For some years now, the debate on artificial intelligence and work has revolved mostly around estimates: how many jobs will be replaced, how many transformed, how many made redundant. These assessments are useful for understanding which professions are most exposed to automation, but they describe only part of the problem. The question of how much work can be replaced should be accompanied by another, probably more important one: where will the productive capacity made available by these technologies go?
Anyone who uses these tools daily can already see considerable productivity gains in some activities. Tasks that used to take hours can be done in much less time. This additional capacity does not necessarily translate into a corresponding reduction in employment. It can be used to produce new services, cut response times, improve quality, clear backlogs, or make economically viable services that previously were not. Of course, it can also be used to cut staff and raise margins. Technology makes both options possible, while the choice between them depends on organizations, on the structure of markets, and on the rules within which those markets operate.
Demand is not fixed
Some of the more pessimistic forecasts implicitly assume that the amount of work to be done remains essentially unchanged. If automation allows five people to do what used to require ten, the other five become redundant. The reasoning is correct if output stays the same, but it is precisely this condition that cannot be taken as a general rule.
In many sectors, actual or potential demand exceeds organizations' capacity to respond. Healthcare systems have waiting lists; courts accumulate pending cases; public administrations struggle to guarantee adequate response times; infrastructure maintenance is often postponed for lack of resources. Many private services too, from customer support to specialist consulting, are delivered in smaller quantities or at lower quality levels than users would ask for if they cost less.
There is therefore a considerable amount of unmet demand. Part of it remains unmet not because the need is absent, but because the cost of satisfying it exceeds the available resources or users' willingness to pay. In these cases, a productivity increase can be absorbed, at least in part, by expanding services rather than by reducing employment.
However, we need to distinguish need from economically sustainable demand. A waiting list, for example, proves the existence of an unmet need, but it does not guarantee that the resources freed by automation will be directed toward meeting it. This distinction matters because it keeps us from turning an economic possibility into an automatic prediction.
There is also a second phenomenon. Reducing the cost and the time needed to access a service can change the demand for it. Someone who gave up on a consultation because it was too expensive may decide to request it when the price drops; someone who avoided a procedure because it was too complicated may undertake it when it becomes simpler; someone who rarely used an unreliable service may turn to it more often when its quality improves. Economic history offers many cases in which falling unit costs and greater accessibility led to an expansion of consumption. It is not a universal law, but it is a mechanism important enough to make any employment forecast based on the assumption of constant demand implausible.
From doing to directing
The increase in productive capacity also changes how activities are distributed between people and automated systems. In many applications human involvement is not eliminated, but it changes in nature. Part of the work shifts from direct execution toward framing the problem, monitoring, verification, and decision-making. A worker can thus supervise a volume of activity that they could not have carried out personally in the same amount of time.
This shift, however, must not be used to dodge the employment problem. If one person can oversee work that previously required ten, the labor needed for that process goes down. Any compensation has to happen elsewhere: through a greater number of services, new projects that have become economically viable, greater personalization of services, or higher quality levels. The presence of unmet demand makes this compensation possible, but it does not guarantee it.
Experience with generative systems also shows how much the quality of the result still depends on the skills of the person using them. A photographer who understands light, composition, and the characteristics of lenses has the categories needed to describe the desired image precisely and to recognize the flaws in the one produced. An experienced programmer can assess not only whether the generated code works, but whether it has architectural problems, vulnerabilities, inappropriate dependencies, or maintainability issues. A lawyer can spot a clause that is incompatible with the applicable regulations or with the client's interests. Similar considerations apply to the systems administrator, the designer, the physician, and many other professions.
The issue is not simply the ability to write effective prompts. In non-trivial problems, formulating a request correctly already presupposes having the conceptual categories of the discipline. Those who do not know the subject struggle not only to evaluate the answer, but also to define the problem adequately.
Artificial intelligence therefore tends to reduce the amount of expertise needed for the mechanical execution of some tasks, without eliminating the expertise required to frame and evaluate them. In certain contexts it can even increase that expertise's value, because a competent professional can apply their judgment to a far greater amount of work than they could have produced directly.
To this we must add the question of responsibility. Signing a deed, making a diagnosis, or approving a project does not just mean producing a technically correct result. It also means taking responsibility for the consequences. An automated system can prepare documents, process information, and draft proposals, but legal and organizational responsibility must still be assigned to identifiable parties. The continued presence of humans in these steps therefore stems not only from the technical limits of automation, but also from the institutional structure within which decisions take effect.
Gains don't distribute themselves
The existence of unmet demand does not, however, solve the problem of where the freed capacity goes. An organization that increases its productivity can use the advantage to increase the quantity or quality of its services, lower prices, cut costs, or combine these options. The choice depends on the incentives within which it operates.
In markets with real competition, part of the productivity gains can be passed on to users through lower prices, higher quality, or additional services. A company that uses automation solely to cut its costs may initially gain a competitive advantage, but that advantage tends to fade when the same technologies become available to other operators. In such a situation, the organizational ability to use the tools better can matter more than mere access to the technology.
The mechanism is different in markets with strong concentration, high exit barriers, or little choice for users. It is different even more so in public services, where citizens normally cannot choose a competing administration to turn to. Without sufficient competitive pressure, higher productivity does not necessarily lead to better service.
In these contexts, different instruments become important: minimum service levels, measurement of processing times, publication of results, and clear assignment of organizational responsibilities. Without such mechanisms, a significant share of the benefits of automation can be absorbed by cost reduction without producing a corresponding increase in the quantity or quality of services.
This lets us avoid two opposite interpretations that share the same determinism. The first attributes to technology the autonomous power to destroy employment; the second assumes that innovation spontaneously generates enough new activity to offset what is eliminated. Both overlook what happens between technical innovation and economic outcome: the decisions of companies, the structure of markets, institutions, and public policy.
A test for organizations
Asking whether artificial intelligence is, in general terms, good or bad for employment is therefore likely to be of little use. It is more interesting to observe how different organizations will use the productivity gains obtained with the same tools.
Some will be able to use the extra capacity to expand their services, make previously too-costly activities sustainable, reduce response times, and tackle backlogs that available resources never allowed them to clear. In these cases the competitive or organizational advantage will derive not simply from owning the technology, which will in all likelihood become progressively more accessible, but from the ability to integrate it with professional expertise, adequate procedures, and effective control systems.
Other organizations may instead prioritize cost reduction, keeping the quantity and quality of services essentially unchanged. This too is an economically rational choice under certain conditions, and there is no intrinsic property of artificial intelligence that prevents this outcome.
It is precisely for this reason that the relationship between artificial intelligence and employment cannot be deduced from the technical capabilities of the systems alone. Knowing how many activities can be automated is not the same as knowing how many jobs will be eliminated. Between the two quantities come demand, prices, investment choices, competition, regulation, and the way organizations decide to use the productivity they gain.
The most interesting question, when facing an AI adoption project, is therefore not only what share of activities can be automated. It is where the capacity made available by automation will go.
If it is used mainly to produce the same amount of services with fewer people, the effect on work will inevitably be negative. If it makes it possible to produce services that are not delivered today, to meet demand that is currently priced out, to raise quality, or to enable new activities, a substantial part of the productivity gain can instead be absorbed by the expansion of output.
Technology determines how much becomes possible to produce with a given amount of labor. It does not determine how much a society will decide to produce. It is in the space between these two quantities that a significant part of the effects of artificial intelligence on employment will play out.