Year five of any given automation. Someone asks in a meeting whether the system still makes sense and the answer arrives without a pause: "it works, and nobody has time to touch it right now". Both things are true. Neither is a justification.
The interesting part is how it got there, because no organization decides to keep an obsolete system. You get there step by step. And you get out with a discipline that almost nobody applies to their artificial resources: the AI lifecycle, the same idea that good management always knew about humans. The role does not belong to its occupant. It belongs to the result.
The complete cycle includes designing, testing, operating, measuring, developing, remediating, expanding, reducing, reassigning, suspending and retiring. Eleven verbs, not three. Automating does not eliminate management: it makes it continuous.
The forever trap
The typical chronology goes like this. Year one: a process is automated and it works; whoever pushed the decision gains credibility. Year two: the first exception the design did not foresee appears and it is resolved with a rule, because a patch costs an afternoon and redesigning costs a quarter. Year three: the exceptions are already eleven, and one person dedicates two hours a week to accommodating strange cases that nobody accounts for as a cost of the system. Year four: the market changes, the automated process stopped being the one the operation needs, and the engineer who designed it already works at another company. Year five: the meeting from the first paragraph.
Look at it in sequence and something becomes clear: there was no moment at which the correct decision was obvious. Each individual step was reasonable. The whole is not. That is why the discipline does not propose better decisions, it proposes a calendar, because the problem was never one of judgment: it was that the question had no date.
The trap does not consist of having automated. It consists of confusing a good past decision with a future obligation.
One distinction sustains everything else: stability is not immobility. A stable system can evolve without every change being a fire. An immobile system only conserves its shape, even though the value diminishes.
The role must re-justify its existence
Human roles without sense sometimes prune themselves: a salary line that the budget review questions, a resignation that forces a decision on whether the replacement is justified. An artificial role eliminates even those weak triggers. Its cost lives fragmented among licenses, inference, integration and human review, so it rarely appears as a line that somebody has to defend. And nobody ever resigns from it. An artificial role without sense does not only persist by default: it loses the last occasions on which somebody would notice it. What accumulates is called organizational debt, and its most dangerous form has a name of its own: the digital zombie, the automation that keeps operating with valid credentials and authority, working for nobody.
The vaccine is a review different from all the others. It does not ask about the efficiency of the role: it asks about its existence. Is this responsibility still necessary in the face of the current strategy?
"It meets its metrics" sounds like a reason not to touch a role. It says nothing about whether it should exist. That is why the existence question goes first, with a calendar of its own, and continuity is never the default: the role is re-justified or it is retired.
The standard behind the book, in its clause HWF-63, sets two triggers for that review: a cadence never greater than twelve months and a declared window after every material change of strategy, policy, regulation, product or structure. A role can end up misaligned the day after a change of course. And it brings a lock against self-deception: a modified scope re-justifies itself on its own merits, it never inherits the previous justification, and work invented to preserve a role is a failed justification, not a redesign. The closing of the clause is quotable: "retirement is the result of a failed justification, never of a missed calendar".
The most uncomfortable test: you
A principle that only applies downward is not a principle, it is a policy. Michael Gerber formulated it in E-Myth Revisited in a way that lands badly the first time: the objective of building a company is that it comes to function without its founder. Not that he leaves. That he can leave.
Read from the lifecycle, it says something sharper: indispensability is not an achievement, it is a design defect. And we almost always celebrate it because it resembles merit. The person without whom a certain process does not walk, the only one who knows why that exception exists, the one who does not take full vacations because "everything falls apart". That is called commitment and it is usually rewarded. It is also a single point of failure with a name, a history and a good reputation.
One precision avoids perverting the argument: making a role not depend on a person is not making the person dispensable. It is the opposite. The indispensable collaborator is trapped: he cannot be promoted, nor get sick, nor take on a new project, because the system needs him exactly where he is. Freeing the role from its dependency frees him as well.
The road runs in two directions
Many AI strategies imagine a one way road: from the human toward the artificial. The discipline recognizes the return. If error increases, supervision becomes excessive, the customer experience deteriorates, regulation changes or the numbers get worse, giving custody back to people can be the most responsible decision.
The most public case is Klarna. In February 2024, the Swedish fintech announced that its AI assistant had handled 2.3 million conversations in its first month: two thirds of all its customer service, in 23 markets and more than 35 languages, the equivalent of the work of some 700 full time agents. For more than a year it circulated as the example of successful automation. In May 2025, the CEO himself, Sebastian Siemiatkowski, reversed course in public: the cost cutting had gone "too far", and Klarna was hiring human agents again so that the customer would always have the option of speaking with a person. According to the press coverage of the turn, customers complained about generic answers and about a system that did not handle complex cases well.
What Klarna did not do is equally instructive: it did not return to the previous model. It rebalanced the work. AI keeps absorbing the routine volume and human agents attend the escalations, the complex cases and the higher value accounts. The reversal does not always return the role to its previous form; sometimes it produces a better distributed workforce.
The operational lesson: define the conditions of return before the pilot. What error rate turns out to be unacceptable? How much human intervention destroys the economics of the role? What incident forces a suspension? Written before knowing the result, those rules reduce the bias of defending the implementation out of pride. Written afterwards they are not criteria: they are justification.
Retirement is executed in full
When the role is indeed retired, the work that almost everybody omits remains: the offboarding. Revoking credentials, disabling tools, stopping routines and queues, rotating secrets, transferring the pending work, preserving the evidence for audit. The exit is as important as the onboarding. The standard says it without ornaments: a retired role whose accesses survive it is not retired, it is unattended.
The capacity that is left free
When an AI Employee takes on a responsibility well, the human hours dedicated to it stop being necessary. The efficiency indicator registers the saving, but it does not explain what happens afterwards. And that is where the most important decision of the entire transition lives: that destination can be growth and it can be cuts, and the word "productivity" is comfortable enough to hide either of the two. Technology can free up hours. It cannot decide their purpose.
That decision no longer belongs to the system: it belongs to the leadership. And it demands a management that directs people and artificial resources at the same time: the Hybrid Workforce Manager.
Frequently asked questions
It is the discipline of managing every role occupied by an artificial resource from its design to its retirement: designing, testing, operating, measuring, developing, remediating, expanding, reducing, reassigning, suspending and retiring. Its central principle: the role does not belong to its occupant but to the result. Continuity is never the default: every role re-justifies its existence with a calendar, it can be reversed toward people if the result demands it, and when it is retired, it is retired in full, with accesses revoked. Automating does not eliminate management: it makes it continuous.
The Hybrid Workforce Standard, in its clause HWF-63, sets two triggers. First, a regular cadence never greater than twelve months. Second, a declared window after every material change of strategy, policy, regulation, product or structure. The review asks about the existence of the responsibility, not about the efficiency of the resource: a role can meet all its KPIs and still not deserve to be kept. And a modified scope re-justifies itself on its own merits: it never inherits the previous justification.
When the result demands it: if error increases, supervision becomes excessive, the customer experience deteriorates, regulation changes, the numbers get worse or the context becomes too variable. Klarna did it in 2025: after automating two thirds of its customer service, it hired human agents again for escalations and complex cases, without abandoning AI on the routine volume. The key is to define the conditions of return in writing, before the pilot. A mature hybrid organization does not measure success by the direction of the transition: it measures it by the quality of its architecture.
It is the orderly exit of an artificial role that is retired: revoking credentials, disabling tools, stopping scheduled routines and queues, rotating secrets, transferring the pending work and preserving the evidence the audit will need. It is the difference between retiring a role and creating a digital zombie, because a retired role whose accesses survive it is not retired: it is unattended. Clause HWF-64 of the standard sums it up: offboarding must revoke accesses and transfer or destroy context securely.
For the extreme case of what happens when this cycle does not exist, read Digital zombies. For the role that manages these transitions, read Hybrid Workforce Manager.
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.