The Human Sub-Agent: What Happens When AI Starts Commissioning People

The provocative simplification

An AI agent may eventually treat a person like a tool call.

Not because people become tools.

Because some work will still require human judgment, physical action, access, trust, or context.

The agent will encounter a task it cannot complete digitally.

It will send a request to a human.

The human will complete the task and return evidence.

The agent will continue.

This is a hypothesis, not a prediction.

The evidence today comes mostly from prototypes, simulations, and early research.

But the direction matters.

We have spent years discussing humans in the loop.

The next question is different:

What happens when the human becomes a sub-agent inside an AI-managed workflow?

That question changes how we think about work, management, operating models, and the value consultants can create.

Wait. Isn’t AI supposed to assist humans?

Yes.

That is still the dominant pattern.

A person asks an AI system to write, search, summarize, analyze, or build.

The AI assists the person.

But agentic systems introduce a different operating model.

The person gives the system a goal.

The agent breaks the goal into tasks.

The agent uses software, tools, data, and other agents to complete the work.

When it reaches a capability gap, it asks for help.

That help may come from another AI agent.

It may come from a person.

It may come from a specialist inside the organization.

It may come from someone who can do something the system cannot do safely or legally.

The direction of delegation changes.

This is not science fiction.

Microsoft Research’s Magentic-UI treats the human user as a special agent inside a multi-agent team.

The system can ask the user to clarify a decision, complete a CAPTCHA, provide information, or take control of a browser.

In one simulated evaluation, adding a user with relevant side information increased task completion from 30.3% to 51.9%.

The system did not need the human for every task.

It needed the human at the right moment.

That distinction is the beginning of the new model.

A real system already shows the pattern

Magentic-UI is not a deployed enterprise workforce system.

It is a research prototype.

That limitation matters.

Still, the architecture makes the idea concrete.

The system has an orchestrator.

The orchestrator creates a plan.

The plan assigns steps to different agents.

One of those agents is the human user.

The system can run a browser, execute code, manipulate files, and coordinate activity.

Then it can stop and say, in effect:

“I cannot complete this step. I need you to do it.”

The human may need to enter information.

The human may need to resolve ambiguity.

The human may need to take over the browser.

The human then returns control to the system.

This is not a chatbot waiting for a prompt.

It is a workflow assigning a bounded task to a person.

Deconstructing the mechanism

The mechanism is simpler than the industry language makes it sound.

1. The agent receives a goal

“Prepare the workforce transition plan.”

“Complete the regulatory submission.”

“Investigate the customer complaint.”

“Launch the new product capability.”

The goal is usually too broad for one action.

2. The agent creates a task graph

The goal becomes a sequence of connected tasks.

Some tasks can happen in parallel.

Some tasks depend on earlier decisions.

Some tasks require specific evidence before the workflow can continue.

This is where the work becomes operational.

Not a prompt.

Not a demo.

A real workflow.

3. The agent checks its available capabilities

The agent may have access to:

• A company knowledge base.
• A browser.
• A CRM.
• A data warehouse.
• A coding environment.
• An approval system.
• Other AI agents.
• A list of human specialists.

The agent then asks a practical question:

“Who or what can complete this step?”

4. The agent finds a capability gap

The system may not know the local history of a team.

It may not have permission to enter a sensitive system.

It may not be able to conduct a difficult conversation.

It may not be able to inspect a physical location.

It may not be trusted to make a regulated judgment.

It may lack the social context required to interpret the evidence correctly.

This is the point where the human becomes relevant.

5. The agent commissions the human

The request must be specific.

Not:

“Please help with this.”

Better:

“Review the proposed workforce transition for the European sales team. Confirm whether the role mapping reflects current responsibilities. Identify any material exceptions. Return the completed review template with supporting evidence by 3 p.m.”

That is a task.

It has a purpose.

It has boundaries.

It has an expected output.

6. The human returns evidence

The human does not simply say, “Done.”

The human returns an artifact.

That may be:

• A completed form.
• A decision.
• A photo.
• A signed approval.
• A short explanation.
• A structured exception.
• A link to a source document.
• A conversation record.
• A risk flag.

The agent needs evidence it can use.

This is the crucial shift.

Human contribution becomes part of the machine-readable workflow.

The familiar analogy

Think of an AI agent as a software system with a growing collection of tools.

A calculator is a tool.

A search engine is a tool.

A database is a tool.

An API is a tool.

Now imagine a tool that can provide local judgment, physical action, or social context.

That tool is a person.

The analogy stops there.

A person is not an API.

A person has rights, preferences, expertise, relationships, and the ability to refuse.

A person also carries responsibility and risk.

So the design cannot treat human delegation as a simple function call.

It needs authority, consent, workload limits, privacy controls, and a clear accountability model.

This is where many vendor demos fall apart.

They show the agent assigning work.

They do not show who owns the outcome.

They do not show whether the human has the time or authority to complete the task.

They do not show what happens when the human disagrees.

They do not show whether the evidence can be audited six months later.

A delegation feature is not an operating model.

The research is useful because it exposes the hard part

A 2025 paper on orchestrating human-AI teams describes a “Manager Agent” that decomposes goals, allocates tasks to human and AI workers, monitors progress, and adapts the workflow.

The research team tested simulated workflows across 20 scenarios.

The results were not a victory lap.

The agent could complete goals.

But it struggled to optimize completion, speed, constraints, and stakeholder preferences at the same time.

In the reported tests, the system assigned 25.8% of tasks to humans.

That increased goal achievement in some workflows.

It also increased execution time.

This is the uncomfortable truth.

A human handoff may improve quality while slowing the system down.

That is not a reason to remove the human.

It is a reason to design the handoff properly.

The question is not:

“Can AI replace this person?”

The better question is:

“Which part of this workflow needs human capability, and how should the system use it?”

A framework for the human commissioning stack

When I think about this problem, I break it into six parts.

1. Mission

What outcome is the agent trying to achieve?

If the outcome is unclear, the delegation will be unclear.

“Improve employee experience” is not a mission.

“Identify the three highest-friction steps in the internal mobility process and recommend which one to redesign first” is closer.

2. Capability gap

Why can the agent not complete the task alone?

The answer should be explicit.

The gap may involve:

• Access.
• Authority.
• Physical presence.
• Specialized knowledge.
• Human trust.
• Judgment under uncertainty.
• A regulated decision.

If the gap is only poor system design, do not create a human handoff.

Fix the system.

3. Bounded instruction

What exactly must the human do?

Keep the request narrow.

Give the human the context required to act.

Do not make them reconstruct the entire problem from scratch.

4. Authority

What can the human decide?

What can they recommend?

What requires escalation?

What can the agent do with the returned evidence?

This is especially important in People decisions.

An HR business partner may validate context.

That does not mean they should silently approve a termination recommendation generated by an agent.

5. Evidence

What must the human return?

Do not accept vague completion signals.

Define the artifact.

Define the quality standard.

Define the source.

Define the exception process.

If the workflow cannot use the output, the task was not properly designed.

6. Accountability

Who owns the result?

Not just who clicked the button.

Who designed the workflow?

Who authorized the agent?

Who approved the delegation?

Who reviewed the evidence?

Who carries the business consequence?

Research on autonomous workflow systems has already raised the risk of a moral crumple zone, where a human operator receives the blame for a failure created by the wider system.

That is not governance.

That is scapegoating with better software.

What this means for People leaders

For senior People leaders, this is not only a technology question.

It is a workforce design question.

If agents can delegate work to people, organizations will need to define:

• Which tasks agents may assign.
• Which employees may receive those tasks.
• How workload is measured.
• How priority conflicts are resolved.
• How human judgment enters the workflow.
• How evidence is recorded.
• How employees challenge an agent’s request.
• How managers support people working inside agent-managed systems.

The People function cannot wait until the software is already live.

By then, the operating model will already be forming.

You will simply be inheriting decisions made by product teams, vendors, and enthusiastic builders.

Not HR policy.

Not a training module.

The real work is designing how human capability enters an AI-enabled organization.

That is a strategic role.

What this means for consultants

This is also where the commercial opportunity moves.

Do not sell another generic “AI readiness” workshop.

Do not charge for a slide deck explaining that agents are important.

Do not build a shiny chatbot and call it transformation.

Clients will pay for consequences.

They will pay for:

• Mapping where human intervention creates value.
• Redesigning workflows around agent and human capabilities.
• Defining evidence standards.
• Creating delegation and escalation rules.
• Building governance into the process.
• Prototyping a safe human handoff.
• Measuring quality, speed, workload, and risk.
• Helping leaders decide what to automate, augment, or preserve.

That is work with a clearer business case.

It affects cycle time.

It affects capacity.

It affects risk.

It affects the cost of operating the function.

It affects whether the organization can actually ship the promised value from AI.

The valuable consultant is not the person who explains the agent.

The valuable consultant is the person who redesigns the work around the agent.

That is the difference between selling information and charging for transformation.

Start with one workflow

Do not begin with an enterprise-wide vision.

Start with one workflow where the value and risk are visible.

For example:

• Internal mobility.
• Workforce planning.
• Policy interpretation.
• Employee case triage.
• Learning needs analysis.
• Regulatory documentation.
• Talent review preparation.

Map the current process.

Mark every step the human performs.

Then ask five blunt questions:

1. Can an agent complete this step with existing data and authority?
2. If not, what capability is missing?
3. Can the human task be made smaller and clearer?
4. What evidence must come back?
5. What happens if the human disagrees with the agent?

Then build a small prototype.

Use a controlled dataset.

Use a low-risk workflow.

Keep an audit trail.

Test the handoff.

Measure the result.

Ship it before you design the future-state operating model for the entire company.

This is where practitioner-led research and roadmaps from the HR AI Institute can help leaders move from AI experimentation to operating architecture.

For builders, this may start as a lightweight tool.

For consultants, it may start as a paid workflow redesign engagement.

For People leaders, it may start as a decision about which process deserves attention first.

Not Yet versus Now

The Not Yet group

You are still collecting AI tools.

You are still asking which model is best.

You are still running isolated demos.

You have not mapped a real workflow.

You are not ready to commission humans through an agent.

You are ready to identify one process where work is already unclear, slow, or expensive.

The Now group

You are already responsible for AI transformation.

Your teams are using agents without a shared design.

Your consultants are being asked to prove measurable value.

Your leaders want productivity gains but cannot explain what work will be removed.

Your organization needs a safe prototype.

You should be designing the human handoff now.

Not after the vendor installs the system.

Not after an employee is blamed for an agent’s failure.

Not after the board asks why the promised savings never appeared.

The deeper implication

The future of work may not be divided into humans and machines.

That is too simple.

The real organization may contain:

• Humans directing agents.
• Agents directing agents.
• Agents requesting human judgment.
• Humans supplying evidence to agents.
• Managers supervising systems that allocate work.
• People building small tools to extend their own capability.

The important question is not whether humans remain involved.

They will.

The important question is who designs the terms of that involvement.

IT will design some of it.

Vendors will propose some of it.

Consultants will influence some of it.

But People leaders understand the human system behind the workflow.

They understand capability.

They understand trust.

They understand identity.

They understand what happens when a process looks efficient on paper but fails in the lived experience of work.

That knowledge is not a soft advantage.

It is operating leverage.

Takeaway

The human sub-agent is a hypothesis grounded in an emerging design pattern.

Today’s systems already show early versions of agents asking humans to clarify, intervene, complete inaccessible steps, and return evidence.

The technology is not mature.

The organizational question is already here.

Do not ask only what AI can do.

Ask what the agent cannot do.

Ask which human capability fills that gap.

Ask how the request should be bounded.

Ask what evidence must return.

Ask who owns the consequence.

The next generation of AI transformation will not be built by automating every human task.

It will be built by architecting the relationship between digital capability and human judgment.

That is the work.

That is where the value is.

And that is where practitioners and consultants can earn the right to charge for consequence, not just conversation.