Imagine an executive meeting in the near future. The team reviews results produced by people and by AI Employees. One figure draws attention and somebody asks who authorized the decision that generated it. Technology explains which agent acted. Operations describes the workflow. The vendor confirms the model worked according to its configuration. And nobody can answer who was responsible for granting that authority or who should have suspended it.
The silence does not reveal a technical gap but an administrative one. The agent is technically integrated and administratively orphaned: it does not appear in the org chart, it has no visible manager, and its authority lives dispersed in configurations that only an engineer knows how to read.
The future of management is directing work through human and artificial intelligence without confusing them. That demands new instruments and a managerial competence that somebody has to exercise.
The org chart no longer tells the whole story
The traditional org chart does its job well as long as the work follows only human lines. When artificial resources sustain recurring responsibilities, it stops describing the real operation: the organization knows who reports to whom, but it cannot show which artificial resource holds which responsibility, which manager answers for it or which decisions it can make.
A hybrid map must answer three questions: who holds each responsibility (human, assisted human or artificial), who keeps the accountability, and where an exception escalates to. It must also show dependencies: an AI Employee can belong to Collections, depend on data from Finance and use a system from Technology. If those relationships live only in the head of one person, continuity is fragile.
The chains can also grow upward: an AI Employee can supervise another. What never travels down the chain is the accountability. Every supervision pyramid ends in an identified human or in a human governance body, capable of seeing the complete chain and intervening at any point without asking the pyramid for permission. Clause HWF-22 of the standard demands that form: walkable and observable end to end.
The role and its mission
Here the hybrid workforce manager appears. Not necessarily a mandatory position: a new managerial competence that somebody must exercise in every organization that grants custody of roles to artificial resources.
His mission is to guarantee that each responsibility is executed by the configuration that produces the best result, considering performance, cost, quality, risk, continuity and human impact. In practice: he coordinates the governed transitions between people and artificial resources, designs the borders between roles, reviews role contracts, makes sure every resource has someone responsible for it, and sustains the governance calendar. He administers, repeated over time, the underlying decision of the HWFA: human or AI?.
Neutrality defines the role, and its incentives define the neutrality. If his success is measured by the amount of AI implemented, he will force automations; if it is measured by human positions preserved, he will block necessary improvements. That is why the role carries an anti-KPI, the metric by which it must never be measured: number of humans replaced or percentage of roles converted to AI. Its true KPIs look at the result: performance delta, cost per correct result, risk, continuity and elevation of human capability.
It is worth stopping on that last point. The elevation of human capability asks what new value the time that automation frees up produces. The book separates six possible destinations, from reinvestment to headcount elimination, each one with an owner and a way of being verified: calling everything "productivity" prevents you from learning what really happened.
Whoever answers for artificial resources needs seven things: competence over the domain, effective authority, time, direct access to the evidence, independence from the pressures that own the result, the real power to stop the system and training against automation bias. If his bonus is tied to the throughput of the resource, he should not be the only authority capable of stopping it.
In exchange, the standard protects whoever accepts the position from the trap that would make it unacceptable: when something fails, the blame follows the causes, not the proximity.
A human placed at the end of a process is not an absorption zone for the blame that belongs upstream: to the design, to the policy or to the vendor.
Without that rule, accepting the position would be accepting to be the designated defendant.
Drawing the hybrid organization
The central instrument of the role is a map. Select a department and draw its main results. Under each result, responsibilities, not names. Identify who holds each one (human, assisted human, deterministic automation or artificial) and add the responsible manager, the level of autonomy, the handoffs and the destination of the exceptions.
Then look for signals of fragility. Two owners for the same outcome? An AI Employee without a manager? An exception without a destination? A continuity that depends on a credential only one person knows? And if a person reviews every action of an agent, the autonomy the design declares does not exist: it is automation sustained by hand, and the autonomy ladder has not been built.
Do not try to transform everything in one session: choose one responsibility and correct its architecture.
When AI Employees work as a team
The map has a second reading that almost nobody does: the lines between artificial resources. Those chains add speed, reach and opacity to risks that management already knew: they can delegate work among themselves many times before a human looks at a single one.
The standard names five risks. Circular delegation: A delegates to B, B to C, and C returns the case to A; the work orbits and nobody owns the fact that it never landed. Feedback cycles: the output of one feeds the input of the other and a small bias compounds on every turn. Memory contamination: a wrong conclusion enters the shared context and the downstream systems inherit it as a fact. And lateral propagation of authority: on each handoff decisions that were already made arrive and the receiver treats them as authorized, and a resource ends up exercising authority that nobody granted it.
The fifth is the most important: correlated failure. Human diversity tends to distribute blind spots; the instances that share model, vendor or configuration concentrate them and can fail in the same way at the same time. If your collections agent, your quality reviewer and your router run on the same model, your "review" shares the exact blind spot of the thing it reviews. The mitigation is the one for monoculture in agriculture: diversity where the risk warrants it, and a fallback for when the common vendor has a bad day.
One border of the role is not technical but ethical: the workforce it directs must not fake humanity. The standard says it in a single line: "An AI identity must never deceive." Courtesy and linguistic empathy remain legitimate as long as they do not assert interiority. "I am sorry about the situation you are going through" is professional courtesy; "I know how painful this is" is a claim of experience the system does not have.
The last resource you manage: yourself
After designing roles, granting authority and measuring results, there remains a variable no framework manages for you: your relationship with learning. A leader can have sufficient data and still defend an automation because it was his idea, or reject a capability because he does not understand it yet. The system that is hardest to correct is usually the identity that needs to look expert.
I say it with my own example, because asking for it without having paid for it would be cheap. Since 2018 I have not written a line of my company's code. I programmed from the age of fifteen and the craft that defined me is practiced today by others better than me. Directing resembles standing in front of an orchestra: the movement of the baton produces no sound at all. The hardest part was not letting go of the keyboard. It was accepting to decide without having all the information, because one hundred percent of the data always arrives late.
There is a difference it would be dishonest not to mark. When you delegate to a person, you delegate to her judgment: somebody who can read a context nobody explained to her and who can refuse. An artificial resource applies declared criteria inside the context and the authority you gave it: what trust resolves with a person has to be written down with it. You can delegate the execution of almost everything. You cannot delegate the judgment that governs it.
Frequently asked questions
He guarantees that each responsibility of the organization is executed by the configuration that produces the best result (human, assisted human or artificial), considering performance, cost, quality, risk, continuity and human impact. He coordinates the governed transitions between people and artificial resources, designs borders between roles, reviews role contracts, makes sure every resource has someone responsible for it and sustains the governance calendar. It is not obligatorily a new position: it is a managerial competence that somebody must exercise wherever custody of roles is granted to artificial resources.
The metric by which it must never be measured: number of humans replaced or percentage of roles converted to AI. That indicator creates a double perverse incentive: it rewards the conversion instead of the result and it guarantees that the person in charge of protecting the quality of the decision gets paid for prejudging it. Its true KPIs look at the result: performance delta, cost per correct result, risk, continuity and elevation of human capability.
The standard names five and demands them as part of the validation (clause HWF-52): circular delegation (the case orbits among resources and never lands), feedback cycles (a small bias compounds on every turn), memory contamination (an error enters the shared context and everyone inherits it as a fact), lateral propagation of authority (unauthorized decisions are treated as authorized on each handoff) and correlated failure (instances that share model or vendor fail in the same way at the same time).
Yes: it can route its work, review its outputs and receive its exceptions. What never travels down the chain is the accountability: every supervision pyramid ends in an identified human or in a human governance body, capable of intervening at any point without asking the pyramid for permission. Clause HWF-22 demands it: a supervision chain must be walkable and observable end to end. Whoever answers but has to ask the pyramid for permission has a petition, not control.
For the decision this role administers daily, read HWFA: human or AI?. For the complete cycle that governs every role, read No role lasts forever.
Want the full method? Read AI Employee. For executive AI consulting or keynotes and workshops.
Go deeper
Want to bring your team to the next belt?
Book a discovery call or explore the full book.