The occupation projected to add the most jobs in the United States between 2024 and 2034 is home health and personal care aides, at 739,800 new positions, according to the Bureau of Labor Statistics table of occupations with the most job growth. It pays a median of $34,900 a year. It also sits in a category of work that barely registers in the measured record of what people actually bring to AI.
Those two facts come from different places, and neither was collected to settle the question in the headline. Laid on top of each other, they answer it better than any prediction does. What follows uses both: where the US government projects jobs will actually be added, and where AI use is actually observed rather than imagined. Five roles come out of the overlap, along with the finding that complicates the list.
What the usage data measures, and what it does not
Anthropic publishes an Economic Index that tracks which occupational tasks people bring to its AI, as distinct from which tasks a model could theoretically handle. Its Cadences report, published June 26, 2026, found that computer and mathematical occupations made up roughly 30 percent of survey respondents against a 4 percent share of US employment, and management occupations 23 percent against a 7 percent employment share. Physical categories, specifically transportation and material moving, food preparation and serving, and construction and extraction, were under-represented in the survey and in the underlying sessions alike, per the same report.
That needs one honest qualification before it gets used for anything. Anthropic states in the report that the survey is not representative of the general population, because it samples that company’s own users. So this is evidence about where AI use concentrates among people already using AI. It is not a census of the American workforce, and a category being thin in it does not prove any job is permanently beyond reach.
It is still worth more than the usual material on this question. Most claims about which jobs are safe are forecasts. This is a record of what has already happened, and it points the same way the labor projections do.
Five roles the numbers are holding up
Every employment figure below comes from the same BLS projections table for 2024 to 2034, last modified August 28, 2025. Each entry names what specifically resists, because that is the part an owner can act on.
1. The person who is physically in someone’s home
Home health and personal care aides are projected to grow from 4,347,700 jobs to 5,087,500, a gain of 739,800 and 17 percent, the largest numeric increase of any single occupation (bls.gov). The Occupational Outlook Handbook entry puts median pay at $34,900 as of May 2024, with typical entry at a high school diploma and short-term on-the-job training.
What resists here is not the conversation. It is that someone has to be in the room, notice the change nobody called in, and carry responsibility for what happens next.
2. The person cooking the food
Restaurant cooks are projected to add 217,000 jobs at 14.9 percent growth, median pay $36,830, and fast food and counter workers another 233,200 (bls.gov). Food preparation and serving is one of the three physical categories Anthropic found under-represented in its usage data.
What resists is execution under time pressure in one specific kitchen, with equipment that behaves the way it behaves rather than the way a manual says.
3. The person diagnosing a fault in place
Electricians are projected to add 77,400 jobs at 9.5 percent growth, with median pay of $62,350 (bls.gov).
What resists is that the diagnosis happens against a particular building, whose wiring no document describes accurately, and where being wrong has physical consequences. Software can help read a code book. It cannot stand in the crawlspace.
4. The person who fixes the machine
Industrial machinery mechanics are projected to add 70,700 jobs at 16.1 percent growth, median pay $63,760, and general maintenance and repair workers another 62,400 at a median of $48,620 (bls.gov).
What resists is the hands, and the repair-or-replace judgment that has to be made while the line is down and the cost of being wrong is running by the hour.
5. The person accountable on the floor
First-line supervisors of food preparation and serving workers are projected to add 73,000 jobs at a median of $42,010, and first-line supervisors of construction trades and extraction workers 49,000 at a median of $78,690 (bls.gov).
This one has a second piece of evidence behind it. In the same Anthropic survey, judgment and management were among the capabilities respondents named as ones AI lacks, and that answer came more often from respondents with more experience, per the Cadences report.
The complication that keeps this honest
Software developers are second on that same growth table: 267,700 new jobs, 15.8 percent growth, median pay $133,080 (bls.gov). That occupation sits squarely inside the most heavily represented category in Anthropic’s usage data.
So the projections do not say AI is emptying out desk work while physical work survives. Both are projected to grow, and the single most AI-saturated occupation in the country is also one of its fastest-growing. Whatever is happening, the tidy version where AI takes the computer jobs and leaves the hands-on ones is not it.
Which means a list of safe job titles is the wrong thing to carry away from this. What the data supports is narrower and more useful to an owner: the unit that gets absorbed is the task, not the job. Anthropic’s own framing makes the same distinction, separating the share of tasks already observed being done with AI from the share a model could in principle do.
The part most advice on this gets backwards
Put the two datasets side by side once more. Measured AI use concentrates in the higher-paid desk occupations. The occupations adding the most jobs, software developers aside, are lower-paid and physically present: $34,900 for home care aides, $36,830 for restaurant cooks, $30,480 for fast food and counter workers, against a median of $49,500 across all occupations (bls.gov).
The standard advice for staying ahead of automation is to move up: more credentials, more analysis, more time at a desk. The measured record runs the other direction. The work currently being done with AI is disproportionately the work of the higher-wage desk occupations.
For an owner that translates into something specific and slightly uncomfortable. The most exposed work in your business is probably the part you think of as the professional part: the quoting, the scheduling, the write-ups, the follow-up emails. The least exposed is the part that appears on your books as a cost line, the person on site.
Four moves that turn this into a decision
The first two are done on paper, in under an hour, with no software at all.
1. Write one week of one role as tasks, not as a job title. Pick the role you have been most worried about. On a single page, list what that person actually did last week, one line per task. Job titles hide the split that matters; a week of lines does not.
2. Mark every line as either has-to-be-there or not. Has-to-be-there means physically present, or accountable in the room with a customer. This is the only split the evidence above actually supports, and it is the one that survives contact with your real schedule.
3. Put AI on the not-there lines, highest frequency first. Not the most annoying task, the most frequent one. Frequency is where the hours are, and the annoying task is usually annoying because it is rare and unfamiliar.
4. Decide what the freed hours are for, and write it down before you start. If the answer is fewer hours paid, you have decided to run a smaller business. If the answer is a job you currently turn down, you have decided to run a bigger one. Both are real choices. Only one of them is a strategy, and making it after the fact is how owners end up with the first while believing they chose the second.
What the third move costs
Three options an owner can price today, all checked on the vendors’ own pricing pages on August 2, 2026.
Claude runs a free tier at $0, Pro at $17 a month on the annual plan with $200 billed up front or $20 billed monthly, and Team at $20 per seat per month billed annually or $25 monthly, for teams of 2 to 150. Google Workspace is the sensible pick if the paperwork already lives in Gmail and Docs: Business Starter at $7.00 per user per month, Standard at $14.00 and Plus at $22.00, with a new-customer discount running at the time of writing, so check the live page before you budget. Otter.ai is the narrower buy, worth it only if your recurring cost is writing things up after site visits and calls: Basic is free with 300 transcription minutes a month, Pro is $16.99 monthly or $8.33 a month billed annually, and Business is $30 monthly or $19.99 a month billed annually.
Pick one and point it at a single task from move three. A subscription bought against a specific line on your page is a tool. One bought against the general idea of AI becomes a renewal you forget to cancel.
Where the evidence stops and the argument starts
Everything above this heading is published data or a price you can check yourself. What follows is judgment, and it is the part I would argue for hardest.
The five roles are not holding because a machine could never in principle do them. They are holding because the work is bound to a place and to a person who answers for it. That binding is an asset, and it is the one most likely to get spent by accident.
The accident runs like this. An owner reads a piece like this one, correctly concludes the desk work is the exposed part, moves AI onto it, and then treats the saving as a headcount question. That is the version where a business gets smaller and files it under efficiency. The version worth having is the one where the paperwork stops eating the evenings and the same people take on the work the shop currently turns down. The capacity you free is an argument for what you can now say yes to, not a case for needing fewer people.
The quiet half of this is keeping the people you already have, which we have written about separately in Stop Employee Turnover: AI Cuts Burnout and Boosts Retention. If your answer to move four is a person stepping up into more, getting them there is its own piece of work, laid out in Professional Development Plan: A Small Business Playbook. And for the longer arc this sits inside, see AI’s Next Decade: A Small Business Playbook.
Pick the role you have been most worried about and write last week down, one task per line. Then count how many lines say has-to-be-there. That number, not a list of safe job titles, is the thing you actually own.
Frequently Asked Questions
Which jobs are safest from AI right now?
The occupations adding the most US jobs through 2034 with the least measured AI use are in-person and physical, led by home health and personal care aides at 739,800 new jobs, the largest gain of any occupation (bls.gov). Treat that as a task-level answer rather than a job-title one: software developers are second on the same table, so heavy AI exposure and job losses are not the same thing.
Does AI take the higher-paid or the lower-paid work first?
The measured record points at higher-paid desk work. Anthropic’s June 26, 2026 Economic Index report found computer and mathematical occupations at roughly 30 percent of respondents against a 4 percent employment share, and management at 23 percent against 7 percent, with physical categories under-represented (anthropic.com). That sample is drawn from AI users rather than the general population, which the report states directly.
What should a small business owner hand to AI first?
The recurring paperwork attached to a role, not the role. Quoting, scheduling, write-ups and follow-up emails are the tasks that repeat most often and require nobody to be physically present. Start with the most frequent one rather than the most irritating one.
Do these BLS projections already account for AI?
They are the agency’s own published estimates for 2024 to 2034, last modified August 28, 2025 (bls.gov). They are projections rather than guarantees and they get revised, so use them as the best official baseline available, not as a forecast that has already priced in every technology shift.
