The Job Is the Wrong Unit of Design for Agentic AI

Why the next wave of AI transformation may depend more on how we organise work than on the technology itself.

There is a slightly absurd version of the AI future in which an agent diagnoses a network fault in seconds, only for the answer to spend the next hour sitting in someone’s inbox waiting for approval.

We could end up with extraordinarily intelligent technology operating inside remarkably inefficient organisations.

And I suspect this is one of the more important challenges facing enterprises as they begin to scale agentic AI.

Most organisations are approaching the opportunity by asking which activities their existing employees perform that could be automated. It’s an understandable starting point, but one that takes the current organisation, its jobs, teams and management structures, largely as given.

What if those structures are themselves the constraint?

We have spent decades organising businesses around the economics of human labour. Jobs bundle together activities, expertise and responsibility. Teams group people with similar skills. Managers coordinate their work, allocate resources and make decisions.

Agentic AI changes some of the fundamental assumptions behind that design.

If execution and expertise can increasingly be separated from the person performing the work, why should the job remain our basic unit of organisational design?

I think answering that question will ultimately matter more than how many agents an organisation deploys.

The job is an organisational invention

For most of the history of the modern enterprise, a job has bundled together three distinct elements: the tasks someone performs, the knowledge and skills required to perform them, and the responsibility they carry for the results.

This made practical sense. If someone possessed the expertise required to undertake a task, it was generally efficient for that person to do the work. Dividing it between multiple people introduced hand-offs, delays and additional management costs.

Our organisational structures developed around that reality.

Agentic AI begins to change the economics.

An agent can execute a task without necessarily possessing all the experience or context of the person who previously performed it. Expertise can increasingly be codified into reusable instructions, tools and decision rules. Multiple agents can apply those capabilities simultaneously, while people retain responsibility for outcomes.

Execution, expertise and accountability no longer have to reside in the same place.

This creates the possibility of redesigning organisations around the work that needs to be done, rather than the jobs we have historically created to do it.

The real value is often between the tasks

Consider what happens when a major fibre route fails in a telecommunications network.

Imagine the fault occurs at 8.30am. Monitoring systems generate dozens of alarms. Engineers begin diagnosing the problem while customer service teams receive their first calls. Enterprise account managers want to know which customers are affected. The communications team needs an assessment of the likely duration and impact.

Everyone is working hard, often with different information.

Now imagine introducing AI agents into that environment.

One correlates network alarms and identifies the likely point of failure. Another calculates which services and customers are affected. A third prepares customer communications. A fourth reviews available restoration options.

Each performs its task in seconds or minutes. Significant productivity gains are possible.

Yet the network operations team may still be waiting for authority to implement a restoration plan. Customer communications may be awaiting approval. The enterprise team may be reconciling conflicting information from different systems. The incident manager is trying to bring everyone together to agree on a response.

The individual tasks have become faster. The organisational machinery surrounding them is largely unchanged.

This is a familiar problem in large enterprises.

Much of the cost and delay in complex work sits between activities, in the hand-offs, competing priorities, approvals and decisions required to connect them.

The bigger opportunity is to redesign the entire incident response.

An orchestration system could maintain a shared picture of the incident, coordinate specialised agents and route information and decisions to the appropriate people. Routine actions could proceed automatically within agreed limits. Decisions with significant consequences would be escalated with the necessary analysis already assembled.

The incident manager would spend less time assembling information and chasing updates, and more time resolving the genuinely difficult questions.

The same logic applies to customer onboarding, service provisioning, product development, procurement and countless other enterprise activities.

Automating individual tasks improves productivity within the existing structure. Redesigning workflows and decision rights creates the possibility of a different structure altogether.

The evidence points the same way. McKinsey’s research finds that AI high performers are substantially more likely to have fundamentally redesigned their workflows.

The implication is that we should be measuring the improvement in the overall outcome, including the coordination costs we remove, rather than simply counting the tasks we automate.

From business capability to human accountability

If jobs are no longer the natural starting point, we need a more precise way to describe how an organisation works.

I’ve been thinking about this through six connected elements. They describe how we move from what a business needs to accomplish, through how the work is performed, to who ultimately has responsibility for the result.

Using the network incident example:

  1. Capability: What the business must be able to do. Service assurance is an enduring organisational capability. It exists regardless of how the work is organised or who performs it.
  2. Workflow: How an outcome is delivered. Detecting, diagnosing, resolving and communicating a network incident forms an end-to-end workflow. It crosses multiple functions, systems and teams.
  3. Task: A discrete unit of execution. Correlating alarms, identifying affected customers or preparing a restoration plan are specific activities with defined inputs, outputs and standards of completion.
  4. Skill: The expertise required to execute a task. Fault-diagnosis rules, network topology analysis or customer-impact assessment are examples of expertise that can be codified, tested and reused by agents across different workflows.
  5. Role: Who is accountable for the outcome. An incident manager retains responsibility for effective service restoration, even where agents perform much of the underlying work.
  6. Decision: Where authority and judgement sit. An agent might recommend rerouting traffic or triggering customer notifications, but the authority to approve actions with significant commercial, safety or regulatory consequences remains explicitly assigned.

Once we separate these elements, we gain much more flexibility in how we design the organisation.

For example, a network diagnosis skill developed for incident management might also be used in proactive maintenance, capacity planning or customer assurance. It no longer needs to belong exclusively to one team or job.

And an accountable manager need not personally supervise every step in a workflow. Their role can be designed around outcomes, exceptions and judgement.

The unit of execution becomes the task. The unit of reusable expertise becomes the skill. The unit of accountability remains the human role.

This distinction is central to building an organisation that can scale an agent workforce.

The distinction between knowledge and judgement

There is a tendency in discussions about AI to divide work into what machines can do and what humans can do.

I don’t find that particularly helpful. The boundary keeps changing, and increasingly capable models will continue to move it.

A more useful distinction concerns the nature of the knowledge required and the consequences of getting something wrong.

Some expertise can be codified. Pricing rules, diagnostic procedures, established policies and repeatable analytical methods can increasingly be expressed in ways that an agent can execute and test.

Other knowledge is highly situational.

It might be a conversation with a supplier yesterday, an emerging concern from a major customer, a change in commercial priorities or an understanding of the personalities involved in a sensitive negotiation.

That information may be decisive, but it is difficult to anticipate and codify in advance.

The second consideration is reversibility.

If an agent makes an error in a routine, reversible task, the consequences may be modest. A pricing commitment, regulatory decision or action affecting customer safety is rather different.

This suggests a practical principle for allocating work: the autonomy given to an agent should reflect both the knowledge it can reliably access and the consequences of an error.

Where the knowledge is codified and the consequences are reversible, agents should be able to operate with considerable autonomy, subject to appropriate controls.

Where decisions are difficult to reverse, agents may still do much of the analysis and preparation, but a human should retain the relevant authority.

And where the required knowledge is largely tacit, the immediate opportunity may be to use agents to support people while learning which elements of their expertise can eventually be made reusable.

The human contribution becomes more precisely defined.

We need people to contribute context that the system cannot see, exercise judgement where the consequences warrant it, and accept accountability for the decisions that matter. In Australia, that accountability is becoming a legal expectation: from 10 December 2026, organisations must disclose when automated systems make, or substantially contribute to, decisions that significantly affect individuals.

That is a more interesting future for human work than simply checking everything the AI produces.

A useful working rule follows: people should step in when the AI is missing something they know, not simply when they would have done it differently.

What happens to management?

This raises harder questions about the future of management.

A surprising amount of management activity is coordination. Managers allocate work, monitor progress, resolve dependencies, move information between teams and ensure that decisions happen.

Increasingly capable orchestration systems could assume a meaningful proportion of this activity.

This doesn’t imply that management disappears. But it may change what managers are principally there to do.

I have been thinking about this in the context of the organisational models that emerged during the agile era, particularly the much-discussed Spotify model.

At its heart, Spotify’s approach sought to reconcile a familiar set of organisational tensions: autonomy and accountability, innovation and standardisation, and alignment and control.

These tensions become sharper when agents are involved.

An agent can operate far more quickly than a human team, potentially across organisational boundaries and at a scale that makes conventional supervision impractical.

Agents change how each tension resolves. Autonomy can be set by how reversible a decision is, rather than by seniority. Standards can still emerge bottom-up, as they did at Spotify, but the unit is now a codified skill that others reuse. And alignment can come from an orchestration layer that routes information and decisions, rather than from managers relaying it.

The answer cannot simply be more approvals and more management.

We need clearer rules about which decisions can be delegated, better mechanisms for evaluating agent performance, and explicit accountability for outcomes.

Human managers may increasingly spend their time defining those boundaries, stewarding the organisation’s reusable expertise and making the decisions that require judgement.

This could have quite profound implications for team structures.

A team might no longer be sized primarily according to the number of tasks it needs to complete. Its size and composition could instead reflect the volume and complexity of human decisions, the need for situational knowledge and the effort required to improve the underlying capabilities.

I wouldn’t suggest that we know what the ideal structure looks like. We have barely begun to experiment with these possibilities.

But it seems unlikely that the optimal organisation for a substantial agent workforce will look exactly like today’s organisation with a few digital colleagues added.

The economics are also changing

There is another reason to revisit organisational design: the economics of AI execution.

Much of the current discussion focuses on model selection, token consumption and infrastructure costs. These matter, particularly as agentic systems can consume substantially more computing resources than simple AI interactions.

But cost per token tells us relatively little about whether an organisation is becoming more productive.

The relevant measure is the cost of achieving an outcome, including human effort, agent execution, coordination, errors, rework and the costs of delay.

It also raises questions about which capabilities an organisation should own.

A generic capability that many businesses need may be economically sourced from an external provider. A proprietary skill that embodies a company’s competitive advantage is a different proposition.

Over time, an organisation’s library of codified expertise could become a significant asset. Skills developed for one workflow can be reused elsewhere, with each additional application potentially reducing the cost of the next.

This creates the possibility of compounding returns.

It also makes ownership of those skills, their underlying knowledge and the improvements made to them an important strategic consideration.

The overlooked problem is incentives

There is, however, a fairly substantial human complication.

We cannot expect people to redesign work in ways that may diminish the apparent importance of their existing jobs without reconsidering how we reward them.

An employee who discovers a way to complete a week’s work in a day may have little incentive to disclose it if the consequence is simply an increased workload or a reduced team.

Similarly, a manager rewarded for the scale of their operation may be reluctant to adopt a system that achieves better results with fewer people.

If we want organisational learning to compound, the incentives need to support it.

People should receive recognition for creating reusable capabilities, sharing knowledge and improving outcomes across the business, not simply for the amount of activity they personally supervise.

There is also a longer-term workforce question.

Much of what experienced professionals know was learned through years of performing relatively routine tasks, encountering exceptions and gradually developing judgement.

If agents perform much of that work, how will the next generation acquire the experience needed to become effective decision-makers?

Codifying skills, testing agents and understanding why they fail could become important forms of apprenticeship. But we will need to design these learning experiences deliberately. Judgement is difficult to develop without exposure to the consequences of decisions.

This may prove to be one of the more important challenges for organisations over the next decade.

An organisational experiment worth running

I don’t think we yet have sufficient evidence to prescribe a universal operating model for the agent workforce.

There are good reasons to be sceptical of claims that a particular organisational structure will produce a particular percentage improvement in productivity. Enterprise complexity varies, technology continues to evolve, and many of the most ambitious claims remain untested.

But the central hypothesis is testable.

Take a substantial workflow and compare the results from automating its individual tasks with those from redesigning the workflow, roles and decision rights around agents.

Measure the overall outcome, including cost, quality, cycle time and the speed at which additional capabilities can be deployed.

Then examine whether reusable skills make the second and third use cases cheaper and faster than the first.

If redesigning roles and decisions produces no material improvement over task automation, the organisational argument becomes much weaker.

If it does, we should be prepared to challenge some fairly longstanding assumptions about how enterprises operate.

For me, this is where the discussion about agentic AI becomes particularly interesting.

Over the past few decades, we have spent enormous amounts of money digitising activities, integrating systems and improving processes, generally within the boundaries of established organisational structures.

Agentic AI offers an opportunity to revisit those boundaries themselves.

The technology will continue to improve, probably faster than our organisations can absorb it. That makes the quality of organisational design increasingly consequential.

And perhaps the most useful question for leaders contemplating an agent workforce is a relatively simple one:

If we were designing our organisation today, knowing that much of its work could be performed by agents, would we create the same jobs, teams and management structures?

I suspect very few of us would.

References

  1. McKinsey & Company, The agentic organization: Contours of the next paradigm for the AI era.
  2. McKinsey & Company, The State of AI (November 2025): AI high performers and workflow redesign.
  3. Hyunjin Kim, Dahyeon Kim and Rembrand Koning (INSEAD and Harvard Business School), randomised field experiment with 515 firms on the AI “mapping problem”, as reported in Harvard Business Review.
  4. Harvard Business Review, on agentic AI orchestration: how human knowledge and AI combine in cross-functional decisions.
  5. Harvard Business Review, on digital labour: integrating AI agents into workforce strategy for HR and procurement.
  6. Tomas Chamorro-Premuzic, Harvard Business Review, on setting the stage for human- AI collaboration: measuring output rather than input.
  7. Michael Mankins and Eric Garton, Time, Talent, Energy: Overcome Organizational Drag and Unleash Your Team’s Productive Power (Harvard Business Review Press / Bain & Company): the three organisational tensions and the Spotify model.
  8. Office of the Australian Information Commissioner, guidance on automated decision- making transparency obligations under APP 1.7–1.9, summarised by Allens (October 2026).

Comments

Leave a Reply

More posts