For most of business history, producing more meant hiring more people.
Ten accountants produced roughly ten accountants’ worth of accounting. An eleventh added capacity. The relationship was not perfectly linear, but it was reliable enough to shape budgets, departments, careers and management itself.
AI weakens that relationship because it changes where capability lives.
Once a piece of work has been abstracted—described, tested and made transferable—it can be executed repeatedly without recreating the same human capability for every instance.
The people who performed that work can begin to look less like a team and more like copies of one capability.
This is the deduplication of labour.
Deduplication is not simply another word for layoffs. A layoff is an employment decision. Deduplication is the structural pressure that appears when a capability no longer needs to be instantiated in the same number of people.
Nor is it the same as abstraction. Abstraction changes the form of the work. Deduplication is one possible consequence for the organisation around it.
The distinction matters because apparent duplicates are not always waste.
Two people who can do the same work may provide cover for illness, a second opinion, contestability, institutional memory or the ability to notice that an automated system is confidently wrong. Repetition may also be how novices become experts. A profession’s ordinary work is often its training ground for extraordinary judgment.
What looks duplicated in an organisation chart may be capacity, resilience or apprenticeship.
The serious task is to distinguish accidental duplication from deliberate redundancy.
A database can remove duplicate records without consequence once the correct record is known. An organisation cannot treat people this way. It has to decide which repeated capability was waste, which was protection and which was carrying a future the current workflow did not show.
When AI releases human capacity, the business receives a dividend. It can spend that dividend in three broad ways.
It can cut. Similar output is produced with fewer people and the saving moves to the cost line.
It can grow. The people remain while the organisation serves more customers, produces more work or enters areas that were previously uneconomic.
It can deepen. Output stays closer to its previous level, but more capacity goes into checking, resilience, difficult cases, relationships and the quality that volume had crowded out.
Most businesses will combine the three. The important distinction is whether the allocation is chosen or merely allowed to happen.
Cut is the easiest result to measure and therefore the easiest to mistake for the whole strategy. A smaller wage bill appears quickly. The loss of judgment, resilience or succession may remain invisible until the business needs it.
Grow is attractive where demand can expand. If cheaper and faster production creates a larger market, capacity released from repetition can be redirected into reach. But not every market grows simply because more can be supplied. Producing twice as many internal reports does not create twice as much value.
Deepen is less dramatic. It uses the dividend to make the work more trustworthy rather than more plentiful. In consequential work, this may be the most valuable use and the least legible on a productivity dashboard.
These choices also change organisational shape.
A conventional department gathers many people with related capabilities and adds managers to coordinate their work. If AI performs much of the repeatable volume, the department may shrink into a compact unit: a small number of people defining the work, handling exceptions, governing quality and developing the next abstraction.
The management layer can change with it. Some management exists because a large group of people requires coordination. Reduce the repeated human workload and part of the coordination job may disappear before anybody formally redesigns the role.
This is why deduplication does not always begin with a restructure.
It can begin with one person quietly becoming much faster. Someone learns to use AI well enough to absorb work that previously justified several positions. No executive programme is announced. The team’s effective shape changes from underneath, one private workflow at a time.
By the time the organisation notices, the capacity decision may already have been made socially. Work has migrated towards the person with the strongest abstraction, while titles, accountability and development paths remain where they were.
Unmanaged deduplication is not neutral. It rewards private leverage while leaving the organisation to discover its new dependencies afterwards.
There is also a labour-market version. If a capability is abstracted across an industry, demand can fall for a category of labour rather than for a role inside one company. The work does not need to disappear entirely. It needs to stop requiring the same number of people to perform each instance.
The evidence is still uneven, and the distinction between exposure and elimination matters. The International Labour Organization’s recent work on generative AI describes transformation as more likely than complete replacement across many occupations. The World Economic Forum’s employer survey similarly contains both intentions: organisations expect automation and some workforce reduction, while reskilling, augmentation and movement into new roles remain more common responses.
That ambiguity is not a weakness in the idea. It is the allocation problem made visible.
A capability can be deduplicated without every person attached to it being removed. Their capacity can be cut, grown into new demand, deepened into better work or redirected into a different capability. Technology creates the possibility; management determines the distribution.
Apprenticeship deserves special care in that distribution.
The repetitive tier of professional work was rarely designed as an education system, but it often functioned as one. Junior accountants learned where the numbers go wrong by seeing many ordinary accounts. Junior lawyers developed judgment through routine documents. Remove the routine work and the organisation still needs a way to produce experienced people.
An abstraction can preserve today’s expertise while quietly consuming tomorrow’s supply of it.
Resilience has the same problem. A second pair of capable eyes looks inefficient until the first pair misses something. Spare human capacity looks redundant until demand changes, a model regresses or a supplier fails. The fact that a capability can be concentrated does not prove that it should be concentrated completely.
Deduplication also hollows out accountability. A department did more than hold labour. Its reporting lines supplied a rough answer to who was responsible for what. A compact team working through AI does not inherit that answer automatically.
If ten people become two people and a system, someone must decide who can authorise the work, who checks it and who answers when it fails. Otherwise the labour is deliberately deduplicated while accountability is accidentally left behind. That problem continues in Outsourcing Accountability.
The old production law was simple: more output requires more people.
The emerging law is less comfortable: more output may require more computation and better direction without requiring proportionally more labour.
That does not make people unnecessary. It changes the argument for each additional person. Capacity, contestability, relationships, resilience, judgment and the production of future experts become reasons that must be made explicit rather than hidden inside the old assumption that volume itself requires headcount.
A business will deduplicate some work deliberately and some by drift. The test is not whether every repeated role survives.
The test is whether the organisation knows what it is removing, what it is retaining and what it intends to do with the capacity it has released.