At 2:13 in the morning, an AI agent identifies a problem.
A customer account appears to be at risk.
Usage has fallen.
Support tickets have increased.
A renewal is approaching.
The agent reviews the customer history.
It summarizes recent conversations.
It identifies an unresolved product issue.
It drafts a recovery plan.
Another agent calculates a commercial concession.
A third creates a technical remediation task.
A fourth sends a message to the customer-success manager.
By the time the employee wakes, the organization has already analyzed the problem, proposed a response, updated several systems, and scheduled follow-up work.
This sounds efficient.
It may be.
But then the questions begin.
Was the usage data complete?
Did the agent interpret the customer’s frustration correctly?
Was it authorized to recommend a discount?
Did it expose confidential information to another model?
Why was the engineering task prioritized above other commitments?
Which employee owns the customer relationship?
Who approves the message?
What happens if the agent acts before a human sees it?
Who is accountable if the customer leaves?
The agents performed the work.
The organization remains responsible.
This is the central tension of the agentic enterprise.
AI agents can increasingly research, interpret, generate, decide within boundaries, coordinate workflows, and take actions across systems.
They are more than passive tools.
They participate in execution.
But they are not employees.
They do not possess moral agency.
They cannot be promoted, prosecuted, inspired, exhausted, ashamed, or trusted in the human sense.
They cannot stand before a customer, regulator, board, or court and accept responsibility.
They can execute.
They cannot be accountable.
The enterprise has spent more than a century designing structures for productive human beings.
Roles.
Departments.
Managers.
Employment contracts.
Performance systems.
Authority.
Career paths.
Policies.
The next enterprise may contain far more productive machine actors than human ones.
Its operating model is not ready.
The Number of Agents Could Grow Much Faster Than Headcount
Human organizations expand slowly.
A company must recruit.
Interview.
Hire.
Onboard.
Train.
Manage.
A new employee creates significant productive potential, but also cost, legal responsibility, cultural impact, and long-term commitment.
An AI agent can be created quickly.
One team may deploy agents for:
-
Research
-
Coding
-
Testing
-
Documentation
-
Customer support
-
Sales preparation
-
Financial reconciliation
-
Compliance monitoring
-
Incident analysis
-
Workflow routing
-
Data validation
-
Knowledge retrieval
Each agent may run continuously.
Agents may create tasks for other agents.
They may operate across hundreds or thousands of cases.
A company with five thousand employees could eventually operate tens of thousands of specialized agents.
A company with fifty employees may use hundreds.
The ratio could increase rapidly because agents are not constrained by the economics of employment.
But this creates an important distinction.
Creating an agent is easy.
Creating a reliable organizational actor is not.
A prompt is not a role.
A workflow is not governance.
A tool connection is not authority.
A successful demonstration is not operational readiness.
The enterprise will discover that deploying agents is much easier than absorbing them responsibly.
Headcount Will Stop Describing Organizational Scale
For generations, employee count provided a rough indication of enterprise capacity.
More employees usually meant more production, more customer coverage, more operational reach, and more management complexity.
That relationship is already weakening, as explored in The Company After Headcount.
When agents perform meaningful work, the number of human employees tells us even less.
Imagine two companies.
The first has one thousand employees and uses AI mainly for occasional drafting and search.
The second has three hundred employees, two thousand production agents, automated verification systems, and deeply integrated human-agent workflows.
Which is larger?
The first has more people.
The second may possess more active productive capacity.
But counting agents creates another distortion.
One simple summarization agent is not equivalent to an autonomous agent that can operate across customer, financial, and operational systems.
Ten poorly designed agents may produce less value than one reliable workflow.
The future company cannot be measured simply by:
Employees + agents.
It must understand the quality, authority, reach, and reliability of its entire execution system.
Organizational scale will include:
-
Human judgment
-
Machine execution
-
Data access
-
Decision authority
-
Workflow coverage
-
Verification
-
External capability
-
Reconfiguration speed
Headcount will still matter.
It will describe the human core.
It will no longer describe the whole company.
AI Agents Are Not Software Seats
Companies are accustomed to purchasing software.
Employees use applications.
The software supports their work.
Responsibility remains with the employee and manager.
An AI agent changes this relationship.
A conventional application waits for a human instruction.
An agent may:
-
Observe a condition
-
Interpret information
-
Choose among permitted actions
-
Use several tools
-
Trigger additional workflows
-
Escalate exceptions
-
Learn from feedback
-
Continue operating without continuous human direction
This makes the agent a productive actor inside the execution system.
It does not make the agent a person.
But it means the company cannot govern it as though it were merely another software licence.
A software seat answers:
Who may use this tool?
An agent raises different questions:
-
What is this agent trying to achieve?
-
Which actions may it take?
-
Which systems may it access?
-
Which data may it read?
-
Which data may it modify?
-
What decisions may it influence?
-
Which conditions require human review?
-
Who owns its performance?
-
Who can suspend it?
-
How is every action recorded?
The enterprise needs a new category between application and employee.
Every Agent Needs an Identity
Employees have identities inside the organization.
They receive email addresses.
System accounts.
Permissions.
Managers.
Departments.
Legal obligations.
Agents also require identity.
Not a generic shared API key.
Not a hidden background process.
A visible organizational identity.
Each agent should have:
-
A unique name or identifier
-
A clearly defined purpose
-
An institutional owner
-
A technical owner
-
Permitted tools
-
Approved data sources
-
Decision boundaries
-
Operating environment
-
Version history
-
Verification requirements
-
Escalation rules
-
Revocation capability
If an agent updates a customer record, the organization should know which agent acted.
If it recommends a payment, the system should preserve the evidence used.
If it sends a communication, the company should know which version of the agent generated it.
If its behaviour changes after a model update, the organization should be able to trace the change.
An agent without identity is shadow labour.
It performs work without occupying a visible place in accountability.
Access Will Become the First Line of Agent Governance
A human employee may receive broad access because the organization trusts the employment relationship and managerial structure.
Even this model is increasingly risky.
For agents, broad access is unacceptable.
An agent can operate rapidly.
It can repeat an error at scale.
It may interact with multiple systems before a human notices.
It can expose information through tool use or generated output.
Agent access should therefore be:
-
Purpose-scoped
-
Task-scoped
-
Data-scoped
-
Time-scoped
-
Environment-scoped
-
Action-scoped
A research agent may read approved documents but not modify records.
A collections agent may generate payment options but not execute a financial transaction above a defined threshold.
A coding agent may create a pull request but not deploy directly to production.
A customer agent may answer routine questions but escalate regulated or emotionally sensitive cases.
The principle is simple:
Agents should receive the minimum authority required to produce the intended outcome.
This is also the direction required by the borderless company described in The Enterprise After Borders.
The future enterprise boundary will not be defined only by employment or location.
It will be defined through precise relationships among identity, responsibility, access, and time.
An Agent Needs an Owner, Not Merely a Creator
The person who builds an agent is not always the person who should own it.
A developer may implement the workflow.
A business leader may depend on the outcome.
A risk leader may define the controls.
An operations manager may handle the exceptions.
Ownership must be explicit.
The owner should be accountable for:
-
The business purpose
-
The quality of the output
-
The appropriateness of automation
-
The escalation design
-
The consequences of failure
-
Continued relevance
-
Suspension or retirement
This is not necessarily day-to-day supervision.
It is institutional responsibility.
An agent without an owner becomes everybody’s tool and nobody’s liability.
When the agent performs well, several teams claim the success.
When it fails, responsibility disappears into technology, business, vendor, and model boundaries.
The company must prevent this.
Every production agent should have a named human or institutional owner.
Managing Agents Is Not the Same as Managing Employees
A manager motivates people.
Builds trust.
Develops capability.
Handles conflict.
Provides feedback.
Creates psychological safety.
Understands personal circumstances.
Agent management is different.
Agents do not require inspiration.
They require:
-
Objective definition
-
Boundary design
-
Tool control
-
Evaluation
-
Monitoring
-
Version management
-
Exception handling
-
Performance comparison
-
Retirement
The future manager may be responsible for both.
They may lead ten humans and govern fifty agents.
The humans contribute judgment, context, relationships, creativity, and responsibility.
The agents perform research, production, monitoring, routing, and analysis.
This is not simply a larger span of control.
It is a mixed operating system.
The manager needs to understand:
-
Which work belongs to people
-
Which work belongs to agents
-
Where collaboration occurs
-
Where verification is required
-
What happens when they disagree
-
Which failures are technical
-
Which failures are organizational
The manager becomes an architect of human-machine execution.
Departments May Become Portfolios of Workflows
The traditional department groups people with related expertise.
Finance.
Marketing.
Engineering.
Legal.
Customer success.
In an agentic enterprise, a department may increasingly operate as a portfolio of human-agent workflows.
Consider finance.
Some work may be handled by:
-
Reconciliation agents
-
Forecasting agents
-
Invoice-review agents
-
Policy-checking agents
-
Narrative-generation agents
Human finance professionals may focus on:
-
Judgment
-
Exceptions
-
Strategic modelling
-
Regulatory responsibility
-
Business partnership
-
Capital allocation
The finance leader no longer manages only a team.
They govern a production system.
The same shift can occur across the enterprise.
Customer support becomes a network of agents, human specialists, escalation rules, and knowledge systems.
Engineering becomes a system of human architects, coding agents, testing agents, security controls, and deployment workflows.
Legal becomes a combination of review agents, knowledge systems, specialist counsel, and human approval.
The department remains.
Its productive composition changes.
The Org Chart Will Become Increasingly Incomplete
The traditional org chart shows human reporting relationships.
It will not show most of the productive actors inside the agentic enterprise.
A customer-success leader may formally manage twenty people.
Operationally, the function may depend on:
-
Five service agents
-
Three analytics agents
-
A renewal-risk agent
-
A support-classification agent
-
A customer-communication agent
-
Several external systems
The org chart shows twenty employees.
The execution system contains far more actors.
This is why the company needs the execution-graph model discussed in From Org Charts to Execution Graphs.
The graph can show:
-
Outcomes
-
Humans
-
Agents
-
Systems
-
Decisions
-
Dependencies
-
Controls
-
Verification
The org chart shows where people belong.
The execution graph shows how work happens.
In the agentic enterprise, both are necessary.
Agent Sprawl Will Become the New SaaS Sprawl
Companies adopted SaaS rapidly.
Teams bought tools independently.
Over time, organizations discovered:
-
Duplicate applications
-
Unused subscriptions
-
Inconsistent data
-
Integration risk
-
Security exposure
-
Vendor dependency
-
Unclear ownership
Agents could spread even faster.
Every team can create one.
Every platform can include them.
Every employee may develop personal workflows.
Soon, the company may contain:
-
Duplicate agents solving the same problem
-
Agents using inconsistent data
-
Agents producing contradictory recommendations
-
Abandoned agents with active credentials
-
Agents built on outdated policies
-
Unapproved model connections
-
Agents whose owners have left
-
Agents triggering one another in loops
-
Agents generating costly activity without measurable value
Agent sprawl is more dangerous than software sprawl because agents act.
An unused application wastes money.
An unmanaged agent can create consequences.
The company will need an agent registry.
A living inventory showing:
-
Purpose
-
Owner
-
Status
-
Access
-
Model
-
Version
-
Cost
-
Usage
-
Performance
-
Risk classification
-
Dependencies
-
Retirement date
Without this registry, the enterprise may not know how many machine actors operate inside it.
Every Agent Will Need a Job Description—But Not a Human One
An agent needs a clear operating contract.
The contract should define:
Objective
What result is the agent expected to support?
Scope
Which tasks fall inside its mandate?
Prohibited actions
What may it never do?
Data
Which information may it access and retain?
Tools
Which systems can it use?
Authority
Which actions can it take independently?
Human review
Which outputs require approval?
Escalation
When must it stop and involve a person?
Verification
How is its work tested?
Performance
Which metrics determine whether it remains useful?
Retirement
Under what conditions should it be suspended or removed?
This is closer to a technical constitution than a conventional job description.
It defines an actor’s rights and limits inside the enterprise.
The Largest Risk May Not Be a Wrong Answer
Much public discussion about AI risk focuses on hallucinations.
Agents may produce incorrect information.
That matters.
But in production environments, the greater risks may involve:
-
Correct information used in the wrong context
-
Authorized data combined in an unauthorized way
-
A locally rational action that harms the wider system
-
Repeated small errors at scale
-
Agents optimizing the wrong target
-
Conflicting agents acting simultaneously
-
Overreliance by humans
-
Invisible shifts in behaviour after model updates
-
Decisions that cannot be explained to affected people
An agent may accurately identify a customer as unprofitable.
It may still be wrong to terminate the relationship.
An agent may correctly identify a fraud pattern.
It may still create unacceptable discrimination.
An agent may accurately optimize delivery speed.
It may do so by bypassing a safety control.
The problem is not only truth.
It is consequence.
Agents Will Create a New Cybersecurity Surface
Every agent with tool access becomes a potential pathway through the enterprise.
It may connect to:
-
Email
-
CRM
-
Financial systems
-
Code repositories
-
Customer databases
-
Internal documents
-
Identity systems
-
Operational platforms
A compromised or manipulated agent can operate using legitimate permissions.
This creates new attack possibilities.
An attacker may not need to breach a system directly.
They may influence the agent through poisoned data, malicious instructions, deceptive content, or compromised tools.
The agent may then perform the harmful action itself.
Security must therefore extend beyond conventional identity and application protection.
Organizations will need to consider:
-
Prompt injection
-
Data poisoning
-
Tool misuse
-
Model manipulation
-
Agent impersonation
-
Cross-agent contamination
-
Unintended information leakage
-
Autonomous privilege escalation
-
Agent-generated malicious code
-
Supply-chain risk from third-party models and tools
The machine workforce will require zero-trust design.
Not because every agent is malicious.
Because agents can act at machine speed inside human systems.
The Human Review Layer Can Become the New Bottleneck
Organizations may initially respond to agent risk by requiring human approval for everything.
This feels safe.
It may make the system unusable.
If agents generate ten times more work but every output requires detailed human review, the company may simply move the bottleneck.
Production becomes fast.
Approval becomes overwhelmed.
Humans begin approving superficially because the review volume is too high.
The appearance of oversight remains.
The quality of oversight declines.
The company must design review proportionately.
Some work may require:
-
Full human approval
-
Sampling
-
Automated verification
-
Exception-based review
-
Dual control
-
Post-action monitoring
-
No human review for low-risk, reversible actions
The right model depends on consequence.
A low-risk internal summary should not require the same oversight as a financial transaction or medical recommendation.
Human review must be meaningful, not ceremonial.
Exception Handling Will Become a Core Human Function
Agents perform best where patterns are clear.
The greatest human value often appears at the boundary.
The unusual customer.
The ambiguous regulation.
The novel failure.
The emotionally sensitive situation.
The conflict between two valid objectives.
The exception that does not fit the workflow.
As routine work moves toward agents, human work may concentrate around exceptions.
This sounds attractive.
It may also be exhausting.
Humans could receive only the hardest, most stressful, and least predictable cases.
The routine work that once provided rhythm, confidence, and learning may disappear.
Organizations must design exception work carefully.
People need:
-
Context
-
Authority
-
Time
-
Support
-
Feedback
-
Recovery
-
Opportunities to shape the system
They should not become an emergency service cleaning up machine uncertainty all day.
The Junior Role Is at Risk Before the Senior Role
Many professions develop expertise through apprenticeship.
A junior employee performs simpler tasks.
Receives feedback.
Observes experienced colleagues.
Gradually takes on greater responsibility.
AI agents can perform many of those entry-level tasks.
Research.
Drafting.
Testing.
Reconciliation.
Documentation.
Basic analysis.
If organizations automate the bottom of the career ladder, where will future experts come from?
This is not only an employment problem.
It is a capability-continuity problem.
A company may reduce junior headcount today and discover later that it has no pipeline of experienced talent.
The senior expert cannot remain senior forever.
The agent may produce outputs without developing human judgment.
The enterprise needs a new apprenticeship model.
Junior professionals may need to learn through:
-
Reviewing agent output
-
Investigating failures
-
Running controlled experiments
-
Working with customer context
-
Participating in decisions
-
Building and governing workflows
-
Rotating across domains
-
Receiving deliberate mentoring
The learning path cannot depend only on the repetitive production tasks AI removes.
This issue connects directly to the human compact discussed in From Loyalty to Leverage.
If companies capture agent productivity while abandoning human development, they weaken both careers and their own future capabilities.
Agent Performance Cannot Be Measured Like Employee Performance
Employees are evaluated through a mixture of:
-
Results
-
Behaviour
-
Collaboration
-
Judgment
-
Growth
-
Leadership
Agents require different measures.
Possible metrics include:
-
Accuracy
-
Completion rate
-
Escalation quality
-
Error severity
-
Cost per outcome
-
Latency
-
Human-review burden
-
Rework
-
Policy compliance
-
Customer impact
-
Drift over time
But local metrics can produce dangerous behaviour.
An agent optimized for response time may reduce quality.
An agent optimized for collections may damage customer trust.
An agent optimized for productivity may create excessive downstream review.
An agent optimized for ticket closure may hide unresolved problems.
Performance must therefore connect to the wider outcome.
The agent should not merely complete tasks.
It should contribute safely to something the organization values.
Agents Will Need Probation, Promotion, and Retirement
The language sounds human, but the concepts are useful.
Probation
A new agent should begin in a controlled environment.
Limited access.
Low-risk cases.
Enhanced monitoring.
Comparison against human performance.
Promotion
As reliability is demonstrated, the agent may receive:
-
Broader case coverage
-
Additional tools
-
Greater autonomy
-
Higher-value workflows
Demotion
If performance declines, permissions should narrow.
Human review should increase.
Retirement
Agents should be removed when:
-
The business need disappears
-
A stronger workflow replaces them
-
The model becomes unsupported
-
Risk becomes unacceptable
-
Ownership becomes unclear
-
Performance no longer justifies cost
The enterprise should not accumulate immortal agents.
Every productive actor should have a lifecycle.
The Economics of Work Will Change
Traditional service economics are based largely on human effort.
Employees receive salaries.
Providers bill hours.
Outsourcing contracts price roles and capacity.
Agents create a different cost structure.
The company may pay for:
-
Model usage
-
Infrastructure
-
Tool calls
-
Data access
-
Monitoring
-
Human oversight
-
Verification
-
Workflow maintenance
-
Vendor subscriptions
The marginal cost of additional machine work may be low.
The cost of error may be high.
This changes how leaders should think about productivity.
The relevant question is not:
“How many human hours did we save?”
It is:
“What is the total cost of producing a reliable, accepted outcome?”
That includes machine consumption, human judgment, governance, correction, and risk.
As argued in The New Labor Arbitrage Is Not Geography. It Is Orchestration., the advantage will come from designing the complete system—not from optimizing one rate or resource.
The Human-to-Agent Ratio Will Be a Weak Metric by Itself
Boards and investors may eventually ask:
How many agents does each employee manage?
A high ratio could signal leverage.
It could also signal chaos.
One person governing one hundred simple, reliable agents may be reasonable.
One person nominally responsible for ten high-risk autonomous agents may be reckless.
The ratio must be interpreted alongside:
-
Agent complexity
-
Authority
-
Risk
-
Verification
-
Exception rate
-
System reach
-
Human expertise
-
Monitoring quality
The goal is not maximum agent count.
It is maximum safe and meaningful leverage.
The Enterprise May Become Faster—and More Fragile
Agents can operate continuously.
They can act in parallel.
They can coordinate across systems.
They can reduce waiting.
This could dramatically improve execution speed.
But speed creates fragility when errors propagate faster than humans can understand them.
An agent may create an action.
Another agent may respond.
A third may interpret the resulting data.
A small initial mistake can spread across the execution graph.
The organization needs circuit breakers.
These may include:
-
Transaction limits
-
Rate limits
-
Confidence thresholds
-
Human escalation
-
Simulation environments
-
Reversible actions
-
Kill switches
-
Independent monitoring agents
-
Separation between recommendation and execution
The agentic company should not assume that more autonomy always creates more value.
Autonomy should increase only where the system can contain failure.
Agents Can Create Activity Without Creating Progress
AI makes output easy.
Reports.
Emails.
Code.
Tasks.
Recommendations.
Analyses.
The organization may become overwhelmed by machine-generated activity.
A hundred agents can create an enormous amount of visible work.
This may produce the illusion of momentum.
The same danger exists with human organizations, but machines can amplify it dramatically.
The enterprise must protect itself from synthetic busyness.
Every agent should connect to:
-
A real outcome
-
A decision
-
A user need
-
A measurable operational state
If an agent generates outputs nobody uses, it is not productive.
If it creates tasks faster than teams can absorb them, it may reduce performance.
If it produces analysis without decision authority, it enlarges the queue.
The company should measure outcome flow, not agent activity.
Human Accountability Must Be Designed Before Automation
Organizations often automate first and ask governance questions later.
A team identifies a repetitive workflow.
Builds an agent.
Shows impressive results.
Then risk, legal, security, and operations become involved.
The company tries to attach accountability after the workflow already exists.
This creates conflict and rework.
The correct sequence is:
-
Define the outcome.
-
Identify the institutional owner.
-
Determine acceptable machine autonomy.
-
Define human decision points.
-
Establish access and data boundaries.
-
Design verification.
-
Build the agent.
-
Test under controlled conditions.
-
Expand authority gradually.
Accountability should shape automation.
It should not be added as an apology afterward.
The Board Will Need an Agent Governance View
Boards currently review:
-
Workforce
-
Technology
-
Cybersecurity
-
Risk
-
Compliance
-
Major vendors
Agents cut across all of them.
Boards should eventually understand:
-
How many production agents operate
-
Which strategic outcomes depend on them
-
Which have high decision authority
-
Which access sensitive data
-
Who owns them
-
How they are verified
-
Which third-party models are involved
-
How failures are contained
-
How agent decisions affect customers and employees
-
How the company preserves human accountability
This does not mean the board should review prompts or technical details.
It must understand the institutional exposure.
An unmanaged machine workforce is a governance risk.
Regulation Will Ask Who Was Responsible
When an automated decision harms someone, saying “the agent did it” will not be sufficient.
Customers, regulators, employees, and courts will ask:
Who authorized the system?
Who defined the objective?
Who approved the data?
Who monitored performance?
Who could have intervened?
Who benefited?
Who is responsible for correction?
The enterprise must be able to answer.
This is why agent governance cannot be delegated entirely to technology teams.
It is an institutional design issue.
Culture Will Shape Agent Behaviour
AI agents do not possess culture in the human sense.
But they operate according to objectives, examples, policies, data, and feedback produced by the organization.
If the company rewards aggressive sales behaviour, its agents may reflect that pressure.
If leaders tolerate weak evidence, agents may generate confident but poorly supported recommendations.
If the organization prioritizes speed above safety, agent workflows may encode that value.
Agents can amplify culture because they operationalize it.
The company must therefore ask:
What behaviours are we teaching machines to repeat?
Agent design is cultural design.
The Agentic Enterprise Needs a Constitution
Policies alone may not be enough.
The organization needs a clear set of principles governing machine participation.
An agent constitution might include:
Human accountability
Every consequential agent has a named human or institutional owner.
Minimum necessary authority
Agents receive only the access and autonomy required.
Visible identity
Every action is attributable to a specific agent and version.
Evidence
Consequential outputs must retain the supporting information and decision path.
Escalation
Agents stop when confidence, risk, or ambiguity crosses defined boundaries.
Reversibility
High-impact actions should be reversible wherever possible.
Independent verification
Production and verification should not rely entirely on the same reasoning chain.
Right to human review
People significantly affected by an automated decision should have an appropriate route to human consideration.
Lifecycle governance
Agents must be reviewed, updated, suspended, and retired.
Human development
Automation should not eliminate the organization’s ability to develop future expertise.
This constitution gives the enterprise a stable framework while specific technologies change.
The Virtual Delivery Center Can Govern Mixed Execution
A Virtual Delivery Center can become a useful operating container for a world in which humans and agents work together.
A VDC is organized around an ongoing execution mandate.
For example:
-
AI modernization
-
Customer implementation
-
Product engineering
-
Compliance operations
-
Supply-chain intelligence
-
Growth execution
Inside the VDC, the organization can combine:
-
Internal outcome owners
-
Employees
-
External specialists
-
Delivery pods
-
AI agents
-
SaaS platforms
-
Verification systems
-
Governance controls
The VDC provides persistence.
The agent and human composition can change.
A coding agent may be replaced.
A specialist may join temporarily.
A verification workflow may be strengthened.
The customer’s systems, context, access rules, and governance remain connected to the VDC.
This creates a governed environment for mixed execution.
The company does not need to treat each agent as an isolated tool or each specialist as a disconnected contractor.
They participate inside a defined execution boundary.
The VDC Must Include an Agent Registry
Every VDC containing agents should know:
-
Which agents are active
-
Which outcomes they support
-
Who owns them
-
What they can access
-
What they can do
-
Which humans review them
-
Which models they use
-
What they cost
-
How they perform
-
When they were last evaluated
This registry becomes part of the execution infrastructure.
It allows the company to reconfigure capability without losing visibility.
Humans Should Own Outcomes. Agents Should Own Bounded Tasks.
This is a useful default principle.
A human or institution owns the outcome.
Agents own clearly bounded tasks inside the outcome.
For example:
Outcome: Successfully onboard the enterprise customer.
Human ownership may include:
-
Relationship
-
Judgment
-
Exception decisions
-
Commercial accountability
-
Final acceptance
Agent tasks may include:
-
Document validation
-
Data mapping
-
Status monitoring
-
Training-content preparation
-
Issue classification
The exact boundary will vary.
But assigning complete institutional outcomes to agents creates accountability problems.
The agent can perform.
The human must remain responsible for what becomes true.
The Enterprise Should Automate Work, Not Abandon Responsibility
There is a temptation to use AI as a distancing mechanism.
The algorithm declined the application.
The model set the price.
The agent sent the message.
The system escalated the employee.
This language hides human choices.
Someone selected the model.
Defined the objective.
Approved the policy.
Accepted the risk.
Automation changes how the decision is produced.
It does not erase institutional responsibility.
The most trustworthy organizations will not hide behind their agents.
They will explain where machines participated and where humans remain accountable.
The Best Agentic Company May Employ More Humans in Different Work
AI does not lead to one inevitable workforce outcome.
A company may use agents to reduce employment.
Another may use them to serve ten times more customers.
Another may create entirely new products.
Another may improve quality and compliance.
Another may shorten the workweek.
The technology creates leverage.
Strategy determines how the leverage is used.
The company may employ fewer people in repetitive production while hiring more:
-
Domain experts
-
Customer leaders
-
Agent architects
-
Verification specialists
-
Security professionals
-
Workflow designers
-
Mentors
-
Exception handlers
-
Governance leaders
The shape of human work changes.
The objective should not be minimizing humans.
It should be maximizing meaningful human contribution.
People Need to Know When They Are Working With an Agent
Transparency will matter.
Employees should know when agent output influences:
-
Performance assessment
-
Work allocation
-
Promotion
-
Monitoring
-
Hiring
-
Termination
Customers should know when agents materially participate in:
-
Advice
-
Support
-
Pricing
-
Eligibility
-
Communication
-
Financial decisions
Not every automated action requires a warning label.
But people should not be deceived about the source of consequential decisions.
Trust depends on appropriate disclosure.
Humans Will Become the Scarce Accountability Layer
When machine production becomes abundant, accountability becomes more valuable.
The person willing and able to say:
“I understand the system.”
“I accept responsibility for this decision.”
“I will explain it to the customer.”
“I will intervene when the model is wrong.”
“I will carry the consequence.”
That person becomes essential.
AI can produce many recommendations.
The enterprise still needs someone to choose.
This is another expression of the argument in Execution Is the New Scarcity.
Intelligence may become abundant.
Responsible conversion of intelligence into outcomes remains scarce.
Leadership Becomes Constitutional Design
The leader of a traditional organization allocates people, budget, and authority.
The leader of an agentic organization must also define the constitution under which machine actors operate.
They decide:
-
Which outcomes may be automated
-
Which decisions remain human
-
What level of evidence is required
-
What access is acceptable
-
What risks can be tolerated
-
Which failures trigger suspension
-
How productivity gains are shared
-
How human capability continues to develop
This is not a technical responsibility alone.
It is leadership.
What Leaders Should Do Now
Inventory every production agent
Do not assume the technology function already knows.
Include embedded platform agents, team-built workflows, and external-provider agents.
Name an owner
Every agent must belong to a business outcome and a responsible leader.
Define the authority boundary
Specify what the agent may recommend, create, modify, approve, communicate, and execute.
Design human escalation
Identify the conditions under which the agent must stop.
Separate production from verification
Do not rely only on the same agent or model to validate its own work.
Measure outcome contribution
Track whether the agent improves a meaningful result—not merely activity or output.
Control lifecycle
Review, update, limit, suspend, and retire agents deliberately.
Protect the apprenticeship pipeline
Create new ways for junior professionals to develop judgment.
Make access granular
Avoid permanent, broad credentials and shared identities.
Prepare the board
Agent governance belongs within enterprise governance.
Questions Every Board Should Ask
-
Which critical outcomes already depend on AI agents?
-
Could we identify every agent with production access?
-
Does each agent have an accountable owner?
-
What is the most consequential action an agent can take without human approval?
-
How are agent failures detected?
-
Can an agent be suspended immediately?
-
Are employees or customers affected by decisions they cannot challenge?
-
Are we reducing junior roles without rebuilding professional development?
-
Are productivity gains improving value—or merely increasing workload?
-
Do we understand our dependence on third-party models and platforms?
The Company After Humans Are Outnumbered
The company will still be human.
It will be founded by people.
Owned by people.
Governed by people.
Serving people.
A machine may perform most of the tasks.
That does not make the enterprise a machine institution.
The purpose, values, consequences, and responsibility remain human.
This is the distinction the next generation of leaders must protect.
The company may operate ten agents for every employee.
Or one hundred.
The number matters less than the architecture.
Are the agents visible?
Governed?
Limited?
Verified?
Connected to real outcomes?
Do humans retain authority where judgment and consequence matter?
Does the system develop people, or merely extract more output from them?
Does machine leverage strengthen the enterprise’s responsibility—or allow it to hide?
The Future Org Chart Will Have Two Layers
One layer will show humans.
Leadership.
Employees.
Institutional accountability.
The second will show the active execution system.
Agents.
Platforms.
Specialists.
Data.
Controls.
Decisions.
Verification.
The first answers:
Who is responsible?
The second answers:
How does the work happen?
An organization that sees only the first will misunderstand its productive capacity.
An organization that sees only the second will lose sight of human accountability.
The two must remain connected.
When the Agents Outnumber Us
The moment may arrive quietly.
No announcement.
No official milestone.
One day, the company will simply have more active agents than employees.
Some will be small.
Some powerful.
Some visible.
Others embedded inside platforms and processes.
The company may not even know the exact number.
That is the danger.
The defining transition is not when the count crosses one.
It is when machine actors become essential to execution while the organization still governs itself as though only humans perform work.
The enterprise must redesign before that moment.
Not because agents are inherently untrustworthy.
Because capability without governance becomes risk.
Intelligence without accountability becomes danger.
Speed without verification becomes fragility.
Automation without human development becomes institutional decline.
AI agents will outnumber humans in many organizations because machine capability can multiply faster than employment.
But the company should never confuse numerical dominance with institutional authority.
Agents may become the majority of productive actors.
Humans must remain the authors of purpose, the designers of boundaries, and the bearers of consequence.
The machines may do most of the work.
The responsibility will still be ours.