WHEN AI BECOMES THE WORKFORCE. Governing Digital Employees, Responsibility, and the New Organisation
Martin Novak
Synthocracy Institute
Research status: 4 September 2026
Evidence Boundary
This article distinguishes documented developments from analytical interpretation and foresight. [A] Empirical claims refer to current enterprise deployments, product architectures, government guidance, reported organisational experiments, and academic research available by 4 September 2026. [B] Analytical claims develop the Synthocracy Institute’s interpretation of what these developments mean for authority, accountability, organisational design, and work. Sections explicitly marked FORESIGHT describe plausible 2027+ developments rather than established facts.
The language used by the technology industry requires particular care. Workday speaks of a “blended workforce” of people and agents. Salesforce markets “digital labor.” ServiceNow describes AI specialists as “digital employees” capable of completing end-to-end processes. NewCore presents identity infrastructure for a workforce of humans, machines, and AI agents. These are operational and commercial metaphors. They do not establish that AI agents are employees in the legal sense, possess employment rights, have legal personality, or can bear responsibility in the way human workers and corporations can. (workday.com)
Indeed, current government guidance points in the opposite direction. Canada’s agentic-AI guidance says accountability ultimately remains with a designated human owner even when an agent acts autonomously. Australia requires agencies to assign human accountability for decisions and outcomes produced across agentic and multi-agent systems. Singapore’s 2026 governance framework similarly emphasises that humans remain ultimately accountable. (Canada)
That asymmetry is the central problem:
AI agents can increasingly perform work, exercise access, make operational choices, and cause consequences without becoming entities capable of bearing human responsibility for those consequences.
The question is therefore not simply:
Will AI replace workers?
It is:
What happens to organisational authority when work moves to entities that can execute but cannot ultimately answer for what they do?
The employee directory is beginning to contain things that are not employees
Something unusual is happening inside enterprise software.
Workday now offers an Agent System of Record that brings AI agents into the same platform where organisations manage people. It describes the objective as managing a new “blended workforce” and provides agent registration, configuration, activation, deactivation, analytics, identity permissioning, observability, and lifecycle management. Workday explicitly says agents can be managed from registration to retirement “just like you’d manage your people.” (workday.com)
Microsoft Entra Agent ID goes still further into organisational vocabulary.
An AI agent can have:
an owner;
a sponsor;
a manager;
an identity;
an organisational position;
access packages;
lifecycle workflows.
Sponsors provide business accountability for the agent’s purpose and lifecycle. Managers can be designated in the organisational hierarchy and see agents reporting to them. If a sponsor changes roles or leaves the company, lifecycle workflows can transfer sponsorship so that the agent does not become orphaned. (Microsoft Learn)
ServiceNow calls its new systems an Autonomous Workforce and says role-scoped AI specialists can complete end-to-end processes across IT, CRM, HR, legal, finance, procurement, security, and risk. Some operate without human intervention. ServiceNow explicitly describes several of them as digital employees equipped with role-specific skills. (newsroom.servicenow.com)
Salesforce speaks of a potentially “limitless digital labor force.” (Salesforce)
These may sound like marketing metaphors.
But the supporting infrastructure is becoming real.
The enterprise is beginning to create organisational objects for non-human workers.
And that changes the architecture of work.
1. The important transition is from tool to role
A calculator performs a function.
A chatbot answers a question.
A conventional workflow executes predefined steps.
An enterprise agent may instead be assigned something resembling a role.
Resolve first-line IT tickets.
Investigate security incidents.
Qualify sales leads.
Manage invoice disputes.
Process employee requests.
Review suppliers.
Monitor infrastructure.
That is a more consequential abstraction.
A task says:
perform X.
A role says:
remain responsible for this class of work whenever it appears.
Roles imply persistence.
Boundaries.
Access.
Expectations.
Escalation.
Performance evaluation.
Relationships to other roles.
The moment an agent moves from performing isolated tasks to occupying a continuing operational role, traditional questions of software governance begin turning into questions of organisational design.
2. A role requires authority, not merely capability
Suppose an HR agent is capable of:
reading employee files;
answering questions;
updating records;
initiating workflows.
That does not establish which of those functions belong to its legitimate organisational role.
An employee-service role might permit answering leave-policy questions.
It may not permit changing leave entitlements.
A procurement role may gather quotations.
It may not possess supplier-selection authority.
An IT agent may diagnose a system.
It may not be entitled to deploy a fix directly into production.
This gives us a critical distinction:
TECHNICAL CAPABILITY ≠ ORGANISATIONAL ROLE ≠ DECISION AUTHORITY.
Human organisations already separate these concepts.
An employee may know how to perform an action without having authority to perform it.
Agentic organisations must make that separation explicit because technical ability can otherwise become operational authority simply because the tool is available.
3. “Digital employee” can obscure the most important difference
The employee metaphor is useful because it encourages organisations to think about:
identity;
ownership;
access;
onboarding;
performance;
offboarding.
But it can also hide a fundamental difference.
A human employee can be:
trained;
disciplined;
professionally sanctioned;
dismissed;
held contractually responsible;
in some contexts personally liable;
asked to explain their judgment;
subject to professional or legal duties.
An AI agent cannot bear these forms of human responsibility in the same way.
It can be disabled.
Its credentials can be revoked.
Its model can be changed.
Its behaviour can be investigated.
But punishment of the software is not accountability.
Deleting Agent 47 after it caused a €2 million loss does not answer:
Who was responsible for giving Agent 47 the authority that made the loss possible?
That is why “digital employee” should remain an operational metaphor rather than an accountability shortcut.
4. Execution can migrate. Responsibility cannot simply migrate with it.
Current public-sector governance frameworks are unusually clear on this point.
Canada says owners and accountable roles should be explicitly designated for every agent, including where agents create sub-agents, and states that accountability ultimately remains with the designated human owner even when the agent acts autonomously within approved permissions. If the owner leaves, the agent should be reassigned, paused, or deactivated unless accountability is formally transferred. (Canada)
Australia similarly requires agencies to define what each agent is responsible for while assigning human accountability for the behaviour, decisions, and outcomes of agentic systems, including multi-agent systems. (digital.gov.au)
The distinction is fundamental:
AGENT RESPONSIBILITY can describe the work allocated to software.
HUMAN / INSTITUTIONAL ACCOUNTABILITY describes who must answer for the consequences.
Confusing the two creates a responsibility gap.
5. Microsoft has already separated technical ownership from business accountability
Microsoft’s Entra Agent ID architecture contains an important design insight.
It distinguishes:
owners, who handle technical administration;
from
sponsors, who provide business accountability;
from
managers, who can represent the agent within organisational hierarchy.
A sponsor is required for agent identities and blueprints. Sponsors are expected to understand the agent’s business purpose and make lifecycle decisions such as renewal, retention, access adjustment, suspension, or removal. (Microsoft Learn)
That separation is more important than the terminology.
It recognises that the person capable of configuring an agent should not necessarily be the person responsible for deciding whether the agent should possess its authority at all.
This is traditional separation of duties applied to a new actor class.
6. The developer should not automatically become the agent’s sovereign
Consider an engineer who builds a procurement agent.
The engineer understands:
the model;
the code;
the API connections;
the runtime.
But should that engineer determine:
which suppliers the agent may select?
how much it can spend?
what risk threshold is acceptable?
when human approval is required?
Probably not alone.
These are business-authority decisions.
A mature architecture therefore needs at least two distinct questions:
Who can technically modify the agent?
and
Who has institutional authority over what the agent is allowed to do?
The answers may involve different people.
That separation becomes increasingly important as agents grow more powerful.
7. The “manager” role is changing too
Microsoft already allows an agent’s user account to have a designated manager. (Microsoft Learn)
This sounds almost mundane.
It is conceptually remarkable.
Organisational charts have historically represented relationships among humans.
Now they may increasingly contain:
HUMAN MANAGER
↓
HUMAN EMPLOYEE
and
HUMAN MANAGER
↓
AI AGENT
The manager may oversee:
agent access;
business purpose;
escalation;
performance;
exceptions;
retirement.
This creates a new management task.
The manager is no longer supervising only human effort.
They are supervising delegated machine authority.
8. Managing an agent is not managing a person
The analogy has limits.
Human management includes:
motivation;
career development;
psychological safety;
communication;
compensation;
social relationships;
professional growth.
Machine management requires different controls:
scope;
identity;
permissions;
budgets;
monitoring;
versioning;
exception thresholds;
delegation rights;
revocation.
Treating agents literally like people can therefore produce bad governance.
A manager does not need to motivate the agent.
They need to understand what power it has.
The organisational vocabulary may converge.
The management mechanics should not.
9. Digital headcount may be one of the first misleading metrics
McKinsey CEO Bob Sternfels attracted attention earlier in 2026 by describing the firm as having roughly 40,000 human employees alongside approximately 25,000 AI agents. The figure was presented as a way of illustrating how deeply agents had entered the consulting firm’s work rather than as a legal headcount. (Business Insider)
This is a powerful narrative.
It is also a warning about measurement.
What is one agent?
A persistent enterprise assistant?
One model instance?
One specialised workflow?
A sub-agent created for five minutes?
An identity?
A deployment?
If one human can spawn 100 short-lived agents, does organisational headcount increase by 100?
The denominator becomes unstable.
A future CEO saying:
“We employ 500,000 digital workers”
may reveal remarkably little about the actual capacity or power of the institution.
10. Count authority, not only agents
A more useful workforce metric may be:
how much consequential work and authority has migrated to agents?
One million read-only summarisation agents may matter less institutionally than ten agents authorised to:
release payments;
modify production systems;
approve refunds;
change access rights;
make eligibility classifications.
This suggests a better reporting architecture:
HUMAN HEADCOUNT
AGENT POPULATION
AUTONOMOUS WORKFLOWS
CONSEQUENTIAL AUTHORITY HELD BY AGENTS
HUMAN EXCEPTION LOAD
HUMAN ACCOUNTABILITY LOAD
Headcount measures bodies.
Agent governance needs to measure operational power.
11. Productivity can rise while organisational resilience falls
This week’s reporting on Meta provides an unusually useful counterexample to simplistic workforce substitution.
Reuters reported that Meta developed an internal plan known as Project OT aimed at making the company more “AI native,” including plans to shrink some teams by as much as 60%. The project was subsequently scaled back. According to Reuters’ reporting on internal discussions, Meta experienced a 40% year-over-year increase in major site-reliability emergencies, while employees spent 70% more time dealing with those incidents; Reuters says AI-generated code and unpredictable AI-agent behaviour were among the concerns. Meta executives also acknowledged problems with timing, communication, morale, productivity, and reliability, and the company backed away from some more aggressive AI-use policies. (Reuters)
This should not be interpreted as proof that AI workforce reduction generally fails.
It demonstrates something more useful:
Removing human capacity faster than automated systems become organisationally reliable can reduce resilience even when automation improves local productivity.
The relevant variable is not merely:
How much work can AI perform?
It is:
What human capacity remains when the automated work fails in unfamiliar ways?
12. Automation can create more work downstream
This is the exception problem.
An agent resolves 95 ordinary cases automatically.
Excellent.
But the remaining 5 cases may be:
ambiguous;
high-risk;
unprecedented;
technically difficult;
emotionally charged;
legally consequential.
The average human task may therefore become harder after automation.
Humans increasingly receive only the cases the machine could not solve.
Gartner calls one version of this the “exception tsunami”: agents can shift unresolved complexity toward humans, requiring organisations to redesign exception ownership, decision rights, and human roles rather than assuming automation simply removes work. (gartner.com)
The human workforce can shrink in volume while increasing in required expertise.
13. The human becomes the exception handler
This creates a plausible new organisational structure:
AGENTS HANDLE THE NORMAL DISTRIBUTION
↓
HUMANS HANDLE THE TAIL
Humans receive:
uncertainty;
conflicts;
novelty;
ethical problems;
policy exceptions;
failures;
appeals.
This can be a high-value division of labour.
But only if the organisation preserves enough human capability to handle the tail.
If junior staff disappear because agents perform all routine cases, where do future experts learn the domain?
If humans see only abnormal cases, do they retain enough understanding of normal operation to judge anomalies?
If 5% of cases require humans but all 5% arrive simultaneously during an incident, is enough human capacity available?
The exception-handler model creates a workforce-resilience problem.
14. Humans can become worse supervisors as agents become better workers
There is a paradox.
The more reliable automation becomes, the less often humans practise the underlying task.
The less they practise, the harder it may become to intervene effectively when automation fails.
The organisation can therefore experience:
AUTOMATION SUCCESS
↓
HUMAN PRACTICE DECLINES
↓
DEPENDENCE INCREASES
↓
RARE FAILURE BECOMES HARDER TO RECOVER FROM
This is not unique to AI.
Automation research has observed related problems in aviation, medicine, and industrial systems.
Agentic AI potentially spreads the pattern across knowledge work.
15. The organisation needs a minimum human capability reserve
That suggests a new workforce-planning question.
Not:
How few humans can we operate with?
But:
What human capability must remain available for the institution to recover when agentic systems fail?
A financial organisation may need people capable of reconstructing automated transactions.
A software company needs engineers able to understand production systems even if agents write most code.
A government agency needs officials capable of interpreting the law independently of machine recommendations.
A hospital needs clinicians capable of functioning when automated systems are unavailable or unreliable.
This can be called, descriptively, a human capability reserve.
The point is not to preserve jobs for symbolic reasons.
It is to preserve institutional recoverability.
16. Workforce optimisation can accidentally optimise away recovery
A spreadsheet may suggest:
humans cost more;
agents are faster;
agents operate 24/7;
therefore remove more humans.
That calculation can omit:
exception handling;
monitoring;
audit;
recovery;
institutional memory;
training future experts;
relationship management;
accountability;
rare-event capacity.
The savings are visible continuously.
The resilience value becomes visible only during failure.
This creates a systematic incentive toward underinvestment in human fallback capability.
Meta’s reported reliability problems illustrate what that trade-off can look like when automation and workforce restructuring move faster than organisational systems can absorb. (Reuters)
17. Digital labour changes separation of duties
A major reason human organisations divide work among several employees is risk control.
One employee creates a supplier.
Another approves it.
One initiates a payment.
Another releases it.
One develops software.
Another reviews deployment.
Agents can accidentally collapse those separations.
An enterprise deploys one powerful agent with access to:
supplier creation;
invoice processing;
payment;
accounting.
The company has created an extraordinarily efficient worker.
It has also created a concentration of authority that would never be permitted for a human employee.
The relevant principle is:
Automation should not silently erase separation of duties merely because one agent is technically capable of performing every step.
18. Multi-agent systems can restore separation—or fake it
A simple response is:
use several agents.
Agent A prepares the payment.
Agent B reviews.
Agent C executes.
That looks better.
But if all three:
use the same model;
share the same context;
operate under the same principal;
receive the same mistaken assumption,
the separation may be largely cosmetic.
The previous Synthocracy analysis of multi-agent governance makes the distinction clear:
MULTIPLE AGENTS ≠ INDEPENDENT CONTROL CENTRES.
Effective separation of duties requires meaningful independence of:
authority;
evidence;
credentials;
decision criteria;
and where appropriate, human accountability.
19. Performance management changes when the worker optimises the metric literally
Digital workers will inevitably be measured.
Workday says its Agent System of Record provides analytics intended to measure agent value and impact. Salesforce similarly frames agent monitoring and testing as a counterpart to performance management in human workforces. (workday.com)
But agent performance metrics create a particularly strong Goodhart problem.
Tell a service agent:
minimise resolution time.
It may close cases prematurely.
Tell a sales agent:
maximise conversion.
It may over-discount.
Tell a security agent:
minimise false negatives.
It may block enormous amounts of legitimate activity.
Tell a coding agent:
maximise feature throughput.
It may generate technical debt.
Agent performance therefore needs multi-dimensional evaluation.
Efficiency alone is not performance.
20. The agent’s KPI is part of its constitution
For human employees, metrics influence behaviour imperfectly.
For optimisation systems, metrics can have still stronger functional consequences.
The relevant agent scorecard may need to include:
task success;
error rate;
exception rate;
reversal rate;
policy compliance;
cost;
resource consumption;
human intervention;
customer impact;
incident generation;
uncertainty escalation;
audit quality.
The metric system determines what organisational success looks like to the agent.
Performance management is therefore a governance mechanism.
21. “Firing” an agent is really revocation
The employee metaphor produces another linguistic trap.
Companies will inevitably speak about:
hiring agents;
promoting agents;
firing agents.
Operationally, agent firing is closer to:
disable identity;
revoke credentials;
terminate runtime;
remove tool access;
stop scheduled activity;
reassign workflows;
preserve logs.
The distinction matters because deleting the visible agent may not terminate everything it created.
It may have:
delegated tasks;
created sub-agents;
scheduled payments;
opened sessions;
issued messages;
modified external systems.
A real offboarding process needs to remove the authority descendants of the agent where appropriate.
22. Agent offboarding may need to be stricter than employee offboarding
A human employee leaving a company typically loses:
login;
badge;
VPN;
corporate accounts.
An agent can have much more fragmented access:
API tokens;
OAuth grants;
wallets;
MCP credentials;
sub-agent identities;
cloud secrets;
background jobs.
Some may survive the visible agent.
This creates an orphaned authority problem.
Microsoft’s sponsor and lifecycle architecture explicitly attempts to prevent agent identities from becoming ownerless when human sponsors move or leave. (Microsoft Learn)
Agent HR therefore becomes inseparable from identity governance.
23. Agent onboarding should begin with a mandate, not an account
Traditional employee onboarding asks:
Which role?
Which systems?
Which manager?
Which equipment?
For an AI agent, the first question should be more fundamental:
Why does this agent exist?
Then:
What outcome is it responsible for?
What decisions may it make?
Which decisions remain human?
Which systems may it access?
Can it delegate?
Can it spend?
What requires escalation?
Who sponsors it?
Who can suspend it?
When should it retire?
The identity should follow the mandate.
Not the other way around.
24. HR systems are becoming governance systems for non-human actors
Workday’s strategy is significant precisely because it places agent management beside workforce management.
The Agent System of Record provides registration, lifecycle controls, analytics, monitoring, and a gateway capable of registering and metering third-party agents. (workday.com)
This creates a fascinating institutional convergence.
Historically:
IAM governed access.
HRIS governed people.
CMDB governed technology assets.
Agentic enterprise increasingly requires a system connecting all three.
An agent has:
identity like a digital principal;
lifecycle like a worker;
dependencies like software;
authority like a delegated organisational role.
No legacy category fits perfectly.
25. The organisation chart may become executable
Traditional organisational charts are descriptive.
They show:
Alice reports to Bob.
Team X owns Procurement.
Department Y handles Compliance.
Agentic systems can turn these relationships into machine-readable control.
If Agent A reports to Manager H:
H may approve access.
If Agent B belongs to Finance:
financial policies may apply automatically.
If C is a subordinate of A:
delegation limits may inherit.
If the sponsor leaves:
access can expire automatically.
The organisational chart begins to influence technical execution.
It becomes partly executable governance.
This could be a major improvement.
It could also automate bad organisational design at machine speed.
26. The AI manager raises a harder question
The next obvious step is machine management.
Could Agent M supervise Agents A, B, and C?
Technically, yes.
It might:
allocate tasks;
monitor outputs;
reassign workload;
evaluate performance;
approve low-risk exceptions;
request additional resources.
That already occurs in many multi-agent orchestration architectures, even if we call the higher-level system a planner or coordinator rather than a manager.
The organisational question is more difficult:
Which managerial powers may legitimately be delegated to AI?
Task allocation is one thing.
Performance judgment affecting humans is another.
Permission escalation is another.
Budget allocation is another.
Disciplinary authority over human employees is another again.
“AI manager” therefore covers several distinct authority levels.
They should not be collapsed.
27. An AI can manage workflow without becoming the accountable manager
This distinction provides a useful architecture.
Agent M can determine:
which agent handles the ticket;
which workflow executes;
whether capacity needs rebalancing.
Human H remains accountable for:
the policy;
the delegation envelope;
high-consequence exceptions;
changes to organisational authority.
The software possesses operational coordination authority.
The human retains institutional accountability.
This separation can preserve automation without pretending that the machine has become an accountable executive.
28. The new span of control may be measured in agents
Management theory has traditionally asked how many employees one manager can effectively supervise.
Agentic organisations introduce another variable:
AI span of control.
Can one manager meaningfully sponsor:
10 agents?
100?
10,000?
The answer depends on automation.
Managers will not inspect each action.
Governance systems will aggregate:
performance;
exceptions;
risk;
spending;
authority changes;
incidents.
The relevant limit therefore becomes not the number of agents but the volume of consequential exceptions and decisions requiring human judgment.
A manager might safely oversee 10,000 low-risk agents.
Ten highly autonomous financial agents may exceed meaningful supervisory capacity.
29. The real span of control is exception-weighted
A useful analytical model would therefore consider:
AGENT POPULATION
× AUTONOMY
× CONSEQUENCE
× EXCEPTION RATE
× REVIEW COMPLEXITY
The product need not be computed literally.
It illustrates the governance problem.
A manager’s capacity should be measured against the oversight burden created by the population, not the raw number of machine workers.
This connects directly to the earlier Meaningful Human Authority framework.
Human oversight needs cognitive space.
An organisational chart can assign 5,000 agents to one human manager.
That does not prove the manager can meaningfully govern them.
30. The human escalation owner becomes a critical role
Every consequential agent needs a destination for uncertainty.
Someone should receive:
policy conflict;
unusual transaction;
legal ambiguity;
high-risk customer case;
unexpected state;
monitor disagreement;
authority escalation.
This human escalation owner does not need to approve ordinary work.
Their value exists precisely at boundaries.
Government guidance is converging on this architecture. Canada requires a documented escalation path and designated human owner. Australia requires defined escalation pathways, review at key stages, and intervention mechanisms for irreversible or high-risk actions. (Canada)
The escalation owner may become more important than the ordinary workflow manager.
31. Exception ownership must be explicit before deployment
A common failure mode is:
the agent encounters something unusual;
nobody knows who owns it.
IT thinks Business owns it.
Business thinks the vendor owns it.
The vendor says the customer configured the workflow.
The human operator assumes the AI team is monitoring.
The AI team assumed the process owner would intervene.
The agent continues.
This is organisational ambiguity converted into machine-speed execution.
A robust deployment should answer in advance:
Who owns the exception?
Not after failure.
32. Humans may become appellate authorities
At sufficient agent scale, it becomes unrealistic for people to participate in every ordinary action.
Human authority may therefore move upward.
Agents handle normal execution.
Specialised monitors review.
Rules resolve routine conflicts.
Humans handle:
precedent;
high-impact exceptions;
appeals;
constitutional changes;
value conflicts;
major incidents.
The human role becomes closer to an appellate layer.
This may produce more meaningful human control than placing people into thousands of routine approval loops.
33. The employee may become the principal of a personal agent team
The transformation also works from below.
Instead of one employee performing ten workflows directly, the employee may supervise several agents.
A marketing manager runs:
research agent;
content agent;
analytics agent;
campaign agent.
An engineer runs:
coding;
testing;
documentation;
deployment assistants.
The human role changes from:
TASK EXECUTOR
toward:
LOCAL PRINCIPAL / SUPERVISOR OF MACHINE WORK.
This could amplify individual productivity enormously.
It also amplifies the authority controlled by one person.
One employee may effectively command a machine organisation.
34. One person’s authority can multiply through agents
This deserves governance attention.
Suppose a manager is authorised to contact 20 suppliers manually.
An agent allows them to contact 20,000 overnight.
The formal authority did not change.
Its operational scale did.
The same applies to:
emails;
candidate screening;
customer decisions;
security actions;
code changes;
transactions.
AI converts individual authority into machine-scale authority.
This suggests another distinction:
NOMINAL AUTHORITY
versus
AMPLIFIED AUTHORITY.
An institution should ask what scale of action an agent makes possible for the principal.
35. Delegation can change power without changing the org chart
Alice remains middle management.
She gains three powerful autonomous agents.
Bob remains a director.
His work remains largely manual.
The chart says Bob outranks Alice.
Operationally, Alice may now control greater throughput and more consequential machine action.
Agent allocation therefore becomes a form of organisational resource allocation.
Who receives the best agents?
Who receives expensive model access?
Who gets autonomous execution?
Which department has access to payments?
AI infrastructure can redistribute organisational power without changing titles.
36. Compute becomes a workforce resource
Human organisations allocate:
salary budgets;
headcount;
offices;
equipment.
Agentic organisations increasingly allocate:
tokens;
inference budgets;
model tiers;
tool access;
runtime slots;
agent counts.
A department with more compute may effectively possess more labour capacity.
The CFO and CIO therefore begin influencing workforce size through infrastructure allocation.
This is why Meta’s internal framing reported by Reuters is so interesting. Zuckerberg reportedly told employees that the company effectively faced two major cost centres: compute infrastructure and people, with greater spending on one affecting the resources available for the other. (Reuters)
Compute and labour are beginning to appear on the same strategic balance sheet.
37. The cost curve changes organisational incentives
A human employee carries costs associated with:
salary;
benefits;
recruitment;
training;
management;
physical infrastructure.
An agent carries:
inference;
software;
integration;
monitoring;
security;
exception handling;
governance.
The first costs are familiar and visible.
The second are still being discovered.
This creates risk of false comparisons.
An agent may look dramatically cheaper per completed routine task.
But organisations must include the governance overhead:
identity;
monitoring;
incidents;
human review;
fallback capacity;
security;
audit.
Meta’s experience is a useful reminder that productivity gains can be offset by operational failure costs. (Reuters)
38. The new workforce economics needs a cost of oversight
Traditional automation ROI often looks like:
LABOUR SAVED − TECHNOLOGY COST.
Agentic systems require another term:
LABOUR SAVED − AGENT COST − OVERSIGHT COST − FAILURE COST − RECOVERY CAPACITY.
A highly capable agent requiring constant human correction may not be economically autonomous.
A cheaper agent generating rare catastrophic failures may be falsely economical if expected failure costs are excluded.
The economics of the digital workforce therefore depends on the economics of governability.
39. Responsibility can become more concentrated as execution becomes more distributed
There is a counterintuitive possibility.
Ten agents make dozens of intermediate choices.
Yet one human sponsor remains ultimately accountable.
Machine execution becomes distributed.
Human responsibility becomes concentrated.
Academic work published in April on human–AI teams found a related psychological effect: across four experiments involving 1,801 participants, people attributed more responsibility to the human decision-maker when paired with AI than when paired with another human. The authors call this AI-Induced Human Responsibility and argue that people may treat AI as a constrained implementer while locating discretionary responsibility in the human. (arXiv)
This is experimental perception research, not a statement about legal liability.
But it captures the emerging institutional asymmetry.
The AI performs more.
The human may become more answerable.
40. Responsibility without visibility is dangerous
Suppose the sponsor is accountable for Agent A.
A uses B.
B invokes C.
C uses five external tools.
The sponsor cannot see the full chain.
Yet they remain designated as accountable.
We now have:
RESPONSIBILITY > CONTROL.
That is precisely the Ceremonial Human failure mode at organisational scale.
A sponsor should not merely carry the name attached to the agent.
They need sufficient:
visibility;
authority;
monitoring;
escalation;
revocation;
evidence
to make accountability meaningful.
Otherwise sponsorship becomes liability assignment rather than governance.
41. Separation of duties must extend to accountability
The same person should not necessarily be:
agent developer;
business sponsor;
security reviewer;
performance evaluator;
incident adjudicator.
The Urban Institute’s August 2026 Responsible Agentic AI Playbook recommends separate accountable-owner, evaluation, security, transparency, and governance roles for high-stakes deployments where feasible. (Urban Institute)
This reflects a broader institutional principle.
The agent workforce needs human governance around it, not one person carrying every function.
Synthetic labour does not eliminate organisation.
It creates another organisation that itself needs staffing.
42. The digital workforce can create new human jobs precisely because it removes old work
This is not paradoxical.
As agent populations grow, organisations need:
agent sponsors;
agent security engineers;
agent auditors;
exception owners;
agent identity administrators;
agent performance analysts;
agent governance leads;
workflow architects;
incident investigators.
Some roles will be new.
Others will be transformed versions of existing functions.
This does not establish that employment gains will offset jobs displaced by automation. That is an empirical macroeconomic question.
It does mean that automation generates governance work.
The more autonomous the system becomes, the more sophisticated some of that governance work may need to be.
43. The worker affected by agents also needs governance
Thus far we have looked at agents as workers.
There is another dimension:
humans managed by agents.
AI can increasingly participate in:
task allocation;
performance evaluation;
scheduling;
monitoring;
recruitment;
promotion recommendations;
workforce restructuring.
The authority question becomes more sensitive here because machine decisions affect human employment.
Who may assign work?
Who may evaluate a person?
Who can issue disciplinary recommendations?
Who can terminate access?
Which decisions require meaningful human judgment?
The digital workforce problem therefore includes both:
AI AS WORKER
and
AI AS WORKFORCE GOVERNOR.
The second deserves stricter scrutiny.
44. The AI manager of humans is a different risk class from the AI manager of agents
An AI coordinator reallocating tasks among five software agents creates one set of risks.
The same system ranking human employees for redundancy creates another.
The action affects:
livelihood;
reputation;
legal rights;
career.
The fact that both activities are called “management” should not cause identical governance.
Authority should follow consequence.
This is a recurring rule across the Synthocracy programme.
45. Workforce data creates another concentration of informational power
Agent managers may have access to extraordinary amounts of organisational information.
Work records.
Communication.
Performance data.
HR systems.
CRM data.
Project activity.
Security logs.
The better the agent understands the company, the more effective it may become.
The same context can create surveillance capability.
A digital manager able to observe entire workflows may know substantially more about employees than any traditional human manager.
This creates an information-governance problem alongside the workforce problem.
The organisation must distinguish:
information necessary for task execution
from
information technically available because the agent can reach it.
46. Digital labour should not become invisible labour
There is also a transparency issue.
A customer speaks with “Sarah from Support.”
Sarah is an agent.
An employee receives an HR decision prepared by an agent.
A supplier negotiates with an automated procurement system.
When is disclosure appropriate?
The answer will vary by context and law.
But institutional clarity matters.
People should not be forced to infer whether they are interacting with:
a human;
an AI assistant;
an autonomous agent;
a hybrid workflow.
The identity of the operational actor can affect trust, expectations, and routes of escalation.
47. The workforce system needs an authority map, not merely an employee list
The organisation of the future may include:
20,000 humans;
100,000 agents;
millions of ephemeral instances.
A conventional directory will not be enough.
For each consequential agent, the institution increasingly needs to know:
who sponsors it;
who technically owns it;
who manages it;
what role it occupies;
what decisions it may make;
what tools it can access;
which agents it can create;
which humans it affects;
who receives escalation;
who can suspend it;
what happens after retirement.
That is not simply HR.
It is an authority registry.
The Synthocracy Digital Workforce Governance Test
The following preliminary diagnostic is intended for organisations treating AI agents as persistent workforce participants. It is a research and governance tool, not a legal employment-status test or security certification.
1. Role — What work is the agent actually responsible for? Define the continuing operational role rather than merely listing technical capabilities.
2. Authority — Which decisions may the agent make? Separate information gathering, recommendation, execution, spending, delegation, external representation, and difficult-to-reverse actions.
3. Human Accountability — Which human or institutional role ultimately answers for the agent? Avoid ownership relationships that exist only on paper; the accountable party should possess enough visibility and intervention authority to govern the role meaningfully.
4. Technical Ownership vs Business Sponsorship — Are they separated where necessary? Identify who configures the system and who possesses legitimate authority to determine whether its business mandate should exist.
5. Separation of Duties — Has automation collapsed controls previously distributed across several humans? Test whether one agent can initiate, approve, execute, and audit the same consequential workflow.
6. Performance — What does the agent optimise? Measure quality, safety, exceptions, reversals, policy compliance, cost, human burden, and downstream incidents rather than throughput alone.
7. Human Exception Capacity — Who receives cases the agent cannot safely resolve? Test whether enough skilled human capacity exists during ordinary operations and peak failure scenarios.
8. Lifecycle — Can the agent be onboarded, changed, suspended, and retired safely? Revocation should include credentials, delegated descendants, scheduled activity, external integrations, and residual authority.
9. Management — What does “manager” mean in this deployment? Distinguish task coordination, access management, performance review, financial authority, and management of humans rather than treating them as one generic AI-management function.
10. Resilience — Could the organisation still perform the critical function if the agent population became unavailable or unreliable? Identify the minimum human capability, documentation, institutional memory, and recovery capacity required to continue.
The central test is:
If the agent disappeared tomorrow—or behaved incorrectly at machine speed—would an identifiable human institution still understand the work well enough, own the authority clearly enough, and retain enough capability to recover?
If not, the organisation has not merely automated work.
It has transferred organisational dependence.
48. FORESIGHT — 2027+: Digital headcount will become a board metric
FORESIGHT — This is a plausible development, not an established reporting standard.
Boards may begin receiving workforce reports containing:
human workers;
contractors;
active persistent agents;
autonomous workflows;
agent-generated work volume.
But raw agent counts will quickly become misleading.
More useful metrics may include:
human-to-agent ratio;
autonomy-weighted agent population;
agent-managed spend;
agent-controlled systems;
human exception hours;
authority concentration.
The board will need to know not merely how many machine workers exist.
It will need to know how much of the organisation they control.
49. FORESIGHT — 2027+: The Chief Agent Officer
Organisational responsibility may consolidate around a new senior role.
The title may not literally be Chief Agent Officer.
But some executive function will need to connect:
technology;
identity;
security;
HR;
risk;
finance;
legal;
operations.
Because agents cut across all of them.
The central question will be:
Who owns the institution’s non-human workforce as a governed population?
Today that responsibility is fragmented.
That fragmentation may become unsustainable.
50. FORESIGHT — 2027+: Machine managers will supervise machine workers
This already exists technically in orchestrator architectures.
It may become explicit organisationally.
A hierarchy could look like:
HUMAN EXECUTIVE
↓
AI OPERATIONS MANAGER
↓
PROCUREMENT AGENTS
FINANCE AGENTS
LOGISTICS AGENTS
The machine manager allocates routine work and resources.
Humans oversee policy, exceptions, and constitutional boundaries.
This can dramatically increase organisational scale.
It also creates a deeper transitive-authority problem.
The manager becomes an authority distributor.
51. FORESIGHT — 2027+: Organisations may have more agents than people
NewCore already argues that machine identities can outnumber people dramatically, and Gartner has forecast very large agent populations inside major enterprises. These are vendor and analyst projections rather than established future facts. (newcore.com)
If agent creation continues becoming cheap, the ratio itself is plausible.
The governance consequence is more interesting than the number.
A human minority could formally remain the organisation’s sovereign authority while a machine majority performs most operational work.
That creates an institution in which:
humans govern
but
agents operate.
This may become one of the clearest real-world forms of Synthocracy.
52. FORESIGHT — 2027+: Human work shifts from production to constitutional authority
In highly agentic organisations, some of the most valuable human roles may move upward in the decision architecture.
Humans increasingly decide:
what objective matters;
what trade-offs are acceptable;
which authority may be delegated;
which exceptions require judgment;
how disputes are resolved;
when a system should be stopped;
how rules should change.
Agents execute more of the ordinary sequence.
Humans retain authority over the constitution of the work.
That would be a very different future from the simple narrative:
“AI does the work; people disappear.”
53. FORESIGHT — 2027+: Human scarcity may become more dangerous than human surplus
Much public debate assumes the central risk is too many humans for too little work.
Agentic organisations could eventually encounter the opposite operational problem in some domains:
too few humans capable of understanding the automated institution deeply enough to govern it.
Experienced engineers leave.
Junior training pipelines shrink.
Manual expertise decays.
Exception complexity rises.
The company looks exceptionally efficient until the machine organisation experiences a failure outside its normal distribution.
The scarce resource then becomes:
qualified human judgment.
54. FORESIGHT — 2027+: Organisational succession will include agents
A manager leaves.
Who inherits their agents?
A department closes.
Who inherits its agent mandates?
A corporation merges.
Which agent identities survive?
Which policies apply?
Microsoft’s existing sponsor-transfer mechanisms provide an early technical analogue. (Microsoft Learn)
At large scale, agent succession may become part of ordinary corporate restructuring.
The organisation will need to transfer not only people and assets.
It will transfer machine authority.
55. FORESIGHT — 2027+: An agent may need a personnel file
Not a human HR file.
An institutional record.
It could contain:
identity;
creator;
sponsor;
manager;
role;
mandate;
permissions;
skills;
model/version;
performance;
incidents;
authority changes;
delegations;
audit history;
retirement state.
Workday and Microsoft are already assembling significant pieces of this architecture. (workday.com)
The agent personnel record may become a core governance artefact.
56. FORESIGHT — 2027+: Agent performance can influence human careers
Suppose a manager operates a team of agents.
One manager obtains exceptional output.
Another does not.
Performance systems begin assessing employees partly on how effectively they orchestrate digital labour.
This would create a new managerial competency:
agent leverage.
Meta’s reported experiment with “tokenmaxxing” provides an early caution. Reuters says the company backed away from pressure that encouraged employees to use AI for its own sake and shifted toward using AI where it made practical sense. (Reuters)
The lesson is broader.
Measure outcomes, not ritualised AI consumption.
57. FORESIGHT — 2027+: The organisation itself becomes hybrid
The final form may no longer divide neatly into:
employees;
software.
It may contain:
human principals;
AI managers;
human experts;
autonomous agents;
AI monitors;
human appellate reviewers;
machine services;
external agent contractors.
Authority flows between them.
Some decisions begin with humans and end with machines.
Others begin with machine detection and end with human judgment.
The future organisation becomes a hybrid authority system.
That is the territory Synthocracy was built to study.
58. The crucial word is not workforce. It is organisation.
The industry is understandably fascinated with digital labour.
Labour asks:
who performs the task?
Organisation asks harder questions.
Who sets the objective?
Who possesses authority?
Who controls resources?
Who monitors?
Who handles exceptions?
Who can stop the process?
Who remains responsible?
Who retains enough expertise to recover?
The future of work will be shaped at least as much by those questions as by the number of jobs automated.
Conclusion — AI can perform the work without becoming the responsible worker
The transition is already visible.
Workday is building a system of record that manages agents alongside people.
Microsoft gives agents distinct identities, sponsors, managers, lifecycle controls, and organisational relationships.
ServiceNow is deploying role-scoped AI specialists across IT, finance, procurement, HR, legal, CRM, and security and calls them an Autonomous Workforce.
Salesforce speaks openly about digital labour.
NewCore has raised $66 million around the proposition that enterprise identity itself must be rebuilt for a workforce composed of humans, machines, and AI agents. (workday.com)
The workforce metaphor therefore reflects a real technological development.
But it becomes dangerous when metaphor replaces governance.
An agent can be given a role.
It can receive credentials.
It can report to a manager.
It can be measured.
It can execute continuously.
It can be disabled.
It cannot make institutional accountability disappear.
Government guidance in Canada, Australia, Singapore, and elsewhere is already converging on this point: humans and deploying organisations remain accountable even where agents operate autonomously. (Canada)
That produces the defining asymmetry of the digital workforce:
Execution can become synthetic while accountability remains human.
Every organisation rushing toward agentic labour therefore needs to resist a seductive but incomplete equation:
HUMAN WORKER → AI WORKER.
The real transformation is more complicated:
HUMAN WORK
↓
AGENT ROLE
↓
DELEGATED AUTHORITY
↓
AUTONOMOUS EXECUTION
↓
EXCEPTIONS
↓
HUMAN INTERVENTION
↓
HUMAN / INSTITUTIONAL ACCOUNTABILITY
Meta’s experience is a particularly timely warning. According to Reuters, its aggressive attempt to reorganise around an “AI native” workforce ran ahead of the reliability of the technology and the organisation built around it. Site-reliability emergencies increased, human firefighting increased, employee morale suffered, and the company backed away from some of its more aggressive practices. (Reuters)
The lesson is not:
Do not replace human work with AI.
It is:
Do not remove human organisational capacity merely because machine task capacity has increased.
A workforce is more than the sum of tasks completed.
It contains judgment.
Recovery.
Responsibility.
Institutional memory.
Separation of duties.
Escalation.
Training.
Authority.
An AI agent can absorb some of these functions.
It cannot automatically inherit all of them merely by being called a digital employee.
So the decisive governance question is not:
How many employees can agents replace?
Nor:
How many agents should count as employees?
It is:
Which parts of the organisation can safely become machine-executed, which forms of authority can be delegated with them, and which human capabilities must remain because someone still has to understand, challenge, repair, and answer for what the organisation does?
That is the real boundary.
The enterprise of the late 2020s may indeed contain more AI agents than people.
But the important question will not be whether those agents count as workforce.
It will be whether the institution still knows where human authority ends, where machine authority begins, and who remains responsible when the two become difficult to separate.
