The fear is real, and it is not foolish. In a Pew Research Center survey of 5,273 employed US adults in October 2024, published February 2025, 52% said they were worried about future AI use in the workplace, and 32% expected fewer job opportunities for themselves in the long term. Only 6% expected more. That worry is the starting point of this business: people who work hard and are good at what they do, afraid the technology will make them irrelevant before they figure it out.
But worry is not the same as a pink slip. Exposure is not displacement. Exposure means AI could touch some tasks in a job. Goldman Sachs Research estimated in 2023 that about 25% of current work tasks could be substituted by generative AI — and that most exposed jobs are only partially exposed, which means complemented more than substituted. The International Labour Organization, in Working Paper 140 (May 2025), found 1 in 4 workers worldwide in an occupation with some generative-AI exposure, but only 3.3% of global employment in the highest-exposure category. The World Economic Forum's Future of Jobs Report 2025 expects 92 million jobs displaced and 170 million created by 2030 across all trends, not AI alone — a net gain of 78 million, with about 22% churn in between. The churn is the painful part. This post is about what happens on the other side of it, industry by industry.
How to read the labels. Every figure below carries its evidence label: Measured means a field experiment, payroll data, or logged records. Survey means workers or employers were asked. Estimate means a projection or model, not a result. Company-reported and Vendor mean the number comes from the company selling or using the system. Where no good industry survey exists, we say so instead of borrowing a general figure.
1. Customer service / call centers
The jobs feeling the pressure: phone and chat agents, support reps, help-desk staff.
The fear: In a Verint survey of 1,000 agents (reported April 2026), only 8% feared being replaced by AI within three years — though nearly all expected their role to change. (Survey; note: Verint sells customer-service automation.) In a Cavell Group survey of 500+ agents in the US and Canada (reported November 2025), nearly half worried AI would replace part of their role, and 55% said task volume and complexity rose after AI took the easier issues. (Survey.)
AI takes: routine FAQs, password resets, order-status checks, drafted replies, call summaries. People keep: complex and escalated cases, de-escalation, judgment calls, exceptions, relationship repair, and checking AI drafts before they go out.
The outcome: In a field experiment with 5,179 agents at one Fortune 500 software firm (Brynjolfsson, Li & Raymond, NBER Working Paper 31161, 2023), a generative-AI assistant raised issues resolved per hour by 14% on average — and about 34% for novice and lower-skilled agents. (Measured — field experiment, single company.) A Verizon Business survey in summer 2025 found 88% satisfaction with human-led digital service versus 60% for AI-only. (Survey.) The cautionary case: Klarna's AI assistant handled 2.3 million conversations in its first month in 2024, work the company equated to about 700 full-time agents, with resolution in about 2 minutes versus 11 — then in May 2025 the company rehired humans after quality fell on complex cases. (Company-reported, mixed outcome.)
Bottom line: the split that works is copilot, not autopilot — AI beside the agent, human in front of the customer.
2. Retail
The jobs feeling the pressure: cashiers, sales associates, stock and inventory staff.
The fear: In a Chadix survey reported in early 2025, 40% of retail-worker respondents feared AI automation would lead to termination of their jobs, and 60% anticipated major industry disruption within five years. (Survey — weak: vendor survey, sample size and method not stated in the reporting found.) The World Economic Forum's 2025 report names cashiers and ticket clerks among the fastest-declining roles. (Estimate.)
AI takes: demand forecasting, replenishment, price and schedule suggestions, inventory counts, "is it in stock?" questions, drafted product descriptions. People keep: greeting and reading the shopper, complaints and returns with discretion, merchandising judgment, fitting and advice, spotting theft or safety issues, and fixing AI mistakes like wrong stock or wrong prices.
The outcome: Walmart reported its "Sparky" shopping assistant grew users 70% year-on-year, with users spending 40% more per order; Lowe's reported its "MyLow" assistant answered 25 million+ questions, with users converting at 3 times the rate of non-users. (Company-reported; association, not proof AI caused the gap.) In a Walmart employee survey of 1.7 million+ responses (June 2026), 84% of US frontline workers said they were confident using digital tools. (Survey — company survey, self-reported.) No randomised or field-experiment evidence comparable to customer support was found for frontline retail associates — the measured evidence here is thin. (Gap.)
Bottom line: the store stays human where judgment and trust live; AI earns its place in the back room first.
3. Hospitality and tourism
The jobs feeling the pressure: front-desk staff, concierges, housekeepers, tour and activity operators.
The fear: In a job-board survey of 3,000 workers across sectors, 59% of hospitality workers feared AI could take their jobs. (Survey — weak: job-board survey, date and method not clearly stated.) Guests worry too: in a Mews survey of 3,250 travellers (July 2026), 67% said they would worry about hotel staff losing jobs to AI. (Survey.) At the same time, hospitality is among the least AI-exposed industries in PwC's 2025 barometer, because so much of the work is physical and in person. (Estimate.)
AI takes: 24/7 answers to routine questions, translation, check-in and check-out, housekeeping scheduling, pricing suggestions, drafted guest messages. People keep: the welcome, local recommendations with real knowledge, complaint recovery, bar and restaurant service, safety judgment, and the personal touch guests say they want.
The outcome: The same Mews survey found 89% of guests want a hybrid stay or AI only behind the scenes, 75% say good hospitality depends on human staff, and 59% say AI should take repetitive tasks so staff can build rapport and trust. (Survey.) Wyndham reported in its Q3 2026 earnings that generative-AI guest-engagement tools at 5,100+ hotels lifted franchisee upsell revenue 16% year-to-date and guest satisfaction by 450 basis points. (Company-reported, not independently measured.) No peer-reviewed study of hotel-staff retention or productivity from AI coexistence was found. (Gap.)
Bottom line: guests want the robot in the back office and the human at the front desk — and they say so in large numbers.
4. Food service
The jobs feeling the pressure: order takers, drive-thru staff, cooks, servers, shift leads.
The fear: No direct "fear AI will take my food-service job" percentage was found. (Gap — no industry fear survey.) The closest worker survey is openness, not fear: in Toast's 2026 survey of 702 US restaurant employees (February–March 2026), about 20% were extremely open to using AI at work and about 10% rejected it outright; 54% were extremely or very open to AI for end-of-shift summaries and shift scheduling. (Survey; Toast sells AI tools to restaurants.) An ILO food-system analysis found most food occupations "not exposed" to generative AI, with waiters and food-delivery drivers at minimal or low exposure. (Estimate.)
AI takes: standard drive-thru and phone orders, upsell suggestions, prep forecasting, shift scheduling, shift summaries. People keep: cooking and plating to standard, complex and custom orders and allergies, greeting and recovering unhappy guests, checking AI orders, and keeping the line moving when AI stalls.
The outcome: In Intouch Insight's 2025 drive-thru mystery-shopping study (120 AI visits across Burger King, Taco Bell, and Wendy's), AI averaged 3 minutes 53 seconds order-to-food versus 4 minutes 15 seconds for humans — but accuracy was 83% for AI versus 87% for humans, about 22% of AI orders needed a person to step in, and 62% of AI errors came from customisations. Hybrid service (AI plus human escalation) was reported at 97% accuracy. (Measured — mystery shopping.) McDonald's "Archy" voice AI (September 2026) was reported at accuracy above 90%, with an expectation of freeing at least 50 labour hours per week per restaurant. (Company expectation, not a measured result.)
Bottom line: AI is faster but less accurate alone; the hybrid is the only version that beat humans on accuracy.
5. Healthcare administration and support
The jobs feeling the pressure: medical billers and coders, schedulers, prior-authorisation staff, medical assistants, documentation support.
The fear: In a job-board survey, 44% of healthcare workers feared AI could take their jobs. (Survey — weak: job-board survey.) Exposure is high for clerical and administrative tasks: the ILO's 2025 exposure index puts clerical occupations highest overall. (Estimate.) Clinical judgment, hands-on care, and patient communication remain human-led in every study below.
AI takes: drafting visit notes from the conversation, chasing prior authorisations, checking eligibility, coding and billing, scheduling and reminders, triaging messages. People keep: reviewing and signing every note, deciding what is clinically true, handling denials and appeals and sensitive conversations, checking AI output for invented details, and keeping patients' trust.
The outcome: A JAMA study of 8,581 clinicians across 5 health systems, using electronic-health-record logs, found ambient AI scribes cut documentation time by only about 13 minutes per day on average — far below vendor claims of 1–2 hours. (Measured — EHR logs.) An Annals of Emergency Medicine study at Mass General Brigham emergency departments found ambient AI saved about 1.6 minutes per note versus 3.2 minutes for a human scribe, with no increase in patients seen, no shorter waits, and no more revenue per patient. (Measured.) An NEJM Catalyst analysis of Kaiser Permanente estimated 15,700+ documentation hours saved across about 2.5 million encounters. (Estimate — analysis, not a randomised trial.) By 2024, 71% of US non-federal acute-care hospitals reported predictive-AI use. (Survey — adoption, not outcome.)
Bottom line: documentation savings are real but much smaller than marketed when measured from logs — and human review remains the safety layer.
6. Finance, accounting, and bookkeeping
The jobs feeling the pressure: bookkeepers, accountants, payroll clerks, financial analysts.
The fear: In a Stanford/MIT survey of 277 accountants (Choi & Xie, 2025), 62% were concerned about AI-generated errors, 43% about data security, and 37% about job stability. (Survey.) In OECD workplace surveys (2023), 60% of finance and manufacturing workers feared losing their job to AI within a decade. (Survey.) Goldman Sachs (2023) put office and administrative support at 46% of tasks susceptible to automation — the highest occupation group. (Estimate.)
AI takes: classifying transactions, matching invoices, reconciling accounts, flagging anomalies, drafting reports, pulling standards and research. People keep: judging exceptions, talking to the client, advising on tax and business decisions, signing off, checking AI work — especially where AI is uncertain — and owning audit and compliance risk.
The outcome: In observational task and survey data from 79 small and mid-sized firms plus the 277 accountants (Choi & Xie, Stanford GSB, June 2025), AI users finalised monthly statements 7.5 days faster, spent 8.5% less time on routine back-office processing, supported more clients per week, and saw 12% more reporting granularity. (Measured/observational — not a randomised trial; firms that adopted AI may differ from those that did not.) The crucial nuance from the same study: senior accountants gained more, treating AI as a collaborator and intervening when confidence dropped; juniors accepted outputs at face value and gained less. Human oversight skill, not just the tool, drove the gain. BCG/OpenAI estimated agents could cut bank costs 30–40% by 2030. (Estimate, not a result.)
Bottom line: the strongest finance evidence is augmentation with senior judgment — not replacement.
7. Legal support
The jobs feeling the pressure: paralegals, legal assistants, document reviewers, junior researchers.
The fear: In an Axiom survey reported in 2026, 76% of in-house legal-team members worried about AI-driven redundancy. (Survey — weak: sample and method not stated in the reporting found.) In a LexisNexis survey of 543 legal professionals (September 2026), 94% used AI for legal work, up from 11% in 2023, while 83% were concerned about inaccurate or fabricated information. (Survey.) Goldman Sachs (2023) put legal work at 44% of tasks susceptible to automation — the second-highest group. (Estimate.) Paralegal and document-review work is the most exposed slice; advocacy, negotiation, client counselling, and courtroom judgment are not.
AI takes: first-pass legal research, summarising documents and transcripts, extracting clauses, triaging e-discovery, first drafts. People keep: verifying every citation — hallucinated cases have led to sanctions in reported cases — exercising legal judgment, advising the client, negotiating, appearing in court, and remaining accountable for the filing.
The outcome: Thomson Reuters projected professionals could save about 4 hours per week within a year and about 12 hours per week within five years. (Estimate — a projection from survey expectations, widely repeated as if measured.) No field experiment for paralegals comparable to the support, writing, and consultant experiments was found. (Gap.) In the LexisNexis survey, 81% of lawyers were more comfortable with AI grounded in trusted legal sources, and human oversight ranked above cost when buying legal technology. (Survey.)
Bottom line: accuracy risk is the binding constraint — which is exactly why the human checker role grows.
8. Marketing, writing, and content
The jobs feeling the pressure: freelance writers, copywriters, content marketers, editors.
The fear: In an ASJA survey of 1,481 working writers (2025), 45% of freelancers reported reduced demand and 40% reduced income from AI, 75% expected opportunities to decline in the next five years, and 25% had considered quitting. (Survey — self-selected writer survey; strong signal for freelancers, not a population estimate.) Exposure is high for routine text: the ILO's 2025 index shows strongly digitised media occupations rising in exposure. (Estimate.)
AI takes: first drafts, outlines, headlines, product descriptions, research summaries, repurposing one piece into many formats, checking consistency. People keep: strategy, interviewing sources, fact-checking, voice and brand judgment, deciding what is true and worth saying, editing AI drafts into publishable work, and owning accuracy and originality.
The outcome: In an experiment with 453 college-educated professionals (Noy & Zhang, Science, 2023), ChatGPT cut writing-task time 40% and raised blind-rated quality 18%, with the weakest writers gaining most. (Measured — experiment; tasks were generic and self-contained.) In an experiment with 758 consultants (Dell'Acqua et al. with BCG, 2023), GPT-4 users completed 12.2% more tasks and worked 25% faster on tasks inside AI's capabilities — but were 19 percentage points less likely to reach the right answer on a task deliberately outside AI's capabilities. (Measured — experiment.) And in freelance-platform data covering 3 million+ postings, writing "About" pages postings fell about 50% after ChatGPT, and exposed freelancers saw about 2% fewer jobs and 5.2% lower monthly earnings. (Measured — platform data, post-ChatGPT shock period.)
Bottom line: AI measurably speeds and improves generic writing in experiments, and measurably shrank demand for routine writing. Both are true — the surviving work is edited, strategic, verified work.
9. Transportation and logistics
The jobs feeling the pressure: truck drivers, dispatchers, warehouse coordinators, route planners.
The fear: In a 2021 study cited in a 2025 EU report, 29% of logistics workers worldwide feared replacement by machines. (Survey — old, pre-generative-AI; label the date.) A PwC sector analysis put transportation and storage at 52% of jobs exposed. (Estimate, via secondary reporting.) Old autonomous-truck displacement forecasts (2–4.4 million EU/US truck jobs by 2030) have not materialised on that timetable and should not be quoted as current forecasts.
AI takes: route optimisation, load planning, demand forecasting, warehouse pick sequencing, predictive maintenance, rerouting around weather and traffic. People keep: driving in complex or unmapped conditions, handling exceptions and customer problems, loading and securing freight, supervising robots, and making safety calls AI cannot.
The outcome: UPS reported its ORION route optimisation saved 100 million+ miles and about 10 million gallons of fuel per year, with $300–400 million in annual savings claimed. (Company-reported, long-running claims.) DHL reported a 12% reduction in European transport spend and 10 million miles saved via smart-truck rerouting. (Company-reported.) In a Mecalux/MIT survey of 2,000+ logistics leaders across 21 countries (November 2025), over 90% of warehouses used some AI or automation, over 75% reported rises in both employee productivity and satisfaction after AI, and over 50% grew their workforce. (Survey — vendor-commissioned, self-reported.) McKinsey estimated AI across logistics at 5–20% logistics-cost reduction and up to 30% inventory reduction. (Estimate.)
Bottom line: routing is the most proven logistics use case — but driver displacement has been slower than forecast, and driver shortage remains the binding constraint in much of the sector.
10. Manufacturing and warehousing
The jobs feeling the pressure: pickers and packers, machine operators, welders, assemblers, maintenance techs.
The fear: In the Chadix survey (early 2025), 45% of manufacturing workers feared AI automation would terminate their jobs — the highest of any sector in that survey. (Survey — weak: vendor survey.) The OECD 2023 figure — 60% of finance and manufacturing workers fearing replacement within a decade — applies here too. (Survey.)
AI and robots take: moving pallets and totes to the person, repetitive picking, heavy lifting, visual defect detection, predicting machine failure, optimising schedules. People keep: setting up, feeding, and supervising machines; fixing exceptions; quality judgment on ambiguous defects; repair and maintenance; and safety calls. The evidence below shows pace pressure is the risk when the robot sets the tempo.
The outcome: In a peer-reviewed study of weekly data from 150+ Amazon warehouses (Greenwood, Burtch & Ravindran, ILR Review, April 2025), robotic fulfilment centres had 40% fewer severe injuries than legacy centres — but 77% more less-severe, repetitive-strain injuries, with pick rates 2–3 times higher. (Measured — peer-reviewed; data from 2016–2019.) Company and union injury counts for Amazon disagree on definitions and periods; report the disagreement, do not pick one side. McKinsey estimated predictive maintenance at 30–50% less unplanned downtime and 10–40% lower maintenance cost. (Estimate.) In an Exotec vendor survey of 400+ warehouse workers, 98% said automation made them more productive and 49% reported pay increases thanks to automation. (Survey — vendor-commissioned, self-reported.)
Bottom line: robots measurably reduce the heaviest lifting and severe injuries, and measurably increase repetitive-strain injuries when pace rises. Both halves of that sentence matter.
11. Education
The jobs feeling the pressure: K-12 teachers, tutors, curriculum developers, teaching assistants.
The fear: In a job-board survey, 44% of education workers feared AI taking jobs. (Survey — weak: job-board survey.) Exposure is the lowest of any sector: PwC's sector analysis put education at 8% of jobs exposed. (Estimate, via secondary reporting.) Teachers' main concern is different — in Gallup–Walton reporting (2025), 57% of teachers said weekly student AI use would decrease independent thinking. The fear is about student learning and workload, not primarily job loss.
AI takes: drafting lesson plans and differentiated materials, instant practice and feedback, tutoring on routine practice, summarising student work patterns, drafting parent messages. People keep: teaching, motivating, spotting confusion and distress, managing the classroom, judging what a student actually understands, designing assessments AI cannot game, and deciding when AI is helping versus harming learning.
The outcome: In a Gallup–Walton survey of 2,000+ US K-12 teachers (2025), 6 in 10 used AI for work and 3 in 10 weekly; weekly users self-reported saving an average of 5.9 hours per week — and 82% had received no formal guidance on applying AI. (Survey — self-reported average, not logged time.) In a two-year cluster randomised trial across 18 Tennessee middle schools (Oreopoulos & Low, NBER, 2025), an AI tutor configured to coach rather than answer saw only about 15% of students with access use it regularly, because many students tried to extract answers and largely stopped using it when it refused and gave hints instead. (Measured — randomised trial.) A World Bank pilot in Edo State, Nigeria (2024) reported 6 weeks of after-school generative-AI English tutoring, with teachers in the room, producing gains equal to nearly 2 years of typical learning. (Measured — single, short, supported pilot; promising, not scalable proof.)
Bottom line: the human teacher is the engagement and judgment layer in both the successes and the failures.
12. Construction and skilled trades
The jobs feeling the pressure: electricians, plumbers, carpenters, welders, equipment operators, site supervisors.
The fear: No construction or trades-specific fear survey was found. (Gap — do not substitute a general figure.) Exposure is the lowest of any group: in an NAHB analysis of Bureau of Labor Statistics AI exposure categories and employment projections 2025–35 (released October 2026), 45 of 47 selected construction occupations — about 96% — were low or moderate exposure; only construction managers and construction/building inspectors were high, and none were very high. (Estimate — BLS categories measure whether AI could assist or complete tasks, not job-loss forecasts.) Goldman Sachs (2023) put construction and extraction at 6% of tasks automatable, and installation, maintenance, and repair at 4% — the lowest groups. (Estimate.) The near-term story is shortage, not surplus: the industry needs about 740,000 workers per year (NAHB estimate), and solar PV installers are projected to grow 37% from 2025–35 (BLS projections).
AI takes: quantity takeoffs and estimating, schedule modelling, clash detection in building models, progress capture from photos and drones, safety-camera flagging, contract and RFI summarising. People keep: building, installing, wiring, plumbing, welding, and finishing in changing site conditions; making safety judgment on the spot; coordinating trades; supervising robots and equipment; signing off quality.
The outcome: A Smartvid.io vendor case across 50+ projects claimed a 45% reduction in recordable incidents and 60% reduction in near-misses over 12 months. (Vendor-published, not independently audited.) In a Bluebeam survey of 1,000+ AEC technology decision-makers (2026), among firms using AI, 68% self-reported saving at least $50,000. (Survey — self-reported.) McKinsey estimated AI workflows could cut total project costs 15–20% and schedule overruns up to 35%. (Estimate, not a measured construction-wide result.)
Bottom line: AI is still mostly in the site office — estimating, scheduling, safety review — not on the tools. Do not promise field-productivity numbers.
13. Administrative / office support
The jobs feeling the pressure: administrative assistants, executive assistants, data-entry clerks, receptionists, office managers.
The fear: No admin-only fear percentage was sourced. (Gap — beyond the general figures in the baseline above.) Exposure is the highest of any group: Goldman Sachs (2023) put office and administrative support at 46% of tasks susceptible to automation — the highest occupation group. (Estimate.) The ILO's 2023 paper found 24% of clerical tasks at high exposure and 58% at medium exposure, with clerical support the only broad group highly exposed. (Estimate.) The World Economic Forum's 2025 report names administrative assistants, executive secretaries, data-entry clerks, and bookkeeping/payroll clerks among the fastest-declining roles. (Estimate.)
AI takes: scheduling, inbox triage, meeting notes and summaries, first drafts, form filling, data entry, expense and travel booking, searching files. People keep: deciding priorities, handling confidential and sensitive matters, managing relationships and office politics, checking AI drafts for errors and tone, following up with people, and keeping the office running when systems fail. Executive support — high trust, high judgment — is far less exposed than routine clerical work. The same title splits in two.
The outcome: In a field experiment with 6,000+ workers across 56 firms (Microsoft 365 Copilot, Dillon et al., NBER), workers with Copilot spent less time reading email; regular users saved 30+ minutes per week on email reading and completed Word documents faster. (Measured — field experiment, but email and Word time, not total job productivity or headcount.) In a small lab task (n=57), summarising a missed 35-minute meeting took 11 minutes 13 seconds with Copilot versus 42 minutes 34 seconds without. (Measured — small lab task; do not scale to a weekly saving.) In an NHS pilot of 30,000+ staff, the average self-reported saving was 43 minutes per person per working day, extrapolated to 400,000 hours per month at full rollout. (Survey — self-report plus modelling, not logged time.)
Bottom line: admin is the highest-exposure, highest near-term-change group — and the group with the clearest measured time savings on narrow tasks. Those minutes only become value when they are deliberately redirected to judgment and relationship work.
Why the numbers get better over time
Everything above is a snapshot. The partnership itself has a mechanism that improves with time — not a promise of specific results, but five reasons the scores move in the right direction when the loop is run honestly, with the human kept in it.
1. Better client models
A system tuned on a client's own conversations, customers, and patterns gets more accurate over time. Generic stays generic. A model that has seen how one business actually talks, what its customers actually ask, and where its edge cases live stops guessing and starts recognising. The improvement comes from fit, not from a bigger generic brain.
2. Faster learning
Every human correction is a lesson: an approved edit, a rejected suggestion, a flagged miss. Each one feeds back, so the system learns the client's voice and edge cases a little faster than it did yesterday. The loop is short — correction today, better draft tomorrow — and that speed is what compounds.
3. More compute at lower cost
Faster, cheaper inference means help in real time instead of after the fact. Answers arrive while the customer is still there. Checks run before the message goes out, not in a weekly review. The same partnership, running at the speed of the conversation, catches more and costs less per check than it did the year before.
4. Humans get better at partnership
People learn what to delegate, what to check, and when to overrule. The human half sharpens with practice like any tool: judgment about when to trust the AI and when to stop it is a skill, and it grows with use. The Stanford accounting study in Section 6 showed exactly this — senior accountants who intervened when confidence dropped gained more than juniors who accepted outputs at face value.
5. The loop compounds
Each cycle, AI handles more routine work, human judgment focuses on harder cases, outcomes improve, trust grows, and more can be delegated safely. Routine work shrinks, judgment work deepens, and the business keeps the people it trained — because the people are the reason the system works, not the cost the system removes.
This only works with the human kept in the loop. Remove the human check and the gains reverse: Klarna rehired humans after quality fell, and healthcare scribe savings measured from logs came in far below vendor claims. The partnership is the product. Take the partner out, and you are back to a tool making mistakes at scale.
APEX