Faced with artificial intelligence, two extremes appear that feed each other: the fascinated leader, who wants to automate anything that is technically possible and measures progress in converted roles, and the defensive leader, who uses every AI error as proof that no important process should change.
They look like opposites. They commit the same error: both start from an ideology about the resource, neither from the result. The narrative of people versus machines sells a lot and directs very little, because a responsibility does not need an ideological winner. It needs a result.
Deciding what to automate with AI requires a method that is immune to enthusiasm and to fear. In the book AI Employee it is called the HWFA, Hybrid Workforce Fit Assessment: the fit assessment for the hybrid workforce, free at hybridwf.com. Before going through it, a boundary has to be drawn that the instrument cannot cross.
First the ethical boundary
The neutrality the discipline requires lends itself to a dangerous misunderstanding: it is neutrality about the type of resource, not moral neutrality. The standard behind the book declares it in HWF-01, the only clause with supremacy over all the others: people are ends; software is a means. The optimization of cost, speed, or capacity never overrides anyone's rights, dignity, safety, or labor protections. In quotable text: "This clause takes precedence over every other clause of this standard."
Said in management terms: first the ethical boundary is drawn, and inside that boundary you optimize without sentimentality. Neither defending a human role out of romanticism, nor handing over a responsibility out of enthusiasm when the decision deteriorates rights or human experience.
The false dilemma, seen up close
The book illustrates it with a composite case: a company does two things that look alike on the outside and are nothing alike on the inside. Classifying service requests: thousands a month, known categories, abundant data; a badly routed case gets corrected and the cost was a delay. And conducting conversations with clients who threaten to cancel: few per week, loaded with anger, history, and commitments. An error there sometimes cannot be reversed with anything.
"Can AI do it?" will probably receive a yes in both cases, and it is useless because it does not distinguish between the two jobs. "Which resource should answer for the result?" separates the answers by itself: in the classification, an artificial resource executes most of it and escalates the ambiguous ones; in retention, AI prepares context while a person keeps the conversation and the final decision.
There is also an empirical reason: the capability of the models advances in patches (the Stanford HAI AI Index describes it as a jagged frontier: a medal in mathematics olympiads, failures at reading an analog clock), so the only actionable question is about a concrete responsibility, with its risk and its exceptions.
Three stages, thirteen dimensions, four outputs
The most important design decision of the HWFA is that it does not return a number. A score allows you to launder through arithmetic a decision that was already made, choosing weights until the formula agrees with whoever already decided. That is why the HWFA returns an argument: an assignment, a risk class, an initial autonomy, the answers that decided the result, and the conditions that would change it.
It runs in three stages, and the order is doctrine. First eligibility, because there are questions that economics has no right to answer. Does the responsibility decide reserved matters: material effects on employment, health, credit, or access to essential services, legal rights, or the treatment of vulnerable people? The decision remains human, as a floor that no assessment can lower. Is the work completely enumerable, with fixed rules and rare exceptions? The honest answer is deterministic software of all time. An instrument that knows how to say "do not use AI here" is an instrument you can trust when it says "here you can".
Then risk: how much the error costs, whether it can be undone, how many people it reaches. And a warning: high volume is not an argument for automating but an amplifier of risk, because at scale the same error rate reaches more people. At the end it asks about economics, only among the configurations that survived the two previous stages.
Thirteen dimensions cover the three stages, from reserved matters to the cost profile. The possible outputs are four: Human, Assisted human, Deterministic automation, or Artificial. A role is occupied by a single resource; the escalation route toward a person is a control, not an identity.
A real responsibility, resolved
The book walks through a complete example with its running collections case, with hypothetical figures declared as such: payment reminders and early follow-up on accounts 5 to 45 days past due. 85% of the cases follow the same pattern, some 1,200 accounts flow per month, exceptions hover around 15%, and there is no reserved matter: no service cutoffs are decided and there is no dealing with people in declared insolvency.
Reading of the instrument: eligibility passed, Moderate risk, economics favorable to the artificial in the repetitive stretch. Result: Artificial, with the escalation boundary at the exceptions: every response that is not a payment confirmation goes out to the analyst with the complete file. Initial autonomy: rung 3, execute with approval, because no responsibility starts above 3. How you go up from there is the subject of the autonomy ladder.
The real discussion was not about the average but about two dimensions: how much judgment each case demands and how many cases fall outside the pattern. Two competent people can estimate 15% or 30% of exceptions, and that difference moves the split: with almost a third of the volume ending up with a person, the responsibility becomes assisted human. That disagreement is not a defect of the instrument. It is its most useful product: the discussion happens over the correct dimension.
When the best worker is not just one
Many good results need two resources; what cannot have two occupants is a role. The combination is put together in one of two ways: an artificial role with an escalation route, where the AI Employee absorbs volume and a person receives whatever falls outside its limits, or an assisted human role, where the person keeps exceptions, relationships, and high-impact decisions. Both are legitimate. What is not legitimate is not choosing: the result has a name, two occupants waiting for the other one to answer.
There is also false automation: the system executes many steps, but a person reviews each one because nobody trusts the design. The activity looks artificial and the cost is still human, hidden in a dashboard that shows the machine's actions and not the reviewer's hours.
The class does not go down because you added controls
The HWFA leaves a risk class in your hands: the dial of all the later obligations, from the autonomy granted to the depth of the audit. The standard defines five classes. Prohibited is not a result of the analysis but a door: uses that no control makes acceptable, such as deceiving people about what the system is. Critical, High, Moderate, and Low grade the rest according to seven factors, not only the cost of the error: rights affected, scale, reversibility, vulnerable people, data sensitivity, adversarial exposure, and concentration of power.
The distinction that prevents the games is the one between inherent and residual risk. The class is evaluated twice, before and after the controls, and it follows the inherent one: the controls lower the residual, the class does not move. Without that rule, every control added would be an argument for reclassifying downward, and the controls would serve to dismantle the controls. A Critical role with excellent controls is a well-governed Critical role. It is not a Moderate role.
Without this rule, every conversation with a vendor ends in "we put a guardrail on it, so now it is low risk". HWF-51 requires the double evaluation and closes that door: controls earn a better residual, they do not buy a better class.
With the boundary drawn, the assignment argued, and the class declared, what remains is the central instrument of the implementation: the role contract. The method already did its job. Nobody decided by ideology.
The correct resource is not the most modern one. It is the one that produces the result best.
Frequently asked questions
The HWFA is the instrument of the WRM framework that decides which resource should occupy a responsibility: human, assisted human, deterministic automation, or artificial. It evaluates each concrete responsibility (not departments or complete roles) in three stages whose order is doctrine: eligibility, risk, and economics, over thirteen dimensions. It is an instrument of structured judgment, not a validated test, and it can be used for free at hybridwf.com.
Because a number allows you to launder through arithmetic a decision that was already made: choosing weights until the formula agrees with whoever already decided. The HWFA returns an argument: the assignment, the risk class, the initial autonomy, the answers that decided the result, and the conditions that would change it. Two managers can conclude differently because they operate under different risks, but both can explain their conclusion.
They are the matters whose decision remains human regardless of the rest of the analysis: material effects on employment, health, credit, or access to essential services, legal rights, or the treatment of vulnerable people. Clause HWF-02 specifies the scope: what is reserved is the decision, not everything else. The analysis, the drafting, and the recommendation can be delegated to an artificial resource; the decision itself, never. That is why a responsibility with a reserved decision almost always lands on assisted human, and a reserved decision is Critical by definition.
No. The class is evaluated twice, as HWF-51 requires: inherent risk before the controls and residual risk after them. The controls reduce the residual, but the class follows the inherent one and does not move. The reason: if every control justified reclassifying downward, supervision and auditing would go down along with the class. A Critical role with excellent controls is still Critical, only well governed. And a Prohibited use does not become compliant through any control.
The instrument presupposes a role already taken apart into responsibilities: that method is in Work Resource Management. The next step is to formalize the role in a role contract.
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.