AI agents are changing the manager's job. Accountability hasn't caught up

Your next boss probably won't be an AI agent.

But your next manager may be responsible for three people, a dozen agents and a workflow that keeps running while they sleep.

That changes the job.

Agents are starting to absorb reporting, meeting preparation, coordination and follow-through. At the same time, they are creating new management work: setting objectives, controlling access, checking outputs, handling failures and remaining accountable when the system gets something wrong.

We are not automating management away. We are unbundling it.

And most organisations are still measuring a manager's job by the number of humans on the organisation chart.

People will work with agents. Working for them is different.

The technology industry increasingly talks about AI agents as "digital workers". That metaphor is useful up to a point.

Workday's global research found that 75% of respondents were comfortable teaming with AI agents, but only 30% were comfortable being managed by one.

People appear far more willing to work with AI than for it.

I think that makes the "AI boss" debate a distraction.

The more immediate question for CIOs and business leaders is this:

What happens when an agent performs work that used to sit inside the manager's role, while the human manager still owns the outcome?

That is already beginning to happen.

Management work is already moving

At CVS, agents are reportedly preparing reports and dashboards that previously consumed much of a two-hour daily management meeting. The meeting has been reduced to around 15 minutes, focused more heavily on exceptions and decisions.

That does not mean an AI system has become the manager. It means part of the manager's work has moved into the system.

At Edward Jones, AI is being used to maintain context, commitments and follow-through across programmes - work that historically depended heavily on managerial memory and coordination.

Broadridge CTO Tyler Derr has raised the scenario of a leader overseeing three human employees and 19 agents. That is a scenario, not an industry benchmark.

And most current enterprise deployments remain bounded and human-supervised rather than examples of autonomous management. But the direction is clear enough: management work is being redistributed between people and machines.

Someone still needs to decide what matters. Someone still needs to investigate the exception. Someone still owns the result.

The governance gap is already visible

This is where the issue stops being a future-of-work thought experiment.

Deloitte's 2026 enterprise research found that 69% of Australian respondent organisations were using autonomous AI agents, while only 22% reported highly advanced agent-governance models.

That is a significant gap between deployment and control.

And it raises three questions I think every organisation should be able to answer before putting an agent into a live operational process:

Who owns the outcome?

What exactly is the agent allowed to do?

When must it stop and escalate to a person?

If nobody can answer those clearly, the agent is not ready for production.

That sounds obvious. In practice, it often isn't.

Business teams may create agents because they solve a local problem quickly. IT may not know they exist. Security may not understand their permissions. Managers may assume someone else is monitoring them.

The old shadow-AI problem was employees using tools the organisation did not know about.

The next problem is potentially more serious: agents taking actions the organisation does not fully see.

A person pasting text into a chatbot creates one kind of risk. An agent with credentials that can retrieve records, update systems, trigger workflows, contact customers or instruct another agent creates another.

That is shadow AI with more authority and potentially much higher error cost.

Headcount no longer tells you a manager's real span of control

Traditionally, span of control meant something fairly simple: how many people report to this manager?

That is becoming a weaker measure.

One manager may have eight human reports. Another may have three people, several AI agents and responsibility for an automated workflow processing thousands of transactions or customer interactions.

The organisation chart says the first manager has the larger team. Operationally, the second may carry substantially more responsibility.

An organisation can cut the number of people reporting to a manager and still increase that person's operational load. Headcount will not show it.

That matters because agents do not simply remove work. They also create work around verification, permissions, exceptions, reliability and consequential decisions.

So before organisations start redesigning management layers around AI, they need to understand what managers are actually being asked to supervise.

This is why agent deployment should sit within a broader AI strategy and roadmap, rather than becoming a collection of disconnected automation experiments.

The CIO cannot own this alone

The technology team has an important role. But IT cannot own the business outcome by itself.

Business leaders need to own the result. Security needs to define access. Risk and legal need to understand consequential decisions. HR needs to understand how roles are changing. Managers need to know how to supervise work that is increasingly performed by a combination of people and machines.

And that management capability is different from learning how to write better prompts.

Managers increasingly need to understand how to define an outcome clearly enough to delegate it, decide what can safely be automated, assess whether an output is reliable, recognise exceptions and know when a process should stop.

Some individual contributors will acquire those verification and exception responsibilities too, even if they never receive a traditional direct report.

That is why building AI capability in leadership teams increasingly needs to go beyond showing executives how to use ChatGPT or Microsoft Copilot.

The bigger question is how to manage work performed partly by humans and partly by machines.

Gartner's latest CIO research highlights the gap. Only 4% of surveyed CIOs said their organisations had explicit plans to redesign roles for a human-agent workforce. Organisations that did were 2.4 times more likely to fall into Gartner's higher-performing AI cohort.

We are moving faster on the technology than on the operating model around it.

You can delegate the work. You cannot delegate the accountability.

AI agents will continue to become more capable. Some will remain assistants. Others will become delegated workers inside defined processes. Some will act with increasing autonomy.

But autonomy should never mean ambiguity.

My operating rule for boards and leadership teams would be:

We will not put an agent into an operational process until we have named the human who owns the outcome, defined what the agent is allowed to do, and established when it must stop or escalate.

That is the line that matters.

The organisations that get agentic AI right will not necessarily be those deploying the most agents.

They will be the organisations that can answer one very simple question when something goes wrong:

Who owns this?