The original forecast is on the record and there is no point dressing it up: within five to ten years, only 20% or 30% of knowledge workers would be needed to carry out 100% of the work being done then. A fivefold gain in productivity, coming from intensive adoption of generative AI by organizations.
That was a hypothesis, not a measurement — and the difference matters more than the number. No part of that reasoning came from counting jobs; it came from looking at what the technology appeared to enable and projecting the line forward.
Nor are we going to settle that account now. The net job balance of a transition like this is not something a software vendor measures, and anyone who says they have measured it is worth doubting. What has changed, and what can be stated with some confidence, is the question.
What the forecast got right: knowledge work came first
The point that has aged well is the most counterintuitive of all. For decades the assumption was that automation would climb from the bottom up — first the repetitive and manual, then, one day, the intellectual.
It happened the other way around. McKinsey's 2023 study, The economic potential of generative AI, showed the shift in numbers: the automation potential of the activity "processing data" rises from 73% without generative AI to 90.5% with it; that of "applying expertise" jumps from 24.5% to 58.5%. Occupations in “Educators and workforce training” and “Office support” record significant increases in automation potential when powered by generative AI — the source does not break out a percentage for them.
These numbers are worth reading for what they are: automation potential for activities, not for jobs. Nobody lays off an activity.
The reason for the inversion is mundane. Manual tasks demand physical dexterity and adaptation to a messy world, something robotics still struggles to replicate. Knowledge work is already digital end to end — it did not have to be converted in order to be reached.
What the forecast missed: capacity is not output
The mistake in the original reasoning — and it was the standard mistake of the time — was treating productivity as an automatic consequence of capacity. If the machine does the work, the work comes out.
It does not. A team where half of it never sleeps does not become more productive on its own. It becomes busier, which is a different thing, and the difference shows up in the first week: reports nobody asked for, three versions of the same material, excellent work done on the wrong priority.
What converts capacity into results is banal and old: someone defines the deliverable, the deadline and the scope, follows what was done and assesses it. That is management. It always was.
This gets concrete in the product we built, and it deserves saying without retouching: an AI worker with no current work plan does work — but its work belongs to no deliverable. It produces, and nobody can add it up. The worker itself is instructed to declare this in its report, because orphan output presented as progress is worse than no output at all.
Supervising is not surveillance
"Supervision" tends to be heard as distrust, which is why it is done badly. Supervising is four things, and none of them is looking over someone's shoulder:
- Scope
- saying which deliverables that worker answers for in the period — and what is left out.
- Deadline
- saying when, with a date somebody took on, not a vague expectation.
- Evidence
- having a way to know what was done without asking. It is not the executor's promise; it is the record.
- Assessment
- closing the period with an explicit judgement, with justification, that stays on the record.
A human gets all four out of company habit. An AI gets none of them by default — and that absence is the explanation we see most often for AI pilots that impress in the demo and vanish by the next quarter.
The role that changes most is not the executor's
The original version of this text imagined how a government would reorganize itself around AI, in layers. One of them said: a layer of specialized civil servants will monitor and review the interactions between virtual agents and citizens, making sure the systems work as intended and pointing out where they need to improve.
That was written as a projection about the public sector. It was the first thing to become reality, and not only in government. When an organization puts AI to work executing, the work that appears immediately is not the executor's — it is the work of whoever defines what it does, reviews what it produced and answers for the result.
Notice what this does to a manager's career. The mechanical part of the role — chasing status, compiling the report, remembering who owed what — is exactly the part the machine takes. What is left is scope, priority, exception and judgement. In other words: the role does not shrink, it concentrates on what has always been its reason to exist.
And it becomes scarcer, not less so. Someone has to sign the plan, decide the exception and assess the cycle. That someone is human. (We wrote separately about what this does to the management layer in Managers are not becoming obsolete. Manual management is.)
The same artifacts, for those who sleep and those who don't
The choice we made is deliberately unoriginal: an AI worker goes through the same ritual as a human, in the same artifacts, on the same screens. No parallel "AI" dashboard, no logs and traces as a form of accountability.
It has a name, a role description written by you, a closed set of tools, a work plan tied to a team, accountability cycles and an assessment at the end of each cycle.
| Artifact | Person | AI worker |
|---|---|---|
| Work plan signed by both parties | Yes | Yes — the manager's signature is enough, and the worker counter-signs on the spot |
| Deliverable outside a signed deliverables plan | Goes through, with a recorded warning | Blocks: the signature is refused, with the list of violations |
| Plan that points to deliverables but no task | Warning, and you carry on | Prevents signing |
| Cycle report | The participant writes and submits it | The worker writes and submits it |
| Cycle assessment, from 1 to 5 stars | The manager is the one who assesses | The manager is the one who assesses — the worker never assesses its own cycle, and the system refuses the attempt |
| Current plans at the same time | Can have more than one | Only one |
The asymmetry in the first two rows is intentional. A person who breaks the rule usually has a legitimate reason and answers for it; an agent has no such judgement — and a worker nobody can hold to account is the worst possible combination.
The same principle governs what the worker can do in the world. Every action is classified by code before it happens and routed down one of three paths: it runs on its own, the AI manager reviews it, or it is held waiting for you. Two default choices say the rest: a tool with no declared classification is treated as if it were contact with a new recipient — the default is to close, not to open — and a review that fails for any reason rejects the action, never releases it by omission.
And the evidence has to hold up in the worst case, which is AI being wrong about what it itself did. A send that failed enters the audit trail as a failure, not as sent. A reply the system decided not to give in an authorized group also becomes a line, because explained silence is better than mysterious silence.
What this does not solve
Nothing here makes AI reliable enough to do without someone supervising it. Our thesis is the opposite: it becomes part of the team because it becomes supervisable, and supervision is human work that still gets paid for.
Daily autonomous execution — the worker waking up, pulling the next task from the plan and working without anyone asking — is in gradual rollout, enabled organization by organization. Outside of it, it works when it is called on.
Two classes of high-risk action, financial and VIP contacts, exist in the matrix and have their rules enforced, but today no tool is automatically classified into them: what is missing is a deterministic signal for "this is money" and "this one is a VIP". They are ready for when there is one.
And the big questions from the original text remain open, because they are not for a product to answer: what to do about those who are displaced, how to reform curricula and training, whether some form of basic income enters the conversation. A vendor that answers those is selling, not thinking.
Conclusion
The question "will AI take the jobs?" presupposed a swap — a chair changing occupant. What you see in operation is something else: a team that gains a kind of participant that does not sleep, does not forget and does not know on its own what the priority is.
That participant does not replace the management structure. It makes it mandatory. The first job AI creates is supervising it — and the role that changes most is not the executor's, it is the manager's.