AI is changing the IT labour market. Who gets hired next?

The boom ended. AI arrived. Now the industry must decide what greater productivity is for, and how the next generation gets a start.

The next phase of the IT labour market may be shaped as much by jobs that never open as by people who lose theirs.

Picture a junior vacancy that a manager decides not to open. A departing engineer whose work the team absorbs. A client project that now needs fewer billable hours. None requires a machine to replace a whole profession. Each can change someone's career.

That is the quiet part of this story. AI does not need to make developers obsolete to make the labour market harder to enter. It does not need to destroy demand for software to change who earns a living from building it.

The reverse matters too. A difficult hiring market does not prove that IT employment is collapsing. Faster coding does not prove that companies need fewer people. And a recovery in job ads does not mean that every candidate has recovered with it.

The evidence available in November 2026 points to a market being reorganised, not one global market moving in a single direction. The EU still has a growing ICT workforce. US software job postings have recovered from a low point, but unevenly. Research raises serious concerns about young workers, while finding much less evidence of broad displacement.

The useful question is not whether AI is good or bad for IT. It is where it reduces the need for labour, where it makes labour more valuable, and who gets the opportunity to move between the two.

The correction did not begin with AI

We need to start before the chatbot.

The pandemic moved more activity online. Businesses expanded digital services and made plans around demand that did not always last. In the euro area, pandemic-era monetary policy supported funding conditions. When funding is easier to obtain, companies can justify projects that tighter budgets would rule out.

There is a revealing contemporary account. In his November 2022 message to Meta employees, Mark Zuckerberg said he had expected the surge in online commerce to continue. It did not. He also cited the macroeconomic downturn, competition and changes to advertising signals. That is one company's explanation, not a complete history of IT. It is still a warning against rewriting the entire correction as an AI story.

The financing environment changed as well. By September 2023, the European Central Bank described tighter financing conditions as dampening demand. More expensive funding changes which investments clear a company's hurdle. It can make a planned expansion less attractive without any change in what an engineer can do.

AI entered a market already dealing with this reset. Hiring plans, client budgets, interest rates and expectations were changing together. Outsourcing decisions and the location of delivery teams added another layer: a role can move to another country without disappearing worldwide.

These forces can reinforce each other. A company under margin pressure may adopt AI to contain costs. It may also cancel a weak product, consolidate suppliers or move work. A later headcount cut cannot, on its own, tell us how much each decision contributed.

Nor can the sources used here say how much of the international IT downturn AI caused. Claiming a single figure would turn a complex adjustment into a convenient story.

A smaller hiring market is not the same as a smaller workforce

Several different measures get compressed into the phrase “the tech jobs market”.

Employment counts people in work. Hires count people starting jobs. Vacancies are positions employers seek to fill. Online postings are job ads on particular platforms, not a census of vacancies or completed hires. Layoff announcements record planned cuts, not net employment after other companies recruit.

A tech company's headcount also includes people outside IT. An IT occupation can sit inside a bank, a hospital, a factory or a public authority. Cutting jobs at a software company and cutting the total number of software professionals are different things.

This distinction changes the picture in Europe. Eurostat's May 2026 release records 10.45 million ICT specialists employed across the EU in 2025, or 5.0% of total employment. Their number grew 2.6% from 2024 to 2025. Growth had slowed, but the workforce had not shrunk. The definition covers people who develop, operate and maintain ICT systems across the economy, not just employees of technology companies.

Now compare a different measure in a different geography. Indeed Hiring Lab's July 2026 analysis found that US software development postings had risen almost 15% since late February 2025. Yet in its series through June 2026, they were still about 27.5% below February 2020. Indeed's live index has risen further since: on 6 November 2026 it stood about 18.6% below February 2020.

Both facts can be true. A recovery can start well below the old baseline. A large existing workforce can coexist with fewer openings for people trying to join it.

The recovery was not evenly shared either. Senior roles accounted for 71% of the net increase in US software development postings between May 2025 and May 2026. Jobs with AI in their title accounted for 37%. The two groups overlap, so the figures cannot be added together. They describe contributions to the increase, not shares of all software jobs.

The timing does not establish that AI caused the rebound. It does challenge the claim that greater AI exposure can only mean falling demand.

Automating a task is not the same as removing a job

The ILO's 2025 global exposure study puts about one in four workers worldwide in occupations with some exposure to generative AI. Its highest exposure category accounts for 3.3% of global employment.

Neither figure is a forecast of unemployment. The study assesses the potential to automate tasks in existing occupations. It does not measure actual adoption, jobs already lost or future jobs created. It is a snapshot of capabilities at the start of 2025, not a live reading of today's tools.

A software role contains more than code production. Someone must work out what the user needs, navigate an existing system, decide which risks are acceptable, test the result and keep it working. AI can change several of those activities without taking responsibility for the whole result.

Yet “jobs are bundles of tasks” should not become a comforting slogan. If a team can deliver the same service with less labour, an employer may hire fewer people even when every remaining job still needs human judgement. The ILO explicitly recognises this possibility.

The labour-market effect depends on what happens next. Does the company keep the saving? Cut prices? Clear a backlog? Enter a market that was previously too expensive to serve? Improve quality? Each choice creates a different demand for people.

A productivity gain does not automatically turn into a redundancy or a new job.

Productivity is real, but the unit matters

The research does not support a universal multiplier for developer productivity.

A study published by Microsoft Research in June 2025 combined randomised field experiments involving 4,867 developers at Microsoft, Accenture and an unnamed Fortune 100 company. The experiments ran in 2022 and 2023. The researchers used randomly assigned access to isolate the effect of taking up the assistant. They estimated about 26% more weekly completed tasks, measured through pull requests, for developers who took up the tool because they were offered access. Less experienced developers showed larger gains.

That is evidence about task output in those workplaces with that tool. It is not evidence that a quarter of developers can be removed. The individual experiments were noisy, and completed tasks are not the same as revenue, customer value or a whole team's capacity.

A different experiment found a different result. In METR's early-2025 study, 16 experienced open-source developers completed 246 tasks in projects they knew well. Tasks took 19% longer when AI use was allowed.

That result is historical, not a verdict on present tools. In its February 2026 update, METR said later data gave an unreliable estimate of the current effect. Developers and tasks were selecting out of the experiment, including people who did not want to work without AI. The researchers thought tools were probably helping more than before, but could not reliably quantify the improvement.

The lesson is not to choose the experiment that confirms our instincts. Context, tool generation, familiarity and quality standards matter. So does the difference between perceived speed and measured results.

The September 2026 revision of “Writing Code vs. Shipping Code”, an NBER working paper on GitHub activity, makes a further distinction. It finds much larger increases in coding activity than in software releases after AI-tool adoption. It is an observational study, not a randomised trial, and public development activity is not the whole software economy. Still, its central question is exactly right: how much extra code becomes software that people actually use?

If writing becomes faster but review, integration or customer adoption remains slow, the bottleneck moves. Hiring against lines of code misses that change. So does cutting a team on the assumption that code generation measures its entire contribution.

The entry route is the most urgent problem

The strongest concern is not that no one will need engineers. It is that fewer employers may want to pay for someone to become one.

The August 2026 revision of Stanford's “Canaries in the Coal Mine?” study examines US ADP payroll records through June 2026. Among workers aged 22 to 25, employment in the two highest AI-exposure groups in the study stood 19% below where it would have been had it kept pace with less-exposed peers, using November 2022 as the baseline.

That is a relative shortfall in this dataset. It is not a finding that AI eliminated 19% of young people's jobs. The study covers exposed occupations across the economy, not just junior programmers. Age is not identical to seniority.

The gap emerged mainly through reduced hiring, not through more people leaving jobs. The authors also find no evidence of widespread, economy-wide displacement. Important caveats remain: some trends predate generative AI, education controls reduce the gap, and the pattern is stronger in the ADP sample than in national survey benchmarks. The researchers describe indicators, not causal estimates.

The result deserves attention without exaggeration. A labour market can close its front door before it empties its offices.

There is an apparent contradiction here. Less experienced developers gained more in the field experiments, yet younger workers face weaker employment outcomes in the payroll evidence. The findings measure different things. A tool can help a junior perform better while reducing an employer's willingness to create junior positions.

This is also a training problem. Routine tasks have often provided a supervised route into harder work. If those tasks change, employers need to redesign the route, not assume that expertise will appear fully formed elsewhere.

A better entry role could combine small production changes, debugging, test design and customer context under a named mentor. AI can assist, but the learner must explain the change, find its weaknesses and show how it behaves outside the happy path. Protected review time is part of the job's cost, not spare capacity to be found later.

Seniority helps only if it means more than years served

In the US postings data, the recent rebound favours experienced roles. It does not make every senior professional secure.

The useful advantage is accumulated judgement: knowing which requirement is incomplete, which dependency will fail, when a migration is unsafe and when a proposed feature is not worth building. A title alone does not demonstrate any of that.

AI may let an experienced engineer explore more options or deliver a wider piece of work. It may also increase the amount of generated material that needs review. Whether this creates leverage or overload depends on how the work is organised.

Managers face a similar test. Producing status summaries is not the same as setting priorities, resolving conflict or developing people. An organisation may choose fewer management layers. It may instead need more deliberate coordination as teams attempt more work. The evidence here does not establish a universal manager-to-engineer ratio for an AI-enabled company.

For hiring, the practical response is to test decisions rather than polish. Ask candidates to investigate a fault, review an AI-generated change or explain an architectural trade-off. State whether tools are allowed. Evaluate correctness, reasoning, risk and communication. These are recommendations, not a claim that all employers have already changed their selection process.

No job family gets an exemption

It would be comforting to divide IT into jobs AI takes and jobs it cannot touch. The evidence is not that tidy.

The ILO places software development, database administration, network work and ICT user support among exposed occupations. Infrastructure is not automatically safe because it is technical. Nor does exposure establish that these occupations will decline.

For software engineering, faster implementation can shift attention towards requirements, integration, maintenance and ownership. For testing, producing test cases is different from deciding which failures matter. The US Bureau of Labor Statistics' role description includes risk assessment, test planning, usability and communication, not just executing checks.

For product and design, a quicker prototype does not prove that the proposed product solves a worthwhile problem. Research, accessibility and choices about conflicting user needs remain distinct from producing an interface.

For data, generating a query is different from determining whether the underlying data is complete, lawful and fit for the decision. For security and infrastructure, generating a configuration is different from accepting its operational risk or recovering from an incident.

For IT support, drafting a standard answer is different from resolving an unfamiliar failure in a customer's environment. The relevant unit is a problem solved, not a reply produced.

These task distinctions are not a ranking of safe occupations. They show why task-level capability cannot settle the employment question. Work that remains difficult can still be concentrated in a smaller team. Or cheaper implementation can make more projects viable and create more demand for that expertise.

Even official projections are conditional. The BLS currently projects 10% growth in combined US employment of software developers, quality assurance analysts and testers from 2025 to 2035. That is a modelled outlook, not an observed increase, a guarantee for graduates or a forecast for Europe.

Romania should not borrow America's numbers

Romania is part of this transition, but US platform statistics cannot tell us how many Romanian engineers will be hired.

Eurostat puts ICT specialists at 2.7% of Romania's total employment in 2025. This measures an occupational share across the economy. It does not isolate outsourcing firms, reveal current vacancies or show how AI affects hiring.

The European Commission's 2026 Digital Decade country report describes strong fixed connectivity and a high number of ICT graduates. It also identifies weak business technology uptake, low basic digital skills and problems retaining ICT talent.

That combination matters. A country can produce technical talent while struggling to apply it across its own economy. There is work to do in business adoption and public services. Unmet need, however, is not the same as a funded contract or a job offer.

For service providers in Romania and other delivery markets, the commercial question is sharp. If a client buys hours for a defined output, faster delivery can reduce billable work. If the client buys a reliable outcome, a provider that delivers it well may win more business. Neither outcome is automatic.

Lower labour cost could remain an advantage. It could also matter less for tasks that require fewer hours, while language, domain knowledge, legal requirements and proximity to the client matter more. These are competing mechanisms to watch, not verified findings that AI has already ended offshoring.

The constructive response is not a national rebrand as “AI talent”. It is a stronger link between graduates, supervised work and customer problems, alongside evidence that providers can own delivery rather than simply supply capacity.

Three plausible markets from here

Our assessment is that uneven change is more plausible than a single global outcome. The following scenarios are conditional ways to read the next few years, not probability forecasts. They can coexist across industries and countries.

1. More output, fewer openings

Companies gain capacity but demand does not expand enough to absorb it. Budget discipline remains strong. Employers meet targets with existing teams, replace fewer leavers and reduce external delivery hours.

This does not require mass layoffs. Employment can weaken gradually through lower recruitment and attrition. Entry-level candidates and suppliers of routine work could feel the pressure first.

What would support this scenario? Sustained gains in delivered output alongside weak hiring, lower replacement rates, shrinking graduate intakes and reduced contractor demand. Postings alone would not be enough; actual hires and payroll employment would need to confirm it.

2. Cheaper delivery opens a larger market

Lower development costs make previously uneconomic projects viable. Businesses modernise old systems, smaller customers can afford useful software, and providers turn efficiency into new sales rather than only lower staffing.

This requires real customer demand, available budgets and reliable delivery. Generating more applications that nobody uses would not qualify. Nor would an AI investment boom that fails to produce paying customers.

What would support this scenario? Broader hiring beyond AI-titled senior roles, more completed and adopted projects, sustained customer spending and renewed entry routes. Growth would need to spread beyond a small set of well-funded companies.

3. A wider gap between workers, firms and regions

Demand grows in some places while access narrows elsewhere. Experienced people with domain knowledge gain scope. Standardised execution work faces price pressure. Some firms build strong learning systems; others recruit only people trained by somebody else.

This is our central concern, not an inevitable destination. It fits the possibility of growing total employment alongside a harder first job. It also leaves room for experienced workers to struggle when their skills or location do not match demand.

What would support this scenario? Persistent gaps in hiring by seniority, longer transitions into first professional roles, diverging pay within comparable occupations and a concentration of gains among particular employers or regions.

Pay is especially easy to overstate. The Danish “Still Waters, Rapid Currents” study, revised in March 2026, found no detectable average effect of chatbot adoption on earnings or recorded hours through December 2024. It still found work being reorganised. Those earlier Danish results do not refute later US warning signs; the countries, periods, tools and outcomes differ. They do show why we should not promise a general AI wage premium.

Employer intentions deserve the same discipline. The World Economic Forum's 2025 survey records plans for both AI-related recruitment and workforce reductions through 2030. Plans are not completed hires or cuts. They tell us what surveyed employers anticipate, not which scenario will win.

A better outcome has to be built

Candidates cannot control company budgets or repair the entry market alone. They can make their contribution easier to assess.

Build evidence around a problem, not a catalogue of tools. Explain what was wrong, how you investigated it, what you changed and how you checked the result. Include a failed approach and the reason you rejected it. If AI helped, identify where it helped and where you had to correct it.

For someone starting out, a small maintained project with tests, documentation and feedback is more revealing than a large unexplained demo. For an experienced professional, a clear account of a difficult decision can show more than a list of technologies. Do not disclose confidential employer or client material to make that case.

Employers have greater power, and therefore greater responsibility. Measure useful output, review time, defects and customer outcomes before turning tool adoption into staffing targets. Look at a full delivery cycle, not an impressive demonstration. Include training, oversight and rework in the cost.

Keep a deliberate route into the profession. Give junior roles defined learning goals, real work and paid mentoring capacity. Offer existing staff a credible route into changed responsibilities before deciding that every new capability must come from outside. Make hiring exercises resemble the work people will actually do.

The positive case for IT is not that everyone will be safe. It is that valuable software still requires choices about people, systems and consequences, and the evidence does not show those needs disappearing. EU ICT employment has continued to grow. Experiments show that less experienced developers can benefit from assistance. Research also shows that work can change before pay or employment changes dramatically.

Those are reasons to invest in the transition, not reasons to ignore its costs.

AI can make implementation less scarce. We can use that capacity merely to narrow the hiring gate. Or we can use it to solve more worthwhile problems and help more people become capable of solving them.

The better future for IT is not a return to indiscriminate hiring. It is a market that rewards sound judgement, creates useful work and still gives people a fair chance to learn. That outcome is not guaranteed. But it is within reach, through decisions employers, educators and professionals can make now.

Research cut-off: 8 November 2026; the Indeed live index figure was retrieved on 11 November 2026. Sources report different countries, periods and measures, identified in the text. Scenarios and practical recommendations are editorial analysis, not measured forecasts.