How much autonomy should we give our AI agents?

Pascal Bornet, global AI authority and co-author of The Human-Agent Orchestrator: Leading and Scaling AI-Driven Organisations (2026), identifies the calibration of human oversight as one of the most consequential and most underestimated decisions leaders face in the agentic era. 

Why this question matters more than most leaders realise

Most enterprise AI discussions focus on capability: what can this agent do, how fast can it do it, and what will it cost? These are legitimate questions. But Bornet’s research, conducted across 432 organisations over four years, points to a different variable as the key differentiator between organisations achieving transformative results from AI and those stuck at incremental gains.

It is not the sophistication of the technology. It is the quality of the management surrounding it.

Organisations that deploy AI agents without a clear framework for human oversight tend to fall into one of two traps. Some apply blanket supervision to every agent action – approving every step, reviewing every output, inserting human judgment into processes where it adds friction rather than value. The agent moves at human speed. The productivity gains disappear. Others go too far in the opposite direction, extending autonomy broadly and quickly, only to discover too late that certain decisions require human judgment that is no longer in the loop.

Both traps are expensive. Both are avoidable. And both stem from the same root cause: treating autonomy as a binary on/off decision rather than a spectrum to be calibrated deliberately.

Autonomy as a spectrum

Bornet invites leaders to think about human oversight not as a fixed setting but as a variable that should be tuned to the specific characteristics of each task, each agent, and each context.

At one end of the spectrum, human judgment is present at every step, reviewing, approving, and directing before any action is taken. At the other end, the agent operates independently within defined boundaries, with humans setting objectives and guardrails rather than approving individual actions. Between these poles lies a range of calibrations, each appropriate for different combinations of risk, complexity, reversibility, and agent maturity.

According to Bornet, the trick is not to prescribe a single correct setting. The true value lies in making the calibration decision explicit, deliberate, and owned, rather than left to assumption, habit, or the path of least resistance.

In his latest book, The Human-Agent Orchestrator, he introduces a key tool, the Autonomy Dial, which helps businesses appreciate that the right level of oversight varies enormously depending on the stakes involved. At one end of the dial, humans are fully in control – every action the agent takes requires explicit human approval before it is executed. At the other end, the agent operates with full autonomy – it plans, decides, and acts without any human intervention in the loop. Neither extreme is correct in practice. The question is always: where on the dial should this particular agent, performing this particular task, in this particular context, sit?

How does the Autonomy Dial work

The Autonomy Dial provides a framework for thinking about human oversight of AI agents as a spectrum rather than a binary choice. 

Some useful starting points:

Risk and reversibility. What is the cost of an error here? Can it be corrected easily, or does it have downstream consequences that are difficult to unwind? Higher stakes and lower reversibility warrant tighter oversight, regardless of how capable the agent is.

Agent maturity. A newly deployed agent with a limited operational history should be supervised more closely than one with a proven track record in comparable conditions. Autonomy should be extended as performance is demonstrated, not assumed from the outset.

Environmental stability. Agents perform most reliably in structured, predictable environments. As complexity and novelty increase, so should the level of human involvement.

Genuine oversight. Human-on-the-loop sounds reassuring, but it requires real monitoring to mean anything. If the humans designated to oversee an agent lack the time, training, or clarity about what to look for, the oversight is nominal rather than real. Nominal oversight is worse than no oversight because it creates false confidence without providing genuine control.

Ownership. For every agentic deployment, there should be a named person or team responsible for the ongoing calibration of the autonomy level. Without clear ownership, the dial stays wherever it was left at launch.

The dynamic dimension: the dial should move

Perhaps the most important aspect of the Autonomy Dial concept is that it is not a one-time setting. It should move deliberately, in both directions, as circumstances change.

As an agent demonstrates reliable performance across a well-defined task set, it is both appropriate and efficient to extend greater autonomy, reducing the overhead of human oversight where the track record justifies it. Equally, when an agent encounters a novel situation, when risk parameters shift, or when the regulatory environment evolves, the dial should move back toward tighter human involvement.

This dynamic calibration is what separates organisations that scale agentic AI safely from those that experience avoidable failures. Autonomy should be earned, maintained, and reviewed – not granted permanently at deployment.

The governance dimension

Bornet also believes it’s crucial for IT leaders to grasp that agents do not take control – humans abdicate it.

In the absence of clear, documented governance – explicit decisions about oversight levels, review cadences, and escalation criteria – autonomy tends to expand by default. Not through dramatic failure, but through gradual drift. Agents do more. Humans check less. Oversight erodes in ways that are invisible until something goes wrong.

The practical implication is that every agentic deployment requires a governance framework as deliberately designed as the agent itself. Who owns the oversight decision? Under what conditions should the autonomy level be reviewed or adjusted? What are the escalation triggers that bring human judgment back into the loop? These are not questions that can be answered once at deployment and then left alone. They need to be live, owned, and revisited as the agent’s operating environment evolves.

For Australian organisations operating in regulated industries – financial services, healthcare, government, legal – this governance dimension is not merely best practice. It is increasingly a compliance expectation, as regulators globally begin to develop specific frameworks for agentic AI oversight.

A leadership shift

The Autonomy Dial reflects a broader shift in what effective leadership looks like in an agentic organisation. As Bornet puts it in The Human-Agent Orchestrator: “Old leadership controlled the work. New leadership designs the system that produces it.”

Calibrating the Autonomy Dial well is an act of system design. It requires understanding risk, agent capability, governance requirements, and the human factors that determine whether oversight will be genuine or merely nominal. It requires revisiting those calibrations as conditions change. And it requires an organisational culture in which humans see their role not as approving every action, but as setting the conditions under which agents can act well and be trusted to do so.

The organisations getting this right are not simply deploying more capable agents. They are developing more capable orchestrators: leaders who understand how to calibrate trust, oversight, and autonomy in a way that unlocks the full potential of human-agent collaboration without exposing the business to unnecessary risk.

That is the real competitive advantage in the agentic era. And it starts with knowing where to set the dial, and who is responsible for turning it.

Who is Pascal Bornet?

Pascal Bornet is not a commentator on AI. He is someone who has spent decades building it, implementing it, and studying its impact at the coalface. Bornet spent over two decades as a senior executive at McKinsey and EY, where he established and led their Intelligent Automation practices and implemented AI and automation initiatives for hundreds of organisations worldwide. He has been consistently ranked among the top 10 global leaders in AI and automation, is a member of the Forbes Technology Council, and has delivered keynotes at over 100 events worldwide each year. He has authored four best-selling books on the subject: The Human-Agent Orchestrator, Agentic Artificial Intelligence, Irreplaceable and Intelligent Automation.

How PhoenixDX helps organisations calibrate their AI governance

Getting the Autonomy Dial right is not just a strategic question; it’s a practical one, and it is precisely where PhoenixDX operates. We work with organisations at the intersection of AI strategy and delivery, helping leaders move beyond the question of which agents to deploy and into the harder, more consequential work of how to govern and lead them well. We combine deep advisory expertise with hands-on implementation capability, working alongside executive teams to define their AI governance frameworks, establish the oversight models and escalation criteria that keep humans appropriately in the loop, and build the practical foundations – technical, organisational, and human – that make agentic AI safe to scale. That is precisely where PhoenixDX can help.