The Human-Agent Organization: redesigning work, teams, management and accountability

The first era of workplace AI was about access. Employees got chatbots, copilots, writing assistants, coding assistants, search tools, meeting summaries. The central question was: how can an employee use AI to perform an existing job more efficiently?

The next era asks something different.

AI systems are beginning to pursue goals over time, operate software, retrieve and interpret enterprise context, coordinate subtasks, communicate with other agents, create and modify business artifacts, monitor conditions, recommend decisions, and execute bounded actions.

That changes the organizational question. It is no longer how should people use AI? It becomes:

How should an organization be designed when productive work is performed by a mixed population of humans, agents, deterministic systems, external services, and machines?

That is the subject of the human-agent organization — and it is not a normal company with more software.

The words deliberately designed are the load-bearing ones, because a company becomes a human-agent organization badly, by accident, with no decision ever being taken. Employees adopt tools. Teams build agents. Agents connect to systems. Management layers automate reporting. Algorithms begin evaluating performance. And machine decisions start shaping who receives work, whose forecast is accepted, which employee is promoted, which customer gets attention, which claim is rejected, which product is prioritized.

The organization may already be reorganizing itself before leadership has acknowledged that anything changed.

Satya Nadella imagines a future organization containing tens of thousands of people and millions of agents operating in a shared loop. The number is not the point. The organizational possibility is: a company may soon have far more active machine actors than human employees — while its org chart continues to show only people.

The goal is not the maximum possible number of agents. It is the strongest possible organization.

It is Monday morning in 2031. A regional commercial director opens her organizational workspace. She formally manages twelve human leaders. Her operating environment also displays 846 active agents associated with her organization.

The director does not review 846 conversations. She sees three decisions requiring her authority, two patterns suggesting a strategic problem, one agent population whose error rate has increased, a warning that review capacity is becoming overloaded, and an alert that an agent attempted to access data outside its permitted customer scope.

Her job has changed. She is not performing the analysis, assigning every task, or reading every report. She is responsible for defining priorities, allocating authority, deciding which exceptions matter, resolving conflicts among objectives, maintaining the quality of both workforces, keeping people capable, protecting trust, and accepting accountability for outcomes.

She manages an organization. The organization is no longer composed only of people.

The same company, without the design

Now remove the organizational design and keep everything else. The director has 846 agents and:

• no complete inventory and no consistent naming
• no clear ownership, and overlapping responsibilities
• several agents using the same data differently
• several managers unaware that agents are acting on their behalf
• employees approving recommendations they do not understand
• no process for retiring agents
• no view of the total cost
• no way to reconstruct a disputed decision

The agents are individually useful. Collectively, they have created organizational disorder.

That is the entire difference between deploying agents and building a human-agent organization.

It is an organization with mixed agency

Agency means the capacity to initiate or direct action toward an objective. In a human-agent organization, three forms of it coexist.

The five defining characteristics

1. 01Agents participate in real workflows. They do more than produce isolated content.
2. 02Agents have bounded operational identities. Their purpose, permissions, owners and tools are known.
3. 03Humans and agents coordinate over time. Work includes handoffs, shared state, review, escalation and learning.
4. 04The organization redesigns roles and authority. It does not merely add AI to existing jobs.
5. 05Outcomes improve both machine and human capability. The organization learns through use.

A traditional tool waits. A spreadsheet does not decide which analysis should be performed; an email application does not initiate a commercial negotiation. An agent can pursue: monitor a condition, decide an event needs attention, retrieve information, call tools, create an artifact, request another agent’s work, wait, resume, escalate.

This does not make the agent conscious. It makes it operationally active — and that is an organizational fact, not a philosophical one.

Delegation changes the psychological contract

When a human uses a tool, the human feels like the performer. When a human delegates, they become principal, manager, reviewer, customer, supervisor. That shift changes attention, trust, responsibility, learning, and identity — all at once, usually without anyone naming it.

Human-agent interaction should not be described on a single scale from “human” to “automated”. Human involvement and agent involvement can bothbe high. Microsoft’s 2026 Work Trend Index describes four broad modes based on the intensity of each — asking, exploration, collaboration, and delegation — and stresses that advanced users select the appropriate mode rather than indiscriminately maximizing agent use.

Enterprise work adds two more.

The correct mode depends on ambiguity, risk, reversibility, expertise, frequency, evidence, authority, and learning value. A strategic decision may deserve high human and high agent involvement. A low-risk administrative task may justify high agent and low human involvement. A safety-critical release may require high human authority even when agents perform most of the investigation.

Traditional organizations design jobs, reporting lines, and departments. AI forces a more granular analysis, because a job contains outcomes, decisions, tasks, relationships, responsibilities, and learning opportunities — and these should not be delegated as one indivisible bundle.

The organization must decide which actor performs each element of work. This is not a one-time automation decision. It is a continuing allocation problem, and it should be revisited as capability moves.

Work best performed jointly

The ten allocation criteria

Effective teams require common ground, shared goals, role clarity, communication, awareness, trust, conflict resolution, and coordination. None of those requirements disappear when one member is an agent.

Microsoft Research’s work on collaboration readiness argues that task success alone is insufficient for evaluating AI teammates — communication efficiency, common ground, workspace awareness, and coordination quality all matter too.

Managing an agent does not require a management title. A professional may assign objectives, evaluate output, correct behaviour, decide what to delegate, supervise execution, and accept responsibility. Microsoft describes advanced AI users increasingly acting as agent bosses — setting direction, applying judgment, owning outcomes.

And agent management is not prompt engineering. Prompting is one interaction skill. Management requires goal definition, context design, work decomposition, delegation, evaluation, authority setting, exception handling, cost control, and learning design.

set intent → allocate work → observe → intervene → evaluate → learn

Span of control becomes span of cognition

A manager can supervise far more productive activity through agents. But the relevant metric was never the number of agents. It is the number and complexity of exceptions, objectives, and accountabilities a manager can genuinely hold.

AI can absorb real managerial activity — information consolidation, scheduling, performance tracking, reporting, task routing. OECD research finds algorithmic-management tools already used to support or automate managerial work, while managers themselves report concerns about explainability, accountability, and worker health.

So does middle management disappear? Some of it shrinks. But:

Less management as information transport. More management as institutional judgment.

A traditional org chart shows positions, hierarchy, and reporting lines. The human-agent organization needs several more maps alongside it — and most companies maintain none of them.

Human roles change before job titles do. The same title can conceal a very different job — a demand planner moving from manual forecast adjustment, spreadsheet reconciliation and data retrieval toward exception prioritization, assumption challenge, commercial interpretation, agent supervision, and decision ownership.

Agents, in turn, may act as analyst, monitor, investigator, coordinator, drafter, tester, scheduler, verifier, tutor, memory, simulator, or operator. An agent role should describe a capability, not imitate an employee title for the sake of it.

People build expertise by observing, practising, receiving feedback, making bounded mistakes, encountering exceptions, developing pattern recognition, and assuming responsibility. Automation changes access to every one of those experiences.

Recent economic research has begun modelling how automation affects not only current productivity but learning-by-doing, the development of expertise, and the transition from worker to manager. The organizational consequences depend heavily on which tasks remain available for skill formation.

Which produces the sharpest failure mode in this whole essay: the reviewer without mastery. A junior employee asked to review an agent’s work without having developed the skills to detect subtle errors. The result is false confidence, ceremonial oversight, and weakened accountability.

Microsoft’s 2026 research reports that advanced AI users are more likely than others to intentionally perform some work without AI to preserve their skills, and to pause before deciding whether a human or an agent should do a piece of work. That should become an organizational practice, not merely an individual habit.

The human does not need to repeat every machine step. The human needs to remain capable of responsible intervention — which means being able to answer eight questions:

1. 01What objective did the agent pursue?
2. 02Which evidence mattered?
3. 03Which assumptions drove the result?
4. 04What uncertainty remains?
5. 05Which tools were used?
6. 06What could make the answer wrong?
7. 07Which decision is mine?
8. 08What did I learn?

Why reviewing at the end is not enough

By the time an agent presents a final output, many intermediate decisions are already embedded inside the artifact. Microsoft Research’s spreadsheet-agent study found that active participation during execution helped users detect errors and understand the task in ways that post-hoc review did not.

A user may feel confident in an agent because the output is fluent, the interface is polished, and the answer was immediate. That is not calibrated trust. Trust should be grounded in proven performance, visibility, predictability, appropriate uncertainty, controllability, aligned incentives, and accountability.

Calibrated trust means humans understand where the system is strong, where it is weak, when to review, when to rely, and when to stop. And crucially, trust is task specific. A person may trust an agent to summarize a meeting but not to evaluate an employee, release a payment, or interpret a safety incident. Trust should not transfer automatically across tasks.

Culture lives in decisions, stories, incentives, exceptions, rituals and informal norms — and agents will increasingly participate in all of them. They can reinforce customer centricity, safety discipline, documentation, financial rigour and learning. They can equally reinforce bureaucracy, bias, short-termism, deference and surveillance.

When policies, examples and prior decisions are built into agents, organizational culture becomes partially executable. Which raises an uncomfortable question:

Which culture are we encoding — the one we claim to have, or the one revealed by our past behaviour?

An agent trained on historical decisions can preserve outdated assumptions, old power structures, discriminatory patterns, and the exact practices the company is currently trying to change. Call it cultural fossilization.

There is a subtler effect too. New employees increasingly learn the company through agents. The onboarding agent explains how decisions get made, what quality means, who holds knowledge, and which behaviours are rewarded. That gives agents significant influence over organizational identity.

A salesperson defines the customer strategy, delegates analysis to agents, uses an agent-generated negotiation package, and closes the deal. Who created the value? The salesperson, the agent designer, the data team, the model provider, or the prior experts whose work trained the system?

The answer is collective — which means performance systems built around individual output become less accurate. Organizations will need to distinguish outcome ownership, execution contribution, capability contribution, and learning contribution.

Watch the incentives encoded in every metric. An agent optimized for case closure will close cases prematurely. An agent optimized for employee productivity will increase work intensity. Neither requires bad intent — only an unexamined objective.

And AI shifts the relative value of execution, expertise, judgment, relationships, ownership, and orchestration. Whether the resulting gains flow into wages, bonuses, reduced hours, employment growth, or shareholder returns is an organizational and political choice — not a technical outcome.

Hiring for current tasks becomes dangerous when a role can change rapidly as agents absorb parts of it. Hiring should target durable responsibility, learning capacity, domain judgment, collaboration, adaptability, and agent-management potential.

AI can support candidate search, screening, interview preparation and assessment. But employment decisions affect rights and opportunities, so the organization has to guard against biased data, invalid proxies, inaccessible assessments, opaque decisions, and the automation of historical discrimination. The ILO warns that AI in human-resource management can be undermined by flawed objectives, biased data, and opaque system design.

Algorithmic management uses software to support or automate managerial functions: task allocation, monitoring, scheduling, evaluation, discipline, compensation. The OECD and ILO both document its spread beyond platform work into conventional workplaces.

That is substantially more powerful, and the risk scales with it.

Some things should never be inferred casually: emotional state, loyalty, personality, future performance, intent, health, political belief. Data availability does not make inference legitimate.

Underneath all of it sits an asymmetry. The organization may know everything about the worker; the worker may know little about what is measured, how it is interpreted, how decisions are made, or how to appeal. That asymmetry corrodes dignity and trust.

OECD experimental research suggests consultation among workers, management and worker representatives can preserve productivity gains while improving perceived job quality — covering purpose, collected data, decision use, appeal, staffing, training, workload, and health.

Human agency is not merely human approval. A worker has agency when they can understand, influence, challenge, develop, refuse inappropriate delegation, appeal decisions, and shape their work.

ILO case studies emphasize that social dialogue can support AI deployment that protects autonomy, develops skills, and maintains labour and social protections.

Two rights follow. The right to explanation: for significant employment decisions, workers need more than “the algorithm produced this score” — they need understandable reasons and meaningful human review. And the right to appeal: an appeal must reach a qualified human, who has authority to change the outcome, access to the relevant evidence, and no incentive to retaliate.

Under the EU AI Act, certain uses of AI in employment and worker management fall within high-risk areas, carrying obligations across risk management, data quality, documentation, traceability, transparency, human oversight, accuracy, cybersecurity and monitoring. Current EU guidance also notes that affected workers and representatives must be informed before certain high-risk workplace systems are deployed. Applicability depends on the specific system, role and legal timeline, and qualified legal review remains necessary.

The EU Platform Work Directive adds protections around transparency of automated monitoring, restrictions on certain personal-data processing, human monitoring, review of significant decisions, and reasons for materially adverse decisions. Those rules target platform work, but they signal the direction of policy concern about algorithmic management generally.

And human oversight has to be real. It does not mean automatic confirmation, an unavailable reviewer, a reviewer without authority, or a reviewer without information.

Productivity is not the only outcome available. Agents can reduce cost, increase output, improve quality, create new products, expand employee capability, improve inclusion, and reduce risk — and which of those a company gets is largely a design choice.

The evidence is genuinely heterogeneous. A major customer-support study reported an average productivity increase of 15%, with larger gains among less-experienced workers and evidence of learning and improved work experience. Other settings show smaller gains, uneven gains, quality trade-offs, limited usage, and organizational barriers.

Headcount is the wrong design variable. The question is what human capability and institutional depth are required to operate the desired level of machine agency responsibly.

On displacement: job exposure does not translate mechanically into job elimination. The ILO’s research emphasizes transformation, with clerical occupations remaining highly exposed while exposure expands into professional and technical roles. But transformation can still mean fewer entry-level roles, job consolidation, wage pressure, and higher performance expectations.

On inclusion, the direction is not predetermined either. Recent research found an AI-based communication system improved productivity, customer ratings, labour supply and retention for deaf and hard-of-hearing workers in the studied setting. The same organization can broaden participation — or exclude workers through inaccessible interfaces and biased evaluation.

OpenAI’s current agent-building guidance recommends explicit human intervention for high-risk or irreversible actions, and whenever failure thresholds are exceeded. That is an architecture requirement, not a policy sentence.

One agent’s risk is not the sum of many agents’ risk. Networks introduce propagation, amplification, coordination failure, trust manipulation, and emergent behaviour. Microsoft’s 2026 research on agent networks observed malicious propagation across agents, amplification of false claims, capture of verification processes, and real difficulty tracing information moving through chains of agents.

The controls are unglamorous and they work: delegation-depth limits, source provenance, independent verification, communication monitoring, network segmentation, identity, rate limits, tool restrictions, and human escalation.

An agent may perform well on each local step and still degrade the total artifact over time. Microsoft Research’s 2026 delegation benchmark found significant content degradation across long document workflows, while noting that production systems can mitigate this through verification, orchestration, and domain-specific tooling.

The mechanisms that make long-horizon delegation survivable are checkpoints, tests, immutable source artifacts, diffs, independent validators, bounded delegation, human review, rollback, and artifact versioning. And authority itself should behave differently over time:

The longer an agent operates, the more conditions change. Authority should expire, be refreshed, and narrow as uncertainty rises.

The principal-agent-agent problem

Classical organizations already face principal-agent problems: shareholders delegate to executives, executives to managers, managers to employees. The agentic organization adds layers — humans delegate to machine agents, and machine agents may delegate to other agents. Every delegation creates information asymmetry, goal misalignment, monitoring cost, and responsibility ambiguity.

Separation of duties still applies. An agent that recommends a payment should not also approve it, execute it, and reconcile it. And incident management needs to cover unauthorized action, incorrect decisions, data exposure, model failure, coordination failure, excessive cost, human overreliance, and network propagation.

A global FMCG company runs a large national grocery account with a team of nine: key account director, two account managers, category manager, trade marketer, demand planner, supply manager, finance partner, and retail-media manager. The customer generates €240 million in annual net revenue, 900 promotional events, 18,000 stores and digital endpoints, and thousands of weekly exceptions.

Today the team spends its time collecting data, preparing reports, reconciling forecasts, monitoring promotions, documenting meetings, chasing actions, and resolving deductions — and correspondingly less time on customer strategy, negotiation, category value, relationships, and difficult decisions.

The redesign

The company builds an Account Orchestration System.

The Monday workflow

Overnight the agents identify 73 availability exceptions, 18 promotion deviations, six deduction mismatches, four commitments approaching deadline, and two material commercial risks. They resolve 54 low-risk exceptions within policy and prepare 14 recommended actions. Five cases require human decisions. Three of them are instructive.

Six months later

Report-preparation time falls. Low-risk exceptions close faster. Account contribution improves. Human meeting time shifts toward decisions. Agent recommendations improve.

And a new problem appears: junior account managers are doing less data investigation, and they struggle to challenge the agents.

So the company introduces weekly case reviews, unassisted scenario exercises, a requirement to explain assumptions, rotation through exception handling, and sampled reconstruction of agent decisions. Only then does the redesign improve both machine capability and human capability.

human-agent value =

    business outcome
  + human capability growth
  + institutional learning
  − agent cost
  − human burden
  − risk
  − capability atrophy

Its purpose is not precision. Its purpose is to make one-dimensional optimization visibly incomplete.

1. 01How many agents operate in the company?
2. 02Which agents can take real action?
3. 03Who is accountable for each agent population?
4. 04Which decisions are being delegated?
5. 05Which responsibilities remain non-delegable?
6. 06Are employees becoming more capable or more dependent?
7. 07What happens to entry-level development?
8. 08How do managers supervise agent work?
9. 09Where is approval fatigue emerging?
10. 10How is worker performance affected by algorithmic management?
11. 11Can employees appeal AI-supported decisions?
12. 12How are productivity gains distributed?
13. 13Which human relationships must be preserved?
14. 14What organizational knowledge is being captured?
15. 15Which agents communicate with other agents?
16. 16What network risks exist?
17. 17Can the company operate without its agents?
18. 18How does the organization measure cognitive coverage?
19. 19Which cultural assumptions are encoded in the systems?
20. 20What kind of institution are we becoming?

Is an agent an employee? Is it just software?

Neither. An agent may perform employee-like activities but has no human rights, moral responsibility, or legitimate authority of its own. It is software — but agentic systems pursue objectives, select tools, maintain state and act over time, so they need stronger operational governance than conventional applications.

Will every employee manage agents? Will managers disappear?

Many employees will delegate to and supervise agents without becoming formal people managers. Administrative management may shrink, while judgment, integration, coaching, conflict resolution and accountability become more important.

What is span of cognition?

The amount of objectives, exceptions, consequences and delegated activity a human can understand and supervise responsibly.

Should companies maximize delegation?

No. Delegation should reflect risk, reversibility, learning needs, evidence, and human authority.

Why is apprenticeship at risk, and how is it preserved?

Agents automate exactly the junior tasks through which future experts traditionally develop judgment. It is preserved through agent-assisted practice, simulations, deliberately unassisted work, trajectory review, counteranalysis, and exception ownership.

Can agents evaluate employees?

They may support evidence organization under strict controls. Material employment decisions require careful legal, ethical and human review.

How should agents appear on the organization chart?

They should not be dropped into human legal reporting lines. Companies need a complementary agent-estate and delegation map alongside the human chart.

Do human-agent teams always outperform humans?

No. Performance depends on the task, the system, expertise, workflow, evaluation and coordination. Some studies find larger productivity gains among less-experienced workers in specific settings, but that pattern should not be generalized to every task or occupation.

Who is accountable when an agent makes a mistake?

The relevant human owners and the organization, according to their legal and operational responsibilities. This does not change because the action was machine-executed.

What is the first implementation step?

Map one important workflow as a governed outcome trajectory, and identify its human roles, agent roles, authority, learning, and worker impact.

The industrial organization combined people, machines, capital and hierarchy. The knowledge organization added software, data and digital communication. The human-agent organization adds something categorically different: machine systems capable of participating actively in the pursuit of organizational objectives — monitoring, investigating, planning, communicating, creating, executing, coordinating, and learning within bounded systems.

This creates extraordinary productive possibility. A small team can pursue ambitions once reserved for large institutions. A large company can distribute expertise across its entire workforce. People can move away from repetitive administration toward judgment, creativity, relationships, responsibility and meaning.

None of it is guaranteed. The same systems can produce surveillance, deskilling, dependency, concentrated power, weakened careers, accountability gaps, social isolation, and institutional fragility.

The difference will be organizational design.

The defining question is not how intelligent the agents are. It is how intelligently the organization has arranged the relationship among its humans, agents, systems, authority, and learning.

The human-agent organization should not be imagined as a company where machines do the work and humans watch. Nor as a company where humans remain unchanged while agents quietly assist them. It is a new institution. Its people will need new capabilities, its managers new disciplines, its career ladders new foundations, its systems new controls — and its culture will have to answer new questions about agency, dignity, trust, learning, power and responsibility.

machine execution   with   human authority
machine scale       with   human meaning
machine memory      with   human judgment
machine learning    with   human development

The company must preserve the distinction. Agents may perform the trajectory. Humans must remain able to choose the destination, change the course, understand the consequence, and accept the responsibility.

The future organization may contain millions of agents. Its greatness will not be measured by how few humans it needs. It will be measured by what its humans are enabled to become.

The source conversation. Grounded in the Satya Nadella–Reid Hoffman discussion, especially its exploration of organizations containing large populations of humans and agents, the artifacts and tacit knowledge produced through their interaction, macro-delegation and micro-steering, cognitive coverage, and agent governance.

Human agency and work modes.Microsoft’s 2026 Work Trend Index presents a framework for asking, exploring, collaborating and delegating with agents, while emphasizing human direction, judgment, responsibility, and the intentional preservation of skills.

Agentic work. Current OpenAI research describes agentic AI shifting knowledge work from short interactions toward longer delegated tasks involving tool use and environmental interaction, with practical guidance emphasizing human intervention for high-risk actions and repeated failure.

Human-agent collaboration. Microsoft Research is developing collaboration-readiness frameworks that evaluate common ground, communication, coordination quality and workspace awareness beyond simple task success.

Long-horizon reliability. Recent delegation research shows that strong model performance on short tasks does not guarantee reliable execution across long, multi-step workflows — reinforcing the need for checkpoints, verification, versioning and human intervention.

Productivity and learning. The customer-support study by Brynjolfsson, Li and Raymond found an average productivity gain of 15%, with larger gains among less-experienced workers and evidence of employee learning in the studied setting.

Work transformation. ILO evidence emphasizes that AI exposure is more likely to transform mixtures of tasks than to eliminate every exposed occupation automatically, while noting significant differences across occupations, countries and worker groups.

Algorithmic management. OECD and ILO research documents the adoption of systems that allocate, monitor, supervise and evaluate work, alongside concerns about explainability, accountability, privacy, health, autonomy and job quality.

Worker consultation. OECD experimental work and ILO case studies indicate that consultation and social dialogue can help align productivity improvements with employee agency and job quality.

Governance and human oversight.NIST’s AI Risk Management Framework treats human oversight and organizational governance as lifecycle responsibilities that should be defined, assessed, documented and monitored.

Employment regulation. The EU AI Act treats certain employment and worker-management uses as high risk, while the Platform Work Directive establishes transparency, data-protection, human-review and explanation requirements in the platform-work context.

Multi-agent risk.Microsoft’s agent-network research identifies propagation, amplification, trust capture and invisibility as risks that emerge when persistent agents interact at scale.