Six Questions That Are a Constitution for Work

When AI agents are in the mix, make sure you answer these six questions.

Start with the act.

Work is something done to a unit — a claim, order, invoice, lead, case, package, part or room — that leaves behind a consequential state.

What act is being performed? What state exists afterward that did not exist before?

A generated recommendation is an artifact.
Approving the claim is an act.
“Claim approved” is a state.

A planned grasp is an artifact.
Seating the part is an act.
“Part seated” is a state.

Two useful cuts:

• The act is not the artifact.
• The act is not the participant.

An artifact is not necessarily an outcome.

Motion is not necessarily work merely because something moved.

Once an act exists, something has to carry the work forward.

Sometimes that is:

• a predefined route
• a manager or planner
• shared state and rules
• a handoff
• a committee
• a market
• a shared board
• a peer network
• a shared floor

I call that arrangement a Formation.

What carries coordination? How do participants know what happens next?

A Formation is not the work itself.

It is the arrangement through which the work progresses.

And that arrangement can be physical as well as digital.

An aisle, zone, fixture, room or charger can constrain coordination just as surely as a queue or workflow can.

Humans, software agents and physical systems can all participate in work.

But knowing who performs the task is not enough.

Who performs, supervises, delegates, reviews, overrides and remains accountable?

Those relationships are not interchangeable.

• Performance is not accountability.
• Participation is not ownership of the work.
• Accountability does not automatically move to whichever participant performed the act.

An agent joins the work as a participant.

It does not become the work itself.

A physical system does the same.

A deployed agent, an agent persona, a robot body, a role, a task and a unit of work are different objects.

Collapsing them into one hides important distinctions in:

• responsibility
• capability
• control

Every act stands on some basis:

• facts
• evidence
• policies
• documents
• observations
• enterprise state
• its own brief, its place in the work and the rules it was given

Some of that basis may be current or stale, authoritative or weak, direct or inferred.

What is this act relying on? How strong, current and authoritative is that basis?

A refund decision may rely on policy, transaction history and account state.

A robot may rely on a camera feed, a map and a current estimate of where a person is standing.

A financial action may rely on enterprise state recorded in a system.

None of those is simply “truth.”

• A policy is normative.
• A camera feed is evidence.
• A system of record is an institutional representation of state.

Grounding tells a participant what it may reasonably believe.

It does not tell the participant what it may do.

A system can be well grounded and still be unauthorized to act.

Capability is not authority.

An agent may be able to issue a refund, approve a claim or place an order without having the organizational authority to do so.

A physical system may be able to enter an aisle, lift a part or apply force without having the grant to do so in that place, on that object, at that moment.

So the question becomes:

May this attempted effect become a real transition, under this grant, now?

A humanoid on an assembly floor makes the distinction concrete.

It may:

• recognize a misaligned component
• plan a corrective grasp
• be physically capable of reseating it

None of that answers whether it may:

• enter an occupied zone
• apply that amount of force
• make the correction without approval

For embodied work, authority includes force, place and proximity.

May this body apply this force to this object, here, now?

Authority is not inferred from capability.

And it does not transfer automatically simply because the work has moved to a different participant.

Not every completed act becomes an accepted fact.

• A recommendation is not a payment.
• A diagnosis is not a discharge.
• A generated purchase order is not necessarily a committed obligation.
• A package placed on a shelf is not necessarily inventory posted.

Something still has to establish what the organization recognizes as having become true.

That is settlement.

What became real? What result will the organization stand behind?

A robot may physically place a component in position, but the assembly step may still require:

• torque verification
• inspection
• a system-of-record update

before the organization recognizes the state as “installed.”

This creates an important temporal cut:

• Grounding: what the participant had reason to believe before the act
• Settlement: what the organization recognizes as having become true after it

That distinction matters even more when the physical world has already changed and the act cannot simply be undone.

Once those six questions are explicit, some operational and economic laws become easier to see.

If 20% of a person's work moves to an agent, that does not mean capacity increased by 20%.

The 20% tells us about allocation.

Capacity comes from the actual operating capability of the participant taking the work.

If an agent can genuinely process 2,000 cases a day, and the Formation allows that capacity to serve the work, then 2,000 a day is a real capacity fact.

But:

capacity ≠ throughput

If only 40 cases arrive, the Work System does not suddenly produce 2,000 completed cases.

The same applies to physical systems.

A machine capable of moving 2,000 packages a day does not create 2,000 packages if only 40 arrive.

A new participant may improve far more than capacity.

It may change:

• cycle time
• first-pass yield
• defect rate
• exception rate
• rework
• service levels
• supervision requirements
• risk exposure

But those effects do not belong to the participant in isolation.

An agent that improves classification accuracy may reduce downstream review.

Another that shortens processing time may matter very little if settlement remains the bottleneck.

A robot that moves material faster may create little value if inspection or loading remains constrained.

So the important question is not simply:

What benefit does this agent or machine bring?

It is:

What properties of the Work System change because this participant is present — and what do those changes cause downstream?

That is where operational benefits become economically meaningful.

A participant can remove work.

It can also create work:

• exceptions
• supervision
• resets
• charging
• recovery
• review
• reconciliation

Changing the arrangement can therefore move several burdens at once:

• execution
• coordination
• review
• intervention
• failure exposure
• transition effort

A twenty-agent arrangement may lower execution cost while increasing coordination or review.

A single capable agent may cost less overall.

A deterministic route may still be preferable when the work is predictable or the consequences of error are high.

So:

released hours are not automatically saved hours

and

a cheaper performer does not automatically produce a cheaper Work System

• Released hours are not agent capacity.
• Agent capacity is not throughput.
• Agent cost is not Work-System cost.
• Operational improvement is not automatically economic value.

Higher capacity, fewer defects, shorter cycle time and better first-pass yield are already real operating benefits.

But economic value requires another step.

Did additional capacity serve more demand?

Did it avoid a hire?

Did better quality eliminate rework?

Did lower failure exposure avoid a loss?

Did faster processing create revenue or margin?

If that causal relationship is known, economic value can be estimated or measured.

If it is not known, the value is unmeasured.

And unmeasured is a statement about what we know.

It is not a substitute for establishing the causal bridge.

AI — and increasingly capable physical systems — dramatically expands who and what can participate in work.

But:

• “agentic” is not itself a way of organizing work
• “swarm” is not itself a way of organizing work
• “humanoid” is not itself a way of organizing work

Those terms describe participants and capabilities.

They do not describe the operating arrangement.

The deeper opportunity is therefore not simply to automate more tasks or add more participants.

It is to reconsider:

how work itself comes together and is made to operate

That is why I keep coming back to the Work System.

It gives us one object from which to reason about:

• the act
• the arrangement
• the participants
• grounding
• authority
• settlement
• capacity
• quality
• economics
• risk
• simulation
• organizational change
• realization