The New Division of Work

For most of the history of the enterprise, the division of work has been a division among people.

The operating model assumes that people are the actors and technology is the infrastructure around them.

That assumption no longer holds.

1. AI can research.
2. →It can analyze.
3. →It can generate.
4. →It can monitor.
5. →It can classify.
6. →It can coordinate.
7. →It can invoke tools.
8. →It can operate software.
9. →It can execute multistep work.
10. →It can maintain state across an assignment.
11. →It can act under supervision or, within defined boundaries, without waiting for a person at every step.

This does not mean that machines can perform every form of work.

It means the enterprise now has another source of execution capacity.

That changes the division of work.

The question is no longer simply how people should use AI.

The question is:

What work should belong to humans, what work should belong to machines, and what work should exist between them?

That is an operating model question.

And there is no durable answer based on the jobs that exist today.

The modern enterprise was designed around the constraints of human labor.

1. People have limited time.
2. →Expertise is scarce.
3. →Attention is finite.
4. →Coordination becomes harder as organizations grow.
5. →Knowledge is distributed.
6. →People forget.
7. →People need rest.
8. →People cannot monitor every condition continuously.

Organizations developed structures to manage those constraints.

1. Jobs bundle related activities.
2. →Teams aggregate capacity.
3. →Functions aggregate expertise.
4. →Managers coordinate people.
5. →Processes standardize execution.
6. →Applications store information and enforce transactions.
7. →Automation removes repetitive work where the rules can be specified in advance.

The resulting operating model contains an implicit division of labor.

Humans provide cognition, judgment and flexible execution.

Software provides storage, calculation, transaction processing and deterministic automation.

That boundary shaped the enterprise for decades.

Generative and agentic AI disturb it because machine capability is moving into territory that previously required human cognition.

1. Language.
2. →Reasoning.
3. →Synthesis.
4. →Interpretation.
5. →Generation.
6. →Planning.
7. →Tool use.
8. →Adaptive execution.

The boundary between human work and machine work is therefore no longer defined by whether a task requires cognition.

Both can now participate in cognitive work.

The operating problem becomes more precise.

One of the easiest mistakes in AI transformation is to treat technical capability as an allocation decision.

1. If a model can perform a task, automate it.
2. →If an agent can execute a process, delegate it.
3. →If the benchmark improves, expand autonomy.

That logic is incomplete.

Capability answers:

Allocation asks:

Those are different questions.

A machine may be capable of drafting a legal analysis.

That does not determine who should own the legal judgment.

An agent may be capable of negotiating a routine commercial term.

That does not establish the authority it should receive.

A model may diagnose a technical failure.

That does not mean it should execute a production change without control.

A system may generate a performance assessment.

That does not establish whether the enterprise should delegate a consequential judgment about a person to it.

Technical capability is therefore one input into work allocation.

It is not the governing principle.

The enterprise must also consider consequence, ambiguity, accountability, reversibility, trust, economics and the availability of effective controls.

Enterprise AI usage is beginning to move beyond assistance.

OpenAI's August 2026 Enterprise Signals reports that enterprise use is becoming more agentic, with work moving from asking AI for help toward delegating substantive tasks. As of June, Codex accounted for 64 percent of combined Codex and ChatGPT output tokens among enterprise customers. OpenAI also reported rapid growth in enterprise Codex use outside engineering, including legal, sales, recruiting and marketing.

Anthropic reports a related shift. Its June 2026 Economic Index describes Claude usage moving from conversations between a person and an assistant toward longer running agentic work. Anthropic distinguishes augmentation, where AI assists a person, from automation, where the system receives more of the task. Its data also shows higher levels of autonomy in Claude Code than in conversational use for comparable outputs.

Microsoft frames the change around human agency. Its 2026 Work Trend Index argues that as agents take on more execution, people can spend more of their capacity directing work, making decisions and owning outcomes. The research draws on Microsoft 365 telemetry and a survey of 20,000 AI users across ten countries.

These sources come from technology providers with commercial interests in greater AI adoption.

That matters.

Their telemetry is also bounded by activity inside their own ecosystems.

But across those ecosystems, the directional signal is consistent.

Machine participation is moving deeper into execution.

The allocation problem is becoming real.

The existing job architecture is a poor starting point for this decision.

A job may contain work that belongs in several different execution modes.

Consider a procurement professional.

The role may include gathering supplier information, comparing commercial terms, evaluating risk, negotiating with vendors, preparing recommendations, routing approvals, updating systems and managing relationships.

Those activities have different requirements.

1. Gathering supplier information may favor machine execution.
2. →Comparing structured terms may favor machine execution with defined validation.
3. →Interpreting an unusual contractual risk may require specialist judgment.
4. →Negotiation may require human participation where relationships, ambiguity and strategic context matter.
5. →Updating systems after a decision may again favor machine execution.

The role does not disappear.

The internal composition of the work changes.

The same logic applies across the enterprise.

An engineer may move from writing every line of code toward defining architecture, constraints and acceptance criteria while agents execute larger portions of implementation.

An analyst may move from gathering and formatting information toward framing the question, interrogating evidence and evaluating conclusions.

A manager may spend less time collecting status and routing work, and more time defining outcomes, resolving ambiguity and intervening in exceptions.

A customer service professional may move from processing routine cases toward handling the cases where policy, emotion or consequence requires human judgment.

This is why Chapter 03 moved the unit of transformation from the job to the workflow.

Once the workflow is decomposed, work can be allocated according to what each part requires.

The new division of work is not binary.

Human or machine is too crude.

A useful operating model needs at least four execution modes.

1. 01

   Human Execution

   The human performs the work.

   AI may provide information or tools, but responsibility for cognition and execution remains with the person.

   This mode remains appropriate where the work depends on forms of judgment, relationship, accountability or context that the enterprise has chosen not to delegate.

2. 02

   Human Led, Machine Assisted

   The human owns the work.

   AI expands the person's capability.

   It may retrieve information, generate alternatives, perform analysis, draft outputs, simulate scenarios or prepare recommendations.

   The human directs the work and determines the result.

   This is where much enterprise generative AI adoption began.

3. 03

   Machine Led, Human Governed

   The machine performs most of the execution.

   A human defines the objective, constraints and authority.

   The system performs the work and returns to a person at defined decision points, exceptions, confidence thresholds or control gates.

   The human is no longer executing every step.

   The human governs the execution system.

4. 04

   Machine Execution Within Authority

   The machine receives an objective and can execute within a defined authority envelope without requiring human approval for each action.

   The system may gather information, make bounded decisions, invoke tools and complete transactions.

   Human accountability remains.

   Human intervention moves to policy definition, monitoring, exceptions, evaluation and redesign.

   This is where agentic execution begins to alter the operating model most directly.

The strongest case for AI Native transformation is not that machines become more capable.

It is that human capacity can move.

Microsoft describes this as an expansion of human agency. In its 2026 research, 58 percent of surveyed AI users said they were producing work they could not have produced a year earlier, rising to 80 percent among the group Microsoft identifies as Frontier Professionals. Microsoft argues that organizations need to redesign work, not merely deploy tools, if they want to convert that capability into value.

Anthropic's survey evidence adds another dimension. Large majorities of surveyed Claude users reported gains in speed, scope and quality, while 68 percent reported learning more with AI and 57 percent said AI had made their skills more valuable. Users who delegated more work to Claude reported more optimism about the value of their skills, although Anthropic cautions that these are self assessments and do not rule out skill erosion.

The important question is what happens to the capacity released by machine execution.

If nothing changes, the enterprise has faster employees.

If the work is redesigned, the enterprise can change what human capacity is used for.

Human contribution can move toward work such as:

• problem definition
• goal formation
• judgment
• exception handling
• relationship
• negotiation
• ethical reasoning
• quality definition
• interpretation
• creative direction
• accountability
• organizational learning

These are not permanent categories of work that machines will never touch.

That claim would outrun the evidence.

They are areas where human participation can carry particular value because the work involves consequence, context, responsibility or forms of understanding that the enterprise may not want to delegate.

The boundary will move.

The operating model has to be able to move with it.

There is a risk in describing the future of work as machines execute and humans judge.

It sounds clean.

The reality is harder.

Research on generative AI and expertise is beginning to expose this tension.

A 2026 perspective in Patterns argues that generative AI can accelerate scientific work while also creating a risk of skill atrophy because the technology can perform some of the same cognitive activities through which expertise develops, including synthesis, experimental design and writing. The authors argue for capabilities such as strategic direction, critical discernment and systematic calibration rather than passive reliance on generated output.

Other empirical work points to uneven complementarity. A 2026 randomized experiment on AI assisted learning found average performance gains from generative AI, but the gains varied with participants' ability to elicit, filter and verify AI output.

The evidence base is still developing.

Different tasks produce different results.

Some studies find that AI narrows performance differences.

Others find that expertise changes how much value people can extract from AI.

The durable operating implication is not that expertise matters less.

It is that the form of expertise may change.

The human who no longer performs every execution step may still need enough domain understanding to define the problem, evaluate the result, recognize failure and intervene when the system leaves its operating envelope.

Traditional organizations develop expertise through participation in work.

1. Junior engineers debug.
2. →Junior lawyers research.
3. →Analysts build models.
4. →Doctors review cases.
5. →Managers learn by resolving operational problems.

Professionals acquire judgment in part by performing the lower level work that exposes them to patterns, exceptions and failure.

AI can absorb some of that work.

That creates an operating model problem that is easy to miss.

If machines take over the work through which humans learned the domain, where does the next generation of human judgment come from?

The answer cannot simply be to preserve low value work so that people can learn.

That would sacrifice operating leverage to maintain an old development model.

But the enterprise also cannot assume expertise will reproduce itself.

Learning has to become part of the architecture.

1. Humans may need deliberate exposure to exceptions.
2. →They may need to review machine reasoning and outcomes.
3. →They may need simulations.
4. →They may need rotations through consequential cases.
5. →They may need responsibility earlier for defining objectives, evaluating evidence and challenging machine output.

The development system has to evolve with the execution system.

The division of work cannot be fixed.

1. Model capability changes.
2. →Costs change.
3. →Workflow volume changes.
4. →Risk changes.
5. →Regulation changes.
6. →The enterprise learns from execution.

A task allocated to a human today may move toward machine execution tomorrow.

A machine executed activity may move back toward human review after a failure.

A workflow may use machine execution for routine cases and human execution for novel ones.

A more capable model may justify greater autonomy for one part of the workflow but cost too much to use across the entire process.

Allocation is therefore not an organizational decision made once.

It is an operating parameter.

The AI Native enterprise needs to know not only who or what performs the work today.

It needs to know why.

That reason should be observable.

It should be testable.

And when the evidence changes, the allocation should be capable of changing with it.

This creates a new management discipline.

The enterprise must manage the boundary between human and machine execution.

Not as ideology.

As architecture.

Once work can move between humans and machines, the boundary between them becomes part of enterprise architecture.

That boundary determines more than who performs a task.

1. It determines where judgment sits.
2. →Where authority sits.
3. →Where accountability sits.
4. →Where cost accumulates.
5. →Where control enters the workflow.
6. →Where expertise is required.
7. →Where exceptions go.
8. →And where the enterprise remains dependent on human capacity.

Consider a workflow with one hundred cases.

Eighty may be routine.

Fifteen may contain ambiguity.

Five may carry material consequence.

There is little reason to require the same execution model for all one hundred.

The routine cases may support machine execution within authority.

The ambiguous cases may require machine analysis followed by human judgment.

The consequential cases may require direct human ownership with AI providing evidence and alternatives.

The workflow can therefore contain different divisions of work at the same time.

This is more precise than deciding whether a process should be automated.

The unit of allocation moves deeper.

The enterprise allocates work according to the characteristics of the work itself.

Not all decisions deserve the same operating model.

A machine choosing which internal document to retrieve is not equivalent to a machine approving a financial transaction.

A system deciding how to format a report is not equivalent to a system deciding whether a customer receives credit.

An agent scheduling a meeting is not equivalent to an agent terminating a supplier relationship.

As consequence increases, the requirements around authority, evidence, review and accountability change.

This does not mean high consequence work must always remain human executed.

Financial markets, industrial systems and infrastructure already contain forms of automated execution where consequential actions occur under engineered controls.

The relevant question is whether the enterprise has designed an operating environment capable of supporting the delegation.

That includes the quality of the system.

But it also includes the quality of the control architecture around it.

1. What evidence does the system use?
2. →How is uncertainty represented?
3. →What actions are permitted?
4. →What thresholds constrain those actions?
5. →What requires approval?
6. →What conditions trigger escalation?
7. →Can the action be reversed?
8. →Can the enterprise reconstruct what happened?
9. →Who remains accountable?

The more consequential the action, the less defensible it becomes to treat machine autonomy as a product feature.

It becomes an institutional design decision.

Consequence is not the only factor.

Reversibility changes the economics of delegation.

Some machine actions are cheap to undo.

1. A draft can be discarded.
2. →A recommendation can be ignored.
3. →A software change can sometimes be rolled back.
4. →A customer communication may be corrected.

Other actions create consequences that are difficult or impossible to reverse.

1. Money moves.
2. →Data is deleted.
3. →A contract is executed.
4. →A regulatory filing is submitted.
5. →A public statement is released.
6. →A person is hired or dismissed.
7. →A physical system changes state.

The allocation model should reflect this difference.

Where actions are reversible, enterprises can often tolerate broader machine experimentation and faster execution.

Where actions are difficult to reverse, authority should narrow and controls should strengthen.

This creates a useful design principle.

Autonomy should expand where consequence is bounded, reversibility is high and observation is strong.

It should contract where consequence is high, reversibility is low or the enterprise cannot reliably observe what the system is doing.

This is not a permanent rule.

Better systems, controls and evidence may change the boundary.

But the burden of design should rise with the consequence of the action.

Machines perform best when the operating environment provides enough information to distinguish a good action from a bad one.

Some enterprise work has clear objectives.

1. Resolve the ticket within policy.
2. →Match the invoice to the purchase order.
3. →Identify the failed service.
4. →Generate the required test cases.
5. →Reconcile the transaction.

Other work contains competing objectives.

1. Preserve the customer relationship while enforcing policy.
2. →Reduce cost without damaging strategic capability.
3. →Respond to an employee problem where policy, context and fairness pull in different directions.
4. →Negotiate a partnership where the value depends on information neither party can fully quantify.
5. →Decide whether an unusual technical risk is acceptable in the context of a larger business objective.

These are not difficult only because they require more intelligence.

They are difficult because the objective itself may be contested, incomplete or contextual.

Human judgment often matters most where the enterprise has not reduced the problem to a stable objective function.

That makes ambiguity a design variable.

The more ambiguous the objective, the stronger the case for human participation in framing the problem, resolving competing values and accepting responsibility for the decision.

Machine intelligence may still perform much of the analysis.

1. It may surface evidence.
2. →Model alternatives.
3. →Identify contradictions.
4. →Simulate consequences.
5. →Challenge assumptions.

But assistance and execution should not be confused with ownership of the judgment.

Enterprises often respond to AI risk with a simple instruction.

Put a human in the loop.

That phrase is useful.

It is also dangerously incomplete.

A human cannot provide meaningful oversight merely by appearing somewhere in the workflow.

1. The person must have enough information.
2. →Enough time.
3. →Enough expertise.
4. →Enough authority.
5. →And a realistic ability to challenge the machine.

If an agent processes ten thousand cases and asks one employee to approve the result, the presence of the employee does not establish effective control.

If the person routinely accepts machine recommendations because reviewing the underlying evidence would take longer than performing the original work, the approval may become ceremonial.

If machine output arrives with apparent confidence and the human lacks the expertise required to identify subtle failure, human review may create the appearance of safety without the substance.

This is an old organizational problem in a new form.

Controls can exist on paper while failing in operation.

The AI Native enterprise therefore has to design human intervention as carefully as machine execution.

1. What exactly is the human evaluating?
2. →What evidence do they receive?
3. →What conditions trigger their involvement?
4. →How much work reaches them?
5. →What happens when they disagree?
6. →What expertise do they need?
7. →How is their decision recorded?
8. →How does the system learn from the intervention?

The answer cannot simply be:

A human approves it.

As machines absorb more execution, human work can move toward a different set of responsibilities.

1. From producing every artifact to defining what good looks like.
2. →From gathering every piece of information to deciding which information matters.
3. →From performing every analysis to challenging the assumptions behind the analysis.
4. →From routing every exception to resolving the exceptions that matter.
5. →From monitoring activity to interpreting outcomes.
6. →From executing policy to determining when policy no longer fits the situation.
7. →From coordinating tasks to designing the system through which tasks are executed.

This is not universally an upward movement.

Some machine systems will perform work that is intellectually demanding.

Some human work will remain procedural because human presence is required or economically rational.

Some occupations will change more than others.

Some forms of human labor may shrink.

Others may expand.

The distributional consequences will not be uniform.

But at the operating model level, one direction is important.

When machines can perform more of the execution, human capacity does not have to remain attached to execution merely because that is where it sat before.

The enterprise can redesign the human role around the places where human participation creates the most value.

Delegating execution does not delegate institutional responsibility.

An enterprise cannot treat an agent as the accountable party in the same sense that it can assign accountability to an executive, manager or regulated professional.

1. The machine can act.
2. →It can produce evidence.
3. →It can maintain logs.
4. →It can operate within policy.
5. →It can even evaluate parts of its own output.

But the enterprise remains responsible for deciding that the system should act.

That distinction is foundational.

The more execution moves into machines, the more explicit human accountability needs to become.

1. Who owns the workflow?
2. →Who determines the objective?
3. →Who defines the authority envelope?
4. →Who approves the controls?
5. →Who accepts the residual risk?
6. →Who responds when the system fails?
7. →Who decides whether the workflow should continue operating?

Machine execution can distribute action.

It cannot eliminate accountability.

The AI Native enterprise therefore separates two concepts that traditional work often bundles together.

The actor performing the work does not have to be the actor accountable for the operating system that produced it.

That separation makes greater autonomy possible.

It also makes governance more important.

Traditional organizations are constrained by human execution capacity.

1. There are only so many engineers.
2. →Analysts.
3. →Managers.
4. →Lawyers.
5. →Designers.
6. →Operators.
7. →Specialists.

Machine execution can loosen some of those constraints.

But it creates new ones.

1. High quality context becomes scarce.
2. →Clear objectives become scarce.
3. →Reliable evaluation becomes scarce.
4. →Authority becomes scarce.
5. →Trustworthy data becomes scarce.
6. →Human attention at the right intervention point becomes scarce.
7. →Compute may become scarce or expensive.
8. →The ability to design effective workflows becomes scarce.
9. →The ability to distinguish good machine output from plausible machine output becomes scarce.

The enterprise does not escape scarcity. The scarcity moves.

This has consequences for organization design.

If machine execution becomes abundant while judgment, context and authority remain constrained, the bottleneck shifts toward the people and systems that provide them.

A thousand agents do not create leverage if every consequential action waits for the same overloaded executive.

Machine execution does not remove the acceptance bottleneck.

Poor operating design can intensify it.

Machine execution has a seductive property.

It can appear almost infinitely replicable.

An enterprise can create another agent without hiring another employee.

That does not mean useful capacity scales without limit.

Every agent consumes something.

1. Compute.
2. →Tokens.
3. →Data.
4. →Tools.
5. →Permissions.
6. →Evaluation.
7. →Observability.
8. →Control.
9. →Human attention.
10. →Exception capacity.

The output of one agent may become the input constraint of another.

A workflow may generate more work for downstream humans than they can absorb.

More execution may create more exceptions.

More generated artifacts may create more review.

More analysis may create more decisions waiting for acceptance.

The system can therefore increase activity without increasing throughput.

This is the same mistake enterprises make when they optimize individual tasks.

Capacity has to be measured at the outcome.

If ten agents produce work faster but the workflow completes no more outcomes, execution capacity has increased while operating capacity has not.

That distinction becomes essential in the AI Native enterprise.

Machine capacity is not valuable because it exists. It is valuable when it changes the economics or performance of the outcome.

Human and machine allocation cannot be determined by capability and risk alone.

It has to make economic sense.

Machine execution has a cost structure.

1. Models consume inference.
2. →Agents may make repeated model calls.
3. →Long running workflows consume context and compute.
4. →Tools and data services carry costs.
5. →Evaluation consumes resources.
6. →Observability creates infrastructure.
7. →Human review consumes labor.
8. →Exceptions create additional work.
9. →Failures create remediation cost.

The cheapest model may create expensive errors.

The most capable model may cost more than the outcome justifies.

A human may remain economically preferable for low volume, high ambiguity work.

Machine execution may dominate high volume, structured work even when human execution remains technically possible.

Mixed execution may produce the strongest economics across many workflows.

The allocation problem therefore has another question:

What combination of human and machine capacity produces the required outcome at the best total economics?

That question cannot be answered at the license level.

It cannot be answered through token price alone.

It requires workflow economics.

The enterprise must understand the cost of producing the outcome under different allocations of work.

That is where productivity becomes operating leverage.

The new division of work is not a prediction about which jobs survive. It is an architectural principle.

Once machines can participate in cognition and execution, the enterprise should stop treating the existing distribution of work as fixed.

Every consequential workflow can be decomposed.

Every component can be examined according to its requirements.

1. Capability.
2. →Consequence.
3. →Ambiguity.
4. →Reversibility.
5. →Observability.
6. →Economics.

The result may preserve human execution.

It may create human led machine assistance.

It may shift toward machine led execution with human governance.

It may support machine execution within explicit authority.

And that allocation can change as the evidence changes.

This makes the boundary between human and machine work a managed part of the operating model.

Not a one time workforce decision.

Not an automation target.

Not a philosophical argument about whether humans or machines are better.

A design choice.

The objective is not to maximize machine work.

The objective is not to protect human work from change.

The objective is to build the strongest execution system for the outcome.

The result is not a human enterprise with AI attached to it.

Nor is it an autonomous enterprise from which people have been removed.

It is a mixed execution system.

That is the operating reality the AI Native enterprise has to learn to govern.

Start with the workflow.

Decompose the work.

For each material unit of execution, ask:

1. Can the machine perform it?
2. →Should the machine perform it?
3. →What is the consequence of failure?
4. →How ambiguous is the objective?
5. →Can the action be reversed?
6. →Can execution be observed?
7. →What authority would be required?
8. →What human capability must remain?
9. →What does each execution mode cost?
10. →What happens to the rest of the workflow if the allocation changes?

Then choose the execution mode.

1. 01

   Human Execution

2. 02

   Human Led, Machine Assisted

3. 03

   Machine Led, Human Governed

4. 04

   Machine Execution Within Authority

Do not treat the answer as permanent.

1. Instrument the workflow.
2. →Observe performance.
3. →Capture exceptions.
4. →Measure human intervention.
5. →Measure machine consumption.
6. →Measure the outcome.

Then change the allocation when the evidence justifies it.

The new division of work is not something the enterprise discovers once.

It is something the enterprise operates.

1. Chapter 03The workflow becomes the unit.
2. Chapter 04The work inside that workflow must be allocated across humans and machines.
3. Chapter 05Decision Rights and Delegated Authority.

Chapter 03 established the workflow as the primary unit of design.

Chapter 04 establishes that execution inside the workflow can no longer be allocated through the existing job architecture alone.

Humans and machines can both perform work.

But capability does not create permission.

The enterprise still has to decide who or what is allowed to act.

A machine may be capable of approving a transaction.

That does not mean it has the authority to approve it.

An agent may be capable of changing a production system.

That does not determine the conditions under which it should be permitted to do so.

A workflow may support extensive machine execution.

That does not answer who remains accountable when the execution creates consequences.

The new division of work therefore creates the next operating problem.

Decision rights.

Machine execution becomes enterprise execution only when those questions have explicit answers.

That is the subject of Chapter 05.

Chapter 05 · Decision Rights and Delegated Authority