Which Jobs AI Is Actually Automating Right Now

In the first six months of 2025, 77,999 tech jobs were directly attributed to AI. Not theoretically exposed, not at risk in a decade, but cut and cited publicly as AI-driven (Challenger, Gray & Christmas). Salesforce eliminated 4,000 customer support roles. IBM restructured its entire HR back office. Duolingo replaced its contract translators with generative models. UPS announced 30,000 cuts on top of the 48,000 it eliminated the year prior, pointing to automated mega-hubs where robots outnumber employees 15 to 1.

These are not projections. They are earnings calls, SEC filings, and CEO interviews. The question is no longer whether AI will automate jobs. It is which ones, how fast, which of the cuts are real versus speculative, and what the labor market actually looks like once the dust settles. This is the full picture: the tasks disappearing first, the workers absorbing the hit, the new roles being created, and the macro numbers underneath the headlines.

The jobs already disappearing

The clearest signal comes from Anthropic's Economic Index, which introduced a measure called "observed exposure": which tasks large language models are actually performing across two million real conversations, not which ones they theoretically could. Ranked by observed exposure, the ten most-affected occupations look like this:

Three things stand out. The top three are not surprises, but the magnitude is new: roughly three-quarters of programming tasks are already being performed with AI assistance. The middle of the list is where white-collar professionals should pay attention, because market research, sales, and financial analysis were supposed to be safe. And the gap between theory and practice is the real story. The theoretical exposure for computer and math occupations is about 94%. The observed rate is 33%. That 61-percentage-point gap is the distance between what AI can do and what organizations have actually deployed, and it is closing.

The iceberg beneath the layoffs

In November 2025, researchers at MIT and Oak Ridge National Laboratory published the Iceberg Index, which put a hard number on current capability: 11.7% of the U.S. workforce is already economically replaceable by current AI systems. Not future systems. The technology that exists right now. That figure translates to roughly $1.2 trillion in annual wages across 151 million workers.

The name is the argument. The visible part of the iceberg is familiar: tech layoffs, shrinking engineering teams, coding assistants replacing junior developers. That slice accounts for about 2.2% of total U.S. wage exposure, roughly $211 billion, concentrated in coastal metro areas. But beneath the surface sits a far larger mass. The remaining 9.5% of wage exposure spans human resources, logistics, finance, office administration, and professional services, geographically distributed across all 50 states. It includes routine document processing, financial analysis, scheduling, and administrative coordination. Nobody writes a press release about eliminating a back-office process, so it never makes headlines. The Iceberg Index is deliberately conservative: it measures only where AI can perform a task at a cost competitive with or cheaper than human labor. These are current economic realities waiting for adoption to catch up.

Customer service: the most visible displacement

Salesforce CEO Marc Benioff said it plainly in September 2025: the company reduced its support workforce from 9,000 to roughly 5,000 because "I need less heads." Agentforce handled over a million consumer conversations and cut support costs by 17%. The customer service roles that survive are the ones dealing with exceptions, escalations, and situations where empathy matters, which turns out to be a larger share of the work than most automation forecasts assumed. (More on why the flagship rollouts got walked back below.)

Translation: the steepest decline

Translation may be the profession most visibly gutted by AI in the current wave. The International Monetary Fund reduced its translator headcount from 200 to 50. Research on Jordan's translation industry found AI drove a 40 to 70% decrease in human translator employment. On freelance platforms, translation postings dropped 19% between late 2022 and early 2024. Duolingo made the shift explicit, replacing contract translators with generative AI and launching 148 new courses in under a year, work that previously took a decade. The company later faced criticism that quality had degraded: "more repetitive, less nuanced, and more like an endless stream of automated exercises." What survives is diplomatic, legal, financial, and medical translation, contexts where a mistranslation carries material consequences. The bulk commodity work is largely gone.

Writing and content: death of the commodity brief

An analysis of 180 million global job postings by Bloomberry found writing roles among the steepest decliners since ChatGPT launched. On Upwork specifically, writing work dropped about a third year over year. Ramp's economics lab tracked the spending shift: freelance marketplace spending as a share of company budgets fell from 0.66% to 0.14%, while AI model spending rose from zero to 2.85%. The decline is not evenly distributed. As the Bloomberry researcher put it, "the copywriter who executes a brief is vulnerable. The creative director who decides what the brief should say is less so." AI writes serviceable first drafts but struggles with brand voice, strategic framing, and editorial judgment.

Data entry and administrative work: quiet, near-total automation

Manual data entry clerks face a 95% automation risk (Brookings Institution/GovAI Occupation Vulnerability Index), with AI systems processing over 1,000 documents per hour at error rates below 0.1%, compared to 2 to 5% for humans. In healthcare, ambient AI that listens to doctor-patient conversations and generates clinical notes in real time is now deployed across the majority of U.S. healthcare systems, and medical scribe postings are declining at 20% annually. Accounting is bifurcating along the same line: transaction categorization (90% automation risk), bank reconciliation (85%), and invoice processing (85%) are shifting to AI, so bookkeeper employment is declining at 5% while accountant employment grows at 5%. The commodity tasks automate; the advisory work requiring judgment grows.

Recruiting and HR screening: automation of the middle

87% of organizations now use AI at some point in their hiring process (SHRM 2025). AI screening reduces resume review time by up to 75%, and teams report 20 to 40% lower cost-per-hire. IBM used AI to take over the work of several hundred HR employees, with its AskHR agent now automating 94% of routine HR tasks. But IBM's total headcount actually increased, redirected into engineering and sales. This is the pattern that keeps repeating: automation does not eliminate the department, it eliminates the routine layer within it and redistributes headcount toward higher-judgment functions.

Freelancers got hit first, and that is the tell

If you want to see where full-time employment is heading, watch the freelance market. It moved first because friction was low. A company posts a task, a freelancer completes it, money changes hands. No long-term relationship, no institutional knowledge. The switching cost from a human freelancer to an AI tool is close to zero, which is exactly why the Ramp exchange rate is so brutal: for every dollar companies cut from freelance budgets, they now spend roughly three cents on AI tools, a 97% cost reduction.

A peer-reviewed study in Organization Science (INFORMS) found freelancers in AI-exposed occupations saw a 2% decline in contracts and a 5% drop in earnings within months of new AI tools entering the market. The counterintuitive part: the highest-earning freelancers were hit hardest. For every 1% increase in a freelancer's past earnings, they lost an additional 0.5% of opportunities and 1.7% of monthly income. When a $50-an-hour writer competes with a tool that produces 80% of the quality at 1% of the cost, the value proposition collapses, because the buyer was paying for the output, not the relationship.

Contrast that with what the Dallas Federal Reserve found for full-time employees, who bring institutional knowledge, team relationships, and accountability. AI substitutes for "codifiable knowledge" (what sits in a textbook) and complements "tacit knowledge" (understanding gained through experience). Since ChatGPT's debut, weekly wages rose about 7.5% nationwide, but jumped 8.5% in the top 10% of AI-exposed industries and 16.7% in computer systems design specifically. Same skills, different relationship structure, opposite outcomes.

And the freelance market did not collapse so much as bifurcate. Bloomberry found several categories growing: video editing and production up 39%, web design up 10%, graphic design up 8%, backend development up 6%. Work involving synthesis, visual judgment, and multi-step creative decisions held up. Work reducible to "produce X text about Y topic" did not. Meanwhile Upwork reported AI-related freelance work grew 60% in gross services volume in 2024, with those freelancers earning 44% higher hourly rates. The question for every worker is which side of that split they land on. Building the career resilience to land on the right side is grounded in relationship depth and domain expertise, not just skill accumulation.

Middle management is being flattened

In March 2026, Meta announced an applied AI engineering team with a 50-to-1 employee-to-manager ratio, double the 25-to-1 figure organizational theorists long considered the outer boundary of effective management. Amazon confirmed it had surpassed its target of increasing the individual-contributor-to-manager ratio by 15%, after cutting roughly 30,000 corporate roles. Google targeted a 10% reduction in manager roles across ad sales. Gartner predicted the trend in October 2024: by 2026, 20% of organizations will use AI to flatten their structures, eliminating more than half of current middle management positions. Roughly 11 million Americans currently hold the title "manager."

The logic starts with how these roles spend their time. A McKinsey survey found the average middle manager spends only 28% of their time actually managing people. Nearly 18% goes to pure administrative work (expense approvals, status reports, scheduling), and another 35% to meetings that mostly transfer information up and down the hierarchy. When a role's core value is aggregating information and relaying it between layers, and software does that in seconds (AI project tools cut time spent on status updates by 90%), the economic argument is hard to refute. The World Economic Forum projects that by 2030 only 33% of workplace tasks will be performed solely by humans, down from 47% today.

What resists automation requires reading humans, not data: mentoring, conflict resolution, culture-building, and navigating organizational politics. Harvard Business Review's 2025 analysis identified three skills that define the surviving manager: agentic AI literacy, domain-specific expertise, and integrative problem solving. But the less-discussed cost is the mentorship gap. Middle managers have traditionally been the primary coaching layer in organizations. Meta's 50-to-1 ratio assumes one manager can meaningfully develop 50 engineers, and organizational research suggests otherwise. Thinning that layer disrupts the pipeline that produces the next generation of leaders, and none of the companies executing these strategies has offered a convincing answer to it.

The entry-level door is closing

If middle management is being compressed from above, the entry level is being sealed from below. In 2021 a new graduate could expect to submit 10 to 15 applications before landing a role. By 2025 that number was 43, and some entry-level roles now attract north of 400 applicants each.

The postings themselves are vanishing. Revelio Labs found entry-level U.S. postings dropped 35% between January 2023 and June 2025, roughly 100,000 fewer junior openings a month, with highly AI-exposed entry roles (data engineers, software developers, customer service reps, financial analysts) down over 40%. Big Tech, per SignalFire, cut new-graduate hiring 25% while increasing hiring of workers with 2 to 5 years of experience by 27%. Stanford's Canaries in the Coal Mine study, built on ADP payroll records, found workers aged 22 to 25 in the most AI-exposed occupations saw a 13% relative employment decline since late 2022, even as employment for workers over 30 in the same occupations grew 6 to 12%. Software developers aged 22 to 25 saw employment fall nearly 20% from its late-2022 peak.

The critical nuance, from the Dallas Fed: this is not layoffs. Separation rates have not increased. What changed is the job-finding rate, which for workers aged 20 to 24 in AI-exposed occupations declined by more than 3 percentage points since its November 2023 peak. Layoffs make headlines. Hiring freezes at the junior level are invisible, showing up as applications that never get responses. The Cleveland Fed found the time it takes a college graduate to find work has roughly doubled, to about four and a half months, matching a high school graduate and erasing an advantage that had persisted for decades.

Then there is the experience paradox. More than 60% of positions labeled "entry-level" in software and IT now require three or more years of prior experience. Entry-level roles historically did two things at once: they produced output and they trained the next generation. AI broke that bargain by automating exactly the routine tasks that justified junior salaries and trained junior workers. A Harvard Business School study of 62 million workers across 285,000 firms found that when companies adopt generative AI, junior headcount drops 7.7% within six quarters relative to non-adopters, while senior headcount holds steady. The decline is driven entirely by slower hiring, an average of 3.7 fewer junior workers per quarter.

The consequence arrives on a lag. Today's senior engineers were yesterday's juniors. Cut the pipeline and the supply of experienced workers contracts a few years later. Andrew Ng has argued that companies which stop investing in junior talent are "borrowing from the future," extracting AI productivity gains now while creating a compounding deficit in institutional knowledge. Organizations that treat junior hiring as a cost to eliminate rather than an investment to protect may find themselves fighting over a shrinking pool of senior talent by 2030.

Most of these layoffs are a bet, not a verdict

Here is the reframe that most coverage misses. A large share of AI-attributed layoffs are not responses to what AI has done. They are wagers on what it might do.

In March 2025, New York became the first state to require employers to disclose whether AI contributed to mass layoffs. Since then, 162 companies filed notices affecting 28,300 workers. The number that checked the AI box: zero. That includes Amazon, which publicly warned AI would reduce headcount across 30,000 roles, and Goldman Sachs, which internally linked reductions to AI productivity gains. Companies attribute cuts to AI in press interviews while declining to make the same claim under legal obligation.

The research explains why. A Harvard Business Review survey of 1,006 executives found AI-attributed layoffs are "almost completely in anticipation of AI's impact," with only 2% of organizations reporting large headcount reductions tied to actual AI implementation. A National Bureau of Economic Research paper surveying nearly 6,000 CEOs and CFOs found more than 90% reported AI had no impact on their employment over the past three years. And the returns are not there to justify the cuts: an MIT report found 95% of companies investing in generative AI are seeing zero return on $35 to $40 billion collectively invested. Even OpenAI's Sam Altman told CNBC that some firms are engaged in "AI washing", "blaming AI for layoffs that they would otherwise do."

The case studies are sobering. At King, the Candy Crush developer, level designers and writers spent months building internal AI tools, and in July 2025 roughly 200 of them were laid off and replaced by those same tools. A 2026 Gartner survey found 64% of organizations used existing employees to create AI training data, but only 22% were transparent about how that data would affect staffing. Block cut roughly 40% of its workforce (about 4,000 people) citing AI, though its headcount had ballooned from 3,800 in 2019 to over 10,000, which analysts read as an overdue cleanup as much as an AI story. And Amazon's retail site suffered four high-severity incidents in a single week, including a six-hour outage, after an engineer followed inaccurate guidance from an AI agent that had pulled from an outdated internal wiki. The company is now introducing what it calls "controlled friction" for AI-assisted changes to critical systems. The humans, it turns out, were performing a function.

Put in context: Challenger attributed roughly 55,000 U.S. layoffs to AI in 2025, a twelvefold jump from 2023, but that is less than 5% of the 1.2 million total cuts announced that year. The trend line (4,600 in 2023, 12,700 in 2024, 55,000 in 2025) matters more than the absolute number. And the majority view among executives runs the other way: the HBR survey found 57% expect AI to increase headcount over the next year versus only 15% who expect a decrease. Walmart made the point concrete in February 2026, committing to free AI training for all 1.6 million U.S. employees with no planned headcount reductions. The layoffs come from a loud minority, not the consensus.

And the bet is already being reversed

The regret is arriving as data. Forrester found 55% of employers now regret laying off workers for AI-related reasons. A February 2026 Careerminds survey of 600 HR professionals found roughly two in three companies that replaced workers with AI are already restaffing: 35.6% have rehired more than half of the eliminated roles, and 52.1% rehired within six months. The financial picture makes the regret concrete. Nearly a third of organizations found rehiring cost more than they ever saved, and another 42% merely broke even. Only about a quarter came out ahead. Gartner predicts that by 2027, half of all companies that cut customer service staff due to AI will rehire for similar functions.

The root cause is the gap between a controlled demo and production at scale. S&P Global found 42% of companies abandoned the majority of their AI initiatives before production in 2025, up from 17% the prior year. Customer service shows the widest gap of all: a Qualtrics XM study of 20,000 consumers across 14 countries found AI-powered customer service fails at nearly four times the rate of other AI tasks. Only 8% of consumers prefer AI over humans for service, 41% believe service has worsened because of AI, and 81% believe companies deploy it to cut costs rather than improve experience.

Klarna is the canonical arc. In 2024 its CEO announced an AI chatbot was doing the work of 700 agents. By mid-2025 the company was rehiring humans. "We focused too much on efficiency and cost," Sebastian Siemiatkowski admitted. "The result was lower quality, and that's not sustainable." IBM followed a similar path: after announcing AI would replace much of its HR department, it found roughly 6% of employee queries required genuine human judgment, and it has since tripled its entry-level hiring for 2026, with new hires supervising AI systems and covering failure points. Yale's Budget Lab found no significant change in unemployment rates for AI-exposed occupations between ChatGPT's release and November 2025. This is Amara's Law territory: overestimating a technology's short-run impact and underestimating the long run. For job seekers, the signal is specific. Demand for human workers has not collapsed, even in the most-exposed sectors. Roles are being redefined around AI collaboration, often under new titles, more than they are being eliminated outright.

The other side: the jobs that didn't exist two years ago

AI is not only subtracting. In January 2024, "AI Ethics Officer" barely registered as a job title. By early 2026, LinkedIn data showed 1.3 million net-new AI roles added to the global economy, positions that did not exist at meaningful scale two years prior. Five of them have gone from nonexistent to six-figure salaries in under 24 months:

• AI prompt engineer ($100,000 to $270,000): designing and testing the instructions given to language models. The standalone title is already being absorbed into broader roles (LinkedIn profiles carrying it dropped 40% from mid-2024 to early 2025, while "AI workflow designer" postings rose 25%), but the underlying skill is in high demand.
• AI ethics officer / head of AI governance ($120,000 to $243,000): bias auditing, fairness testing, and regulatory compliance. Demand is surging because the EU AI Act's high-risk requirements become enforceable in August 2026, and Forrester predicts 60% of Fortune 100 companies will appoint a head of AI governance by the end of 2026. The AI governance market is projected to grow from $840 million in 2025 to $26.9 billion by 2035.
• AI trainer / RLHF specialist ($15 to $30 an hour at entry, up to $120,000 to $180,000 for senior specialists): the most accessible entry point into AI work, actively recruiting people from teaching, research, and law rather than pure engineering.
• AI auditor ($72,000 to $170,000): evaluating AI systems for accuracy, bias, and compliance. Only 18% of organizations have fully implemented AI governance frameworks despite 88% using AI operationally, and that gap is where auditors live.
• Human-AI interaction designer ($70,000 to $225,000): designing how people actually use AI products, including how to communicate uncertainty and handle failure modes.

The wage signal underneath all of this is unambiguous. PwC's 2025 Global AI Jobs Barometer, analyzing close to a billion job advertisements, found workers with AI skills earn a 56% wage premium over comparable peers, up from 25% the prior year, and jobs requiring AI skills grew 7.5% year over year even as total postings fell 11.3%. LinkedIn reports more than 10% of professionals hired today hold job titles that did not exist in 2000. The most durable of these roles sit at the intersection of AI capability and domain knowledge: an ethics officer who understands healthcare regulation, an auditor who knows financial compliance. Because large language models have not been around for a decade, no one has a decade of experience, which is a genuine and time-limited window for career switchers. It closes as the field matures.

What the aggregate numbers actually hide

The World Economic Forum's 2025 Future of Jobs Report is the number everyone quotes: 170 million new roles created and 92 million displaced by 2030, a net gain of 78 million jobs, with 22% of all jobs facing structural disruption. That framing is not wrong. But it obscures a harder question: progress for whom? The workers most likely to lose their jobs and the workers most likely to land the new ones are not the same people.

Look at the composition. The new jobs skew in two directions at once: high-skill technical roles (big data specialists, fintech engineers, AI and machine learning specialists, software developers, security specialists) and physical-presence work (the green transition alone is projected to add 34 million farmworker roles, plus delivery drivers, construction, and care-economy roles driven by aging populations). What disappears sits squarely in the middle. In absolute terms, cashiers and ticket clerks face a net loss of 13.7 million, administrative assistants and executive secretaries 6.1 million, stock-keeping clerks 2.64 million, and accounting and payroll clerks 1.65 million. The fastest percentage declines are postal clerks (-34%), bank tellers (-31%), and data entry clerks (-26%). The pattern is stark: the new jobs require either technical specialization or physical presence, while the vanishing ones are structured, routine, office-based work.

The clerical and administrative category faces the single largest absolute decline of any occupational group, and it is one of the most gender-concentrated segments of the workforce. A Brookings study measured "adaptive capacity" (savings, transferable skills, local job-market density, age) and found roughly 6.1 million U.S. workers face both high AI exposure and low adaptive capacity, and about 86% of them are women. The largest groups are office clerks, secretaries, receptionists, and medical secretaries. The International Labour Organization corroborates it globally: 4.7% of women's jobs fall into the highest AI-risk category versus 2.4% for men, and in high-income countries the gap widens to 9.6% of female employment versus 3.5% for men. As economist Noreena Hertz noted, women make up the majority in more than half of the 40 occupations most at risk, and a workplace-adoption gap (36% of men use generative AI daily at work versus 25% of women) leaves them simultaneously more exposed and less positioned to capture the upside.

Age cuts a second fault line. Historical BLS data shows workers aged 55 to 64 who lost jobs during the Great Recession were 16 percentage points less likely to find re-employment than workers aged 35 to 44. The Dallas Fed's finding that AI hits entry-level workers while complementing experienced ones means AI is hollowing out both ends toward the middle: juniors lose the entry ramp, and senior workers without technical fluency lose the safety net. And the whole transition runs into a skills bottleneck. The WEF estimates 39% of core job skills will change by 2030. If the global workforce were 100 people, 59 would need retraining, and 11 are unlikely to receive it, which translates to over 120 million workers at medium-term risk. Understanding the skills gap and which capabilities to prioritize first is the practical version of this problem.

None of this points to apocalypse. The ILO's study explicitly concludes that job transformation, not wholesale replacement, is the most probable outcome, and Census data offers the long view: about 60% of Americans today work in occupations that did not exist in 1940. Technology transitions do create jobs, eventually. The open question is always what happens in the gap.

What this means for job seekers

Pull all of it together and the takeaways are consistent, and more actionable than the headlines suggest.

The threat is task-level, not job-level. If 40% of your role is automatable, that is not a 40% chance you lose your job. It means 40% of your time gets reallocated, either to higher-value work or, if your organization handles it poorly, to a smaller headcount doing the remaining 60%. Audit your task composition, not your job title. The 61-point gap between what AI can do and what organizations have deployed is your window to reposition, and it is closing, not permanent.

Speed and AI fluency compound. The workers who learn to handle the automatable portion of their work with AI tools become more valuable, not less. Across the freelance data, the entry-level data, and the new-role data, the same divergence appears: AI-augmented, judgment-heavy, relationship-driven work grows, while commoditized output-focused work compresses. The augmentation ratio is the hopeful signal here, with Anthropic finding that 52% of AI usage is people doing their existing jobs better versus 45% doing tasks independently.

The hiring freeze is the bigger risk than the layoff. Companies are not dramatically firing so much as quietly not backfilling, which shows up as a measured drop in job-finding rates for young workers in exposed fields. For job seekers that means higher competition for fewer open roles, longer timelines, and a premium on moving fast when a position opens.

Precision beats volume in a saturated market. Employers now receive 250-plus applications per posting, and only about 2.4% of applicants reach the interview stage. Submitting more generic applications is not a strategy. The advantage has shifted to quality: tailored materials, targeted companies, and fast response to new postings. Two tactics that help in a tighter funnel are targeting small companies, which attract fewer applicants and move faster, and writing resume bullets that show impact rather than list duties. It is also worth knowing that a meaningful share of postings are ghost jobs with no real intent to hire, another layer of friction that has nothing to do with AI.

The labor market is not collapsing. It is reshuffling, unevenly, and faster than most application processes were built to handle. The workers who fare best are the ones who understand their exposure, invest in the skills AI cannot replicate, and approach the search with precision rather than sheer volume.

---

Nox finds, evaluates, and applies to jobs on your behalf, matching your experience against open roles and submitting tailored applications through employer platforms while you focus on the parts of your search that require human judgment.

Try Nox free. No credit card required.

---

Sources: Anthropic Economic Index, MIT / Oak Ridge National Laboratory Iceberg Index (2025), World Economic Forum Future of Jobs Report 2025, Brookings Institution: Measuring US Workers' Capacity to Adapt to AI-Driven Job Displacement, St. Louis Federal Reserve (2025), Dallas Federal Reserve (Feb 2026), Dallas Federal Reserve (Jan 2026), Stanford Digital Economy Lab: Canaries in the Coal Mine (2025), Harvard Business School (2025), Cleveland Federal Reserve, Revelio Labs, Goldman Sachs, McKinsey Global Institute, Bloomberry 180M Jobs Analysis, Ramp Economics Lab, INFORMS / Organization Science, Upwork State of AI, PwC 2025 Global AI Jobs Barometer, LinkedIn / WEF: 1.3M New Jobs, Harvard Business Review: Companies Are Laying Off Workers Because of AI's Potential, NBER (w34836), MIT GenAI Divide report, Sam Altman on AI washing, Careerminds: Cost of AI Layoffs, Gartner rehire prediction, S&P Global Market Intelligence, Qualtrics XM Institute, Yale Budget Lab, International Labour Organization (2025), Noreena Hertz, Project Syndicate (2025), Randstad, Challenger, Gray & Christmas via CNBC, Greenhouse 2025 Workforce Hiring Report, SHRM 2025 AI in Hiring Survey.