living reference, and a fast-moving one: it is reviewed at least every 90 days as agent capability changes.