How Is AI Changing IT Roles? What the Evidence Shows in 2026

Aggregate IT headcount has held up. Entry-level hiring has not. The distance between those two facts is the decision most IT leaders have not yet made.

Published by MASSIVUE, an enterprise AI transformation and capability-building firm. Last reviewed: August 2026. Reading time: about 11 minutes.

The short answer

AI has not produced widespread job displacement in IT. What it has changed is who gets hired. The clearest measured effect in the labour data falls almost entirely on early-career workers in occupations where AI substitutes for human tasks, and software development and technical support sit near the top of that exposure ranking. Experienced workers in the same occupations show no comparable gap.

For an IT organisation that produces a specific and uncomfortable situation. The routine work agents absorbed first was also the work that trained the next generation of senior engineers. Reducing the junior intake looks like an efficiency gain on this year's budget and reappears as a capability shortage three to seven years later, at which point it cannot be solved by hiring, because every comparable employer made the same decision in the same window.

The useful response is not to freeze headcount, and not to protect roles that no longer have work attached to them. It is to separate two questions most organisations answer as one: which work has moved to agents, and which roles should therefore change. Those have different answers, and the second is a design decision rather than an automatic consequence of the first.

Key takeaways

• There is no evidence of economy-wide job displacement from AI in the best available payroll data through June 2026.
• Employment of workers aged 22 to 25 in AI-exposed occupations stands 19 percent below where it would be had it kept pace with less-exposed peers. Experienced workers show no equivalent gap.
• The adjustment runs through reduced hiring rather than separations, which is why it is nearly invisible from inside an organisation.
• Declines concentrate where AI substitutes for tasks. Where AI complements the work, employment is flat or rising.
• Case deflection is not a headcount number. ServiceNow deflects roughly 75 percent of its customer cases with a support team no smaller than two years earlier.
• The strategic risk is the loss of the apprenticeship path, not the loss of jobs.

What the evidence actually shows

The strongest current evidence comes from Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at Stanford, whose paper Canaries in the Coal Mine? was updated in August 2026 using administrative payroll records from ADP covering millions of United States workers through June 2026. Administrative payroll data matters here because it avoids the two weaknesses that make most AI-and-jobs commentary unreliable. Job-posting counts measure intent rather than employment. Surveys measure belief.

Their first finding is the one most often dropped from summaries: they find no evidence of widespread, economy-wide job displacement. Overall employment remains robust.

Their second finding is where the story is. Employment of young workers aged 22 to 25 in AI-exposed occupations now stands 19 percent below where it would be had it kept pace with that of their less-exposed peers. Experienced workers show no comparable gap. In absolute terms, employment of 22 to 25 year olds in the two most exposed occupation quintiles fell about 11 percent between November 2022 and June 2026, while employment of the same age group in the three least-exposed quintiles grew about 10 percent.

Three further details change what an IT leader should do about this.

It operates through hiring, not firing. The divergence comes primarily from reduced hiring of young workers rather than increased separations. Nobody was marched out. The graduate intake was simply smaller each year, which is exactly the kind of change that never triggers a governance conversation.

Substitution and complementation point in opposite directions. Declines concentrate in occupations where AI usage primarily substitutes for human tasks. Where usage primarily complements workers, employment is flat or rising, and rising fastest for experienced workers. Exposure alone does not predict the outcome. How the tool is deployed does.

Pay did not absorb the shock. Adjustment is occurring through employment rather than base compensation. Salaries did not fall. The number of openings did.

Is it really AI? The honest dispute

This finding is contested, and an article that hides that is not worth reading.

In January 2026 Zanna Iscenko and Fabien Curto Millet published Looking for the Ladder through the Economic Innovation Group, arguing the diagnosis is premature. Their central objection is about timing. The decline for young workers in exposed occupations begins in November 2022, immediately after the public release of ChatGPT, and by June 2023 roughly half of the observed decline had already materialised. They argue it is implausible that firms across the economy evaluated AI, redesigned workflows, cleared security review and changed hiring at national scale within six months of a consumer chatbot launching. The more parsimonious explanation, they suggest, is the sharpest monetary tightening cycle in four decades.

That is a serious argument, and for much of 2026 it was the strongest available counterweight.

The August 2026 update to the Stanford paper answers it directly. The authors report that the divergence has continued to widen through mid-2026, long after interest rates peaked, that it persists when controlling for occupational exposure to interest-rate increases, and that it survives excluding technology firms, excluding computer occupations, and accounting for remote work. They also note that by November 2022 the relative position of exposed occupations had already returned to approximately its pre-pandemic level, so the subsequent decline moves the gap below that baseline rather than merely back toward it.

They are careful in a way most coverage is not. They state plainly that these are descriptive facts rather than causal estimates, that the pattern attenuates when controlling for occupational education levels, and that the divergence is more pronounced in the ADP sample than in national survey benchmarks.

The practical reading for an IT leader is that the cause is still being argued and the consequence is not. Whether the missing junior cohort is explained by agents or by interest rates, the cohort is missing, and the capability it would have become is missing with it.

Where AI substituted, and where it did not

Because substitution and complementation lead to opposite employment outcomes, the deployment question stops being a tooling decision and becomes a workforce decision. Two findings are worth holding alongside each other.

The first is that measured productivity gains are less reliable than they feel. In a randomised controlled trial published in July 2025, METR gave 16 experienced open-source developers 246 real tasks in their own repositories, randomising whether AI assistance was allowed. The developers completed tasks 19 percent slower with AI available. They estimated afterwards that AI had made them 20 percent faster. The authors are explicit about the limits: this is a snapshot of early-2025 tooling in one setting, with experienced contributors working in codebases they know well, and it does not establish that AI fails to speed up most developers. What it does establish is that self-reported speedup is not evidence of speedup, which matters when a headcount decision rests on it.

The second is that the organisation, not the tool, determines the return. The 2025 DORA State of AI-assisted Software Development report from Google concludes that "AI's primary role in software development is that of an amplifier. It magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones." A team with unclear priorities and a brittle legacy architecture does not become a good team with agents. It produces its existing output faster.

Both point the same way. If AI is deployed into a well-run system as a complement, the evidence associates that with stable or growing employment and better work. If it is deployed as a substitute for a headcount line, the gain is assumed rather than measured, and the capability cost lands later.

MASSIVUE has examined the same disconnect on the delivery side, where faster building has not translated into better outcomes, in Why AI Hasn't Improved Product Outcomes.

Deflection is not a headcount number

The most common error in IT operations planning right now is treating a deflection rate as a staffing calculation. If agents resolve 70 percent of tier-one tickets, the reasoning goes, tier-one staffing can fall by something close to 70 percent.

The clearest counterexample comes from a company with every incentive to claim otherwise. Speaking on Bloomberg Live in August 2025, ServiceNow President and Chief Operating Officer Amit Zavery said the company deflects roughly 75 percent of its customer cases while its support team is no smaller than it was two years earlier. Over the same period case volume rose about 40 percent. The freed capacity went to complex cases and to customers who needed a person, customer satisfaction moved in a positive direction, and retention on the team improved.

That is a customer service organisation rather than an IT service desk, and the analogy should not be pushed further than it goes. But the mechanism is the one that matters, and it is general. Automating the simple half of a queue does not remove half the work. It removes the half that was easy, raises the average difficulty of everything left, and usually reveals demand that the old queue was suppressing. The people who remain are doing harder work than before, which is an argument for investing in them rather than for counting them down.

The real problem: the apprenticeship pipeline broke

The Stanford authors offer a mechanism that deserves more attention than the headline number. Employment declines for young workers concentrate in occupations that involve codified knowledge, the kind of knowledge that can be written down, taught from a manual and checked against a known answer. Occupations that involve tacit knowledge see faster employment growth for experienced workers.

Now consider what entry-level IT work has always consisted of. Tier-one tickets with a documented resolution path. Runbook execution. Standard configuration changes. Boilerplate code and test scaffolding. Log triage against known signatures. All of it codified. All of it precisely what current agents do competently.

Here is the part that rarely gets said out loud. That work was never valuable primarily as output. It was valuable as training. A junior engineer closing three hundred routine tickets is not producing three hundred tickets' worth of value. They are building the pattern library that lets them recognise, four years later, that an incident does not look like the incident it appears to be. Judgment in IT is tacit knowledge accumulated by doing large volumes of codified work under supervision.

Agents have taken the codified work. They have not taken the judgment. But the judgment was a by-product of the codified work, and the organisation was getting it for free.

This is why the labour data shows senior workers doing well. It is not evidence that seniority is safe. It is evidence of scarcity, and the scarcity gets worse on its own, because the pipeline that replenishes it has been narrowed at the intake end while demand for judgment rises.

The hiring market is already reacting to the symptom. Gartner's October 2025 predictions for IT organisations include two that read very differently once the mechanism is clear. By 2027, Gartner expects 75 percent of hiring processes to include certifications and testing for workplace AI proficiency. And through 2026, Gartner expects atrophy of critical-thinking skills from generative AI use to push 50 percent of global organisations to require "AI-free" skills assessments, isolating a candidate's unaided reasoning.

Employers are building tests for unassisted judgment at the same moment they are removing the work that used to produce it. Both moves are individually rational. Together they describe a system that has stopped manufacturing its own seniors.

Which IT roles change, and how

The honest version of this section contains no invented job titles and no salary figures. What can be said with confidence is which work has moved and what that does to the shape of a role.

One pattern runs through every row. The residual human work is disproportionately review and judgment under uncertainty, and that is the work that requires the most experience. The composition of an IT organisation shifts toward tasks its shrinking junior intake is least equipped to do and slowest to learn.

What IT leaders should do now

Six decisions follow from the evidence above. They are ordered by how quickly they become irreversible.

1. Measure intake, not attrition. The adjustment runs through hiring. If your workforce reporting shows headcount, attrition and vacancy rates but not the size and seniority mix of your annual intake, you cannot see the thing that is changing. Add a standing metric for entry-level hires as a share of total hires, tracked over three years rather than quarters.

2. Separate the work question from the role question. Establish which tasks agents have genuinely absorbed, evidenced by measured outcomes rather than vendor deflection rates or self-reported speedup. Then decide, as a separate deliberate act, what each affected role should become. Collapsing these two into one step is how organisations cut a role because a task moved.

3. Stop converting deflection into headcount. Require that any proposed reduction justified by automation states what happened to the residual work, what the new average case difficulty is, and who now owns the agent's failure modes. If those cannot be answered, the saving is not established.

4. Rebuild apprenticeship deliberately, because it will not happen by accident. The old curriculum was volume of routine work. The replacement has to be constructed. In practice that means giving juniors structured exposure to the review path rather than the execution path: supervising agent output and being accountable for accepting it, sitting in on incident command before leading it, and being handed the escalations the agent could not close rather than the tickets it could. This is more expensive per head than the model it replaces. It is also the only mechanism that produces the senior capability the same organisation is bidding for in the market.

5. Segment roles into upskill, reskill and transition, explicitly. Not every affected person has the same path. A capable tier-one analyst whose queue has been automated may be an excellent agent supervisor, a poor developer, and unemployable in the role they currently hold. Making that assessment role by role, with stated criteria, is uncomfortable and considerably better than the alternative, which is a hiring freeze that silently makes the decision by attrition.

6. Test for unassisted judgment before the market forces you to. If Gartner is right that half of organisations will require AI-free assessment through 2026, the organisations that benefit are the ones that already know which of their people can reason without assistance, and have developed rather than merely measured it.

Where MASSIVUE fits

MASSIVUE is an enterprise AI transformation and capability-building firm. The part of this problem we work on sits between the deployment decision and the workforce consequence, which is where most of the damage described above is done quietly.

Our AI Workforce Transformation service covers exactly this ground: AI skills assessment and gap analysis, AI role design and organisational restructuring, and capability development delivered through MASSIVUE Academy. The role design element is the one relevant here. It is the difference between deciding what a service desk analyst becomes and letting the vacancy rate decide.

Where decision rights over AI are the underlying issue rather than roles, that belongs to an operating model question. Our Protum™ AI Operating Model addresses it, and the reasoning is set out in What Is an AI Operating Model?

To be clear about attribution: the employment figures above belong to Stanford, the Economic Innovation Group, Gartner, METR, Google's DORA programme and ServiceNow, and are attributed at the point of use. What is MASSIVUE's own contribution is the synthesis: the argument that entry-level IT work was functioning as an apprenticeship mechanism rather than as output, the role-change table, and the six decisions above, drawn from advisory work on AI workforce transformation.

Limitations of this article

The core employment evidence is United States payroll data. The direction is likely to generalise to comparable labour markets, but the magnitudes should not be assumed to hold in Singapore, the European Union or India, where hiring norms, graduate supply and employment protection differ materially.

The Stanford findings are descriptive rather than causal, by the authors' own statement, and the Economic Innovation Group critique of the causal story has not been fully resolved. The METR trial involved 16 developers on early-2025 tooling and should not be read as a general productivity finding. The ServiceNow figures are a single company's account of its own performance, given by an executive of that company, and are reported here as an illustration of a mechanism rather than as a benchmark. Gartner's 2026 and 2027 figures are predictions, not measurements.

The role-change table reflects the current state of agent capability. It is the section of this article most likely to age quickly.

Frequently asked questions

Is AI reducing IT headcount?

Not at the aggregate level, on the best available evidence. Administrative payroll data covering millions of United States workers through June 2026 shows no widespread, economy-wide job displacement. What it shows instead is a sharp divergence by age within AI-exposed occupations: workers aged 22 to 25 are employed 19 percent below where they would be had they kept pace with less-exposed peers, while experienced workers show no comparable gap. The change is in hiring, not in separations.

Which IT roles are most exposed to AI?

Exposure concentrates where the work is codified: tier-one support with documented resolution paths, routine infrastructure monitoring and remediation, boilerplate and test code, standard database optimisation. But exposure alone does not predict job loss. The evidence shows declines where AI is deployed to substitute for tasks and flat or rising employment where it is deployed to complement people, so how the tool is introduced matters more than which function it touches.

Should we stop hiring junior engineers?

Reducing the junior intake is the decision with the longest and least reversible consequence. Entry-level work in IT functioned as an apprenticeship: juniors acquired tacit judgment as a by-product of high volumes of routine work. Agents have absorbed the routine work but not the judgment, and demand for judgment is rising. If you reduce the intake, the replacement development path has to be built deliberately, through supervised review of agent output and structured exposure to escalations, rather than assumed.

Does AI make developers faster?

Sometimes, but self-reported speedup is not evidence. In a randomised trial by METR published in July 2025, 16 experienced open-source developers working on 246 real tasks in familiar codebases were 19 percent slower with AI available, while estimating they had been 20 percent faster. The authors caution against generalising from one setting and early-2025 tooling. The practical implication is narrow and important: do not base a headcount decision on perceived productivity gains.

If our AI agents deflect 70 percent of tickets, can we cut tier-one staffing by 70 percent?

No, and the assumption is the most common planning error in IT operations today. Deflection removes the easy half of a queue, which raises the average difficulty of everything remaining and often surfaces demand the old queue suppressed. ServiceNow reports deflecting roughly 75 percent of customer cases with a support team no smaller than two years earlier, over a period when case volume rose about 40 percent. Any reduction justified by automation should have to state what happened to the residual work and who owns the agent's failure modes.

How do we know whether our AI deployment is actually working?

Measure outcomes rather than activity, and measure them against a baseline taken before deployment. Google's 2025 DORA research concludes that AI acts as an amplifier, magnifying the strengths of high-performing organisations and the dysfunctions of struggling ones, which means adoption rates and tool usage tell you very little on their own. Useful signals include change in end-to-end resolution or delivery time, change in rework and escalation rates, and whether review capacity has become the constraint.

Related MASSIVUE resources

• What Is an AI Operating Model? covers who funds, owns, governs and stops AI, which is the decision-rights layer beneath the role changes described here.
• How Do You Manage Change During Enterprise AI Transformation? addresses the delivery problem created by AI programmes that have no go-live date.
• Do You Need a Chief AI Officer? examines executive accountability for AI, one level above the IT organisation.
• AI Upskilling Companies in Singapore compares named providers and funding routes for organisations acting on the capability gap.
• Why AI Hasn't Improved Product Outcomes looks at the same speed-versus-outcome disconnect at team level.

Sources

Each figure above was checked against the publisher's own material at the last review of this article. Where a source is preliminary, single-firm or narrowly scoped, that is stated at the point of use.

• Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford Digital Economy Lab, August 2026. ADP administrative payroll data through June 2026. Authors state findings are descriptive, not causal. https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/
• Zanna Iscenko and Fabien Curto Millet, Looking for the Ladder: Is AI Impacting Entry-Level Jobs?, Economic Innovation Group, January 2026. https://eig.org/wp-content/uploads/2026/01/TAWP-Iscenko-Millet.pdf
• Gartner, Gartner Unveils Top Predictions for IT Organizations and Users in 2026 and Beyond, 21 October 2025. Figures cited are predictions. https://www.gartner.com/en/newsroom/press-releases/2025-10-21-gartner-unveils-top-predictions-for-it-organizations-and-users-in-2026-and-beyond
• METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, 10 July 2025. Randomised controlled trial, 16 developers, 246 tasks. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
• Google, DORA 2025 State of AI-assisted Software Development Report, 2025. Amplifier finding quoted from the published abstract. https://research.google/pubs/dora-2025-state-of-ai-assisted-software-development-report/
• Amit Zavery, President and Chief Operating Officer, ServiceNow, on Bloomberg Live, August 2025, as reported by CX Today, 1 September 2025. Single-company account. https://www.cxtoday.com/contact-center/servicenow-hasnt-cut-its-customer-service-headcount-despite-deflecting-75-of-cases/