The AI employment mirage
Artificial intelligence promises extraordinary improvements in productivity. But the argument that the jobs created by implementing AI will compensate for the jobs AI eliminates deserves considerably more scrutiny. History suggests that technology implementation booms are temporary; the people displaced by them are not necessarily the people who benefit.
We are in favour of artificial intelligence. We are not in favour of pretending that its economic benefits will automatically translate into employment opportunities for those whose jobs it replaces. AI can make businesses more productive, improve services, reduce costs and eliminate tedious work. These are genuine benefits. But when a customer service department replaces twenty employees with an AI agent, the resulting savings are not twenty new jobs waiting to be discovered elsewhere. Nor is the consultant who installs that agent guaranteed a lifetime of work.
The first casualties are not necessarily the best placed to adapt
Consider a customer service operation. A business can use an AI agent to answer routine enquiries, check orders, process straightforward requests and escalate exceptions. It may then need fewer customer service employees. Whether any particular deployment actually reduces headcount depends on demand, service quality, cost and management decisions; nevertheless, reducing the labour required per enquiry is often central to the commercial case.
The employees at risk are frequently in relatively modestly paid roles. Their work may provide an accessible entry into employment, predictable hours or a foothold in a career. Telling them that AI is creating jobs in machine learning, agent architecture or systems integration is scarcely a credible remedy. Those occupations require different skills, experience and opportunities. Even where retraining is possible, it costs time and money, and the new work may not be available locally.
The same problem reaches the small business. The plasterer may still be needed to prepare a wall and finish it beautifully, but the person who manages the diary, sends invoices, chases payments and answers the telephone may be replaced by a subscription. We can celebrate the reduction in paperwork while regretting the disappearance of a perfectly respectable job. The two sentiments are not contradictory.
The implementation boom is real; its permanence is not
Upwork reported in February 2026 that demand for AI integration skills rose 178% year on year, AI chatbot development 71% and AI data annotation and labelling 154%. These figures are evidence of a lively implementation market. They are not forecasts of how many human specialists businesses will require once implementation becomes routine. Upwork also reported continued demand for conventional human skills, an important counterweight to sweeping claims of immediate mass unemployment.
A business commissioning its first agent may require discovery, workflow mapping, integration, testing, security review and staff training. The second business can often reuse the design. The hundredth may purchase a packaged product and configure it in an afternoon. Once deployed, the agent needs maintenance, oversight and improvement, but not necessarily another large project.
There is a further twist: AI is increasingly used to write code, configure integrations, test workflows and document systems. The same technology generating implementation work can reduce the human hours required to perform it. Demand for AI services can therefore rise while demand for AI implementation labour eventually falls.
We have seen this film before
Y2K created an intense market for auditing, correcting and testing computer systems against a fixed deadline. The exceptional demand did not persist after the deadline passed. Windows rollout programmes required deployment engineers, compatibility testing and hands-on support; standard images and centralised management gradually reduced much of that effort. Exchange Server demanded installation, maintenance, backup and upgrade expertise; hosted email concentrated many of those tasks with providers. Microsoft 365 and other cloud office migrations generated another wave of consultancy, followed by more routine ongoing administration.
These histories are analogies, not statistical proof that AI employment will follow an identical timetable. They do, however, expose a recurring error: treating the temporary labour required to adopt a technology as though it were a permanent characteristic of using that technology. Email usage increased even as the work of maintaining individual email servers diminished. AI adoption could likewise increase while the market for building bespoke agents contracts.
Even continuous improvement may not save the consultants
It is sometimes argued that businesses constantly change, so agents will constantly need specialists to rebuild them. Yet many changes are small: an approval limit, a new field, a revised invoice rule or a different reporting format. An adequately designed agent may accommodate such changes with limited human involvement, subject to controls. Indeed, if an agent continuously identifies waste and suggests improvements, it may reduce the need for periodic external process redesign.
More frequent change does not necessarily mean more paid work. If a company once commissioned twenty changes at forty specialist hours each, it bought 800 hours. If it later makes fifty changes at two human hours each, it buys 100 hours. That is an illustrative 87.5% fall in labour demand despite doing more improvement work.
Not everything is a minor adjustment. Acquisitions, legislation, contested priorities, safety critical decisions and genuinely new business models can demand substantial human judgement. But the existence of such work is not evidence that it will absorb everyone displaced elsewhere.
The plumber, the carpenter and the plasterer
Physical trades offer a useful contrast. A plumber must work in an unpredictable physical environment, diagnose faults and make a safe repair. A carpenter must deal with actual materials, imperfect buildings and practical tolerances. A plasterer needs skill, coordination and physical presence. Robotics may eventually change these occupations, but automating their full range of tasks is not equivalent to generating an email or answering a routine customer enquiry.
This is not an argument that trades are immune to technology, nor that everyone should retrain as a plumber. It is an argument for distinguishing digital tasks from embodied work and for being realistic about how quickly different occupations can be substituted. And the trades themselves are not untouched: the carpenter may retain the workbench while losing the administrator who organised the appointments. That loss deserves attention too.
What the evidence establishes, and what it does not
The International Labour Organization estimated in 2025 that one in four workers globally held occupations with some exposure to generative AI. It stressed that exposure is not the same as redundancy, and that most jobs are more likely to be transformed than eliminated outright. Clerical occupations were among the most exposed. The OECD likewise identifies automation, new tasks and productivity as competing channels whose combined employment effects remain uncertain.
These findings caution against declaring that a particular number of jobs will vanish by a particular date. They also caution against the opposite certainty: that new AI jobs will automatically replace lost ones. The Upwork marketplace is informative about freelance demand, not a representative census of the entire labour market. Its reported growth rates do not establish the duration of the boom or its net employment effect.
Our thesis is therefore a risk assessment, not a numerical forecast: implementation employment may peak, standardisation may compress it and the workers most readily displaced may be poorly positioned to capture the temporary gains.
The policy question is distribution, not whether AI should exist
The sensible response is not to oppose AI. It is to ask who receives the productivity gains and who bears the adjustment costs. Employers can invest in genuine redeployment, offer training tied to actual vacancies and share some of the benefits of increased productivity. Governments can strengthen practical transition support, assess regional and occupational impacts and avoid presenting speculative future occupations as a substitute for present livelihoods.
We should also measure the right things. Count not just advertised AI jobs or the number of agents deployed, but paid human hours per implementation, maintenance hours after deployment, changes in customer service staffing, wage progression, job quality and the outcomes of displaced workers. Track cohorts over several years. Otherwise, a short-lived boom in expensive technical contracts may be mistaken for an enduring solution to unemployment.
There may be entirely new markets that we cannot yet foresee. That possibility deserves serious consideration. But a possibility is not a plan, and hope is not an employment policy.
Conclusion: embrace the technology, abandon the comforting arithmetic
AI can be a remarkable economic advance while still causing painful employment losses. Those statements can both be true. It is entirely reasonable for a business to adopt a useful agent, and equally reasonable for society to ask what happens to the people whose wages that agent replaces.
There is an opportunity for specialists now: learn the tools, understand the business problems and take part in the implementation wave. But do not assume that today's scarcity of AI skills guarantees tomorrow's demand. Standard products, reusable agents and automated implementation are precisely what successful technology markets tend to produce.
The plumber may still have a leaking pipe to fix. The carpenter may still have a door to hang. The plasterer may still have a wall to finish. But the plasterer's administrator may no longer have a job. We should not dismiss that loss with the cheerful suggestion that she become an AI architect.
We can be enthusiastic about artificial intelligence without being careless about the people it leaves behind. Indeed, if we want the technology to succeed on terms that society can sustain, we must be capable of both.
Companion article: Read the shorter tabloid commentary, The AI Jobs Miracle? Don't Bet Your Wages on It.