Accountants, Lawyers, Radiologists: The White-Collar Professions AI Is Quietly Dismantling

In 2016, Geoffrey Hinton, the computer scientist who would later win a Nobel Prize for his work on neural networks, made a prediction so confident it bordered on reckless. Within five years, he declared, artificial intelligence would outperform radiologists at reading medical images. Universities should "stop training radiologists now." The five-year mark came and went. By 2026, the Mayo Clinic had expanded its radiology staff by 55 percent, to four hundred physicians. Hinton, to his credit, publicly admitted he had been wrong. But the interesting thing about his failed prophecy is not that it was incorrect. It is that a version of it is quietly coming true, just not in the profession he singled out, and not in the way anyone expected.

Across accounting, law, and medicine, artificial intelligence is reshaping white-collar work with a selectivity that defies the simple narrative of machines replacing humans. The transformation is uneven, sometimes counterintuitive, and rarely as dramatic as either the doomsayers or the optimists would have you believe. But the data now arriving from major firms, hiring platforms, and workforce surveys tells a story that is impossible to dismiss: certain tiers of professional work are being hollowed out, while others are, paradoxically, becoming more valuable than ever.

The Apprenticeship That Disappeared

The most telling statistic in professional services right now is not a layoff number. It is a hiring number. In the United Kingdom, accounting graduate job listings fell 44 percent year-on-year, according to Adzuna, the labour market platform. That figure alone would be alarming. But drill into the firm-level data and the picture sharpens considerably. KPMG slashed its graduate intake by 29 percent, from 1,399 to 942. Deloitte cut 18 percent, bringing its cohort down from 1,700 to 1,400. EY reduced by 11 percent. PwC trimmed 6 percent. These are not rounding errors. They represent a structural contraction in the pipeline that has fed the accounting profession for decades.

The Big Four have not been shy about explaining why. PwC's UK leader explicitly cited generative AI as the reason for cutting approximately 200 entry-level roles, noting that AI was reshaping what mentors assigned to first-year staff. EY has delayed graduate start dates for three consecutive years; its 2025 hires will not begin work until March 2026. Meanwhile, PwC invested over one billion dollars in generative AI tools and training, and Deloitte committed three billion. KPMG pledged 4.2 billion in technology spending, with AI at its centre. The firms are not shrinking their ambitions. They are shrinking their need for the humans who once executed the routine work that justified those ambitions.

What does that routine work look like? Tax preparation, bank reconciliations, ledger analysis, compliance checks. These are the tasks that occupied the first three to five years of an accounting career, the apprenticeship through which juniors learned the profession's rhythms and judgments. Agentic AI systems now rolling out at Deloitte and EY can upload documents for tax returns, analyse financial statements, and flag anomalies without human intervention. Firms report that AI-advanced practices generate up to 39 percent more revenue per employee. The metric itself is telling: "revenue per employee" is replacing "billable hours" as the industry's north star, a shift that implicitly values fewer, more productive workers over larger teams.

The mid-career and senior ranks have not been spared, though the cuts there feel more surgical. PwC eliminated 1,500 U.S. audit and tax positions, about 2 percent of its workforce. KPMG cut 330 U.S. audit roles. EY removed 30 UK partners. Deloitte trimmed 180 UK advisory positions. PwC's global headcount dropped below 365,000 after shedding 5,600 workers in the first half of 2025 alone, quietly abandoning its 2021 pledge to add 100,000 employees by 2026. Yet the firm simultaneously announced it was creating a formal career track for engineers, data scientists, and software developers, and struggling to hire enough of them. The profession is not dying. It is being recomposed.

A Brief That Once Took Twenty-Five Hours

Harvey, the legal AI platform that has become something of a bellwether for the profession, was valued at three billion dollars in February 2025. By June, it was worth five billion. By December, eight billion. In early 2026, the company entered talks to raise capital at an eleven-billion-dollar valuation. That trajectory alone tells you something about where capital believes the legal industry is heading. Harvey's annual recurring revenue hit an estimated 195 million dollars by the end of 2025, nearly four times its figure from the previous year. More than 100,000 lawyers across firms like A&O Shearman, Latham & Watkins, Mayer Brown, and Paul Weiss now use the platform.

The productivity numbers are striking. According to an RSGI adoption study of 40 law firms, power users of Harvey save 36.9 hours per month; standard users save 15.7 hours. In-house legal teams report similar gains, with power users reclaiming 28.3 hours monthly. Thomson Reuters has documented cases where a legal brief requiring 25 hours of work can now be completed in 10, a 60 percent reduction in billable time. A midsize firm reported slashing contract review times by the same margin. Firms using AI across their operations have seen a 30 to 50 percent drop in overhead costs associated with manual legal processes.

The paradox, and it is a genuinely strange one, is that law firm profits soared 13 percent in 2025. Hourly rates climbed. Technology spending rose nearly 10 percent. Talent costs increased 8.2 percent. Rather than passing AI-driven efficiency gains to clients, many firms are treating the technology as a profit multiplier, completing work faster while maintaining or raising their billing rates. Nearly 60 percent of in-house counsel surveyed said they had seen "no noticeable savings" from their outside counsel's use of generative AI. The firms are getting richer. The clients are getting restless. And the junior associates who once billed those hours? Their ranks are thinning. The traditional pyramid model, in which large cohorts of junior lawyers generated revenue through document review and legal research, is under sustained pressure from clients demanding efficiency and AI accelerating the reimagination of that structure.

Six out of seven leading AI platforms, when asked whether AI would result in more or fewer legal profession jobs by 2030, said fewer. The question is not whether the legal workforce will contract. It is whether the contraction will be managed as a gradual recalibration or experienced as a sudden dislocation.

The Radiologist Who Would Not Vanish

Radiology was supposed to be the canary in the coal mine. It has instead become the profession that refused to cooperate with the narrative. By the end of 2025, the FDA had authorised 1,104 AI-enabled radiology devices, representing 76 percent of all AI medical device approvals. GE HealthCare alone holds 120 authorisations. Siemens Healthineers has 89. The technology is not theoretical. It is deployed, approved, and operational in hospitals worldwide.

And yet radiologists are earning more than ever. Median compensation reached approximately $580,000 in 2025, with a 7.5 percent year-over-year increase pushing average salaries from $532,000 to $572,000. The American College of Radiology reports that imaging volumes are rising 3 to 4 percent annually, driven by an ageing population and expanding diagnostic protocols. Radiologist attrition rates have jumped 50 percent since 2020. It takes an average of 130 days to fill a radiology position, according to the 2024 AAPPR Benchmarking Report. The workforce shortage is projected to worsen.

Why has AI failed to displace the one profession everyone predicted it would devour? Several factors converge. First, the models perform brilliantly on benchmarks but stumble in real hospital conditions, where image quality varies, patient histories are messy, and edge cases abound. Most approved tools can only diagnose abnormalities well represented in their training data. Second, diagnostic image reading accounts for only a fraction of what radiologists actually do. The rest involves consulting with other clinicians, communicating with patients, integrating clinical context, and making judgment calls that no algorithm can yet replicate. Third, regulators and medical insurers remain reluctant to approve or cover fully autonomous radiology AI. The legal liability questions alone are enough to keep a human in the loop for decades.

Hinton, in his retraction, said he "didn't make it clear" that he was speaking purely about image analysis, and that he was "wrong on the timing but not the direction." Perhaps. But the radiology story offers a more useful lesson than a prediction about timelines. It demonstrates that the relationship between AI capability and workforce displacement is not linear. A technology can be extraordinarily good at a subset of a job's tasks and still fail to replace the person holding that job, if the remaining tasks require skills the technology cannot approximate.

The Uneven Geography of Displacement

Goldman Sachs estimates that AI could replace the equivalent of 300 million full-time jobs globally, though the firm projects only a mild unemployment increase of about 0.5 percent during the transition period. McKinsey, somewhat more soberly, estimates that current technology could automate approximately 57 percent of U.S. work hours, with 30 percent of hours currently worked likely to be automated by 2030, concentrated in data processing, customer service, and documentation-heavy functions. Entry-level hiring in roles classified as "AI-exposed" has already dropped 13 percent since large language models began proliferating.

The demographic skew is striking. Approximately 79 percent of employed women in the United States hold positions categorised as high-risk for automation, compared with 58 percent of men. This is not because the technology discriminates; it is because the occupational landscape does. Women are disproportionately represented in administrative, clerical, and para-professional roles that involve precisely the kind of structured, document-heavy work AI handles well.

The Klarna episode offers an instructive cautionary tale. The fintech company's AI customer service agent was handling the work of 853 human agents by 2025, saving the company 60 million dollars. CEO Sebastian Siemiatkowski celebrated the company's shrinkage from 5,000 to 3,000 employees. Then, quietly, Klarna reversed course and began hiring human agents again, acknowledging that AI had failed to meet the company's standards for customer experience. The lesson was expensive and public: automation that looks transformative on a spreadsheet can prove brittle in practice when the work involves nuance, empathy, or the kind of situational judgment that emerges from human experience.

What Remains When the Routine Is Gone

The pattern across all three professions is remarkably consistent. AI is not eliminating the profession. It is eliminating the apprenticeship. The entry-level work that once served as both revenue generator and training ground, reconciling ledgers, reviewing contracts, triaging scans, is precisely the work most amenable to automation. This creates an uncomfortable structural problem: how do you produce experienced professionals if there is no longer a path of gradually increasing responsibility through which they can develop judgment?

Deloitte's research suggests AI agents could cut costs by 25 percent and boost productivity by 40 percent in finance teams. That is an extraordinary figure, but it raises an immediate question about where the next generation of finance leaders will come from if the work that once trained them no longer exists. The same question applies to law. If junior associates no longer spend years grinding through document review, developing an intuition for which clause matters and which is boilerplate, what replaces that education? The firms have not answered this convincingly. Most are betting on "upskilling," a word that has become so reflexively deployed in corporate communications that it has lost nearly all meaning.

For radiologists, the dynamic is inverted. AI has not eliminated the entry-level path; it has enriched it. Radiology residents now learn to work alongside AI tools, using them to prioritise scans, enhance image quality, and cross-reference findings. The technology augments their training rather than replacing it. This may explain why radiology, despite being the profession most frequently cited as AI's first victim, has proven most resilient. The integration happened collaboratively, not as a substitution.

The Firms That Ate Their Young

There is a particular irony in the Big Four's position. For years, these firms warned of an accounting talent shortage, lobbied for expanded CPA pipelines, and recruited aggressively on university campuses. Now they are investing billions in technology that renders much of that pipeline unnecessary. PwC trained 315,000 employees in AI while simultaneously cutting thousands of positions. The message to young professionals is dissonant: we need you to be AI-literate, but we may not need you at all.

Law firms face a parallel contradiction. Profits are at record highs. Partners resist structural change because the current model is working beautifully, for them. But the Thomson Reuters analysis is blunt: "firms deploying technology that can accomplish in minutes what once took hours, then trying to bill for it by the hour" creates what it calls an "almost absurd tension." The shift toward alternative fee arrangements, projected to rise from 20 percent of law firm revenue in 2023 to substantially more by 2026, will eventually force a reckoning between efficiency and employment.

The professions are not collapsing. They are stratifying. At the top, experienced practitioners with strong client relationships, deep judgment, and the ability to wield AI as a force multiplier will thrive. Their compensation will likely increase. At the bottom, the entry points are narrowing, the training paths are eroding, and the competition for remaining positions is intensifying. In the middle, a precarious zone is forming where professionals who cannot demonstrate value beyond what an algorithm provides will find themselves quietly surplus to requirements.

The Silence Before the Restructuring

What is perhaps most notable about this transformation is how quietly it is proceeding. There are no factory closures to photograph, no picket lines to film. The displacement happens in hiring freezes rather than layoffs, in delayed start dates rather than termination notices, in roles that simply stop being posted. A 44 percent decline in graduate accounting listings does not make the evening news. A law firm that completes a brief in ten hours instead of twenty-five does not issue a press release about the fifteen hours of associate work that vanished. A radiology department that uses AI to triage scans faster does not announce that it decided not to hire the additional radiologist it had budgeted for.

McKinsey estimates that 10 to 20 percent of entry-level white-collar jobs could be eliminated in the next one to five years, potentially affecting nearly 50 million U.S. positions. That is a staggering figure, but it will likely materialise not as a single dramatic event but as a slow accumulation of incremental decisions made by individual firms, each one rational in isolation, collectively reshaping the professional landscape in ways that will take years to fully comprehend.

Hinton was wrong about radiologists, but he was asking the right question. The professions that built the modern middle class, the careers that parents pointed their children toward as safe harbours, are being rewritten by technology that is, in many cases, genuinely remarkable. The question is no longer whether AI will change these professions. It is whether the people currently training for them will arrive to find the door still open, or merely ajar.