Jobs AI Won't Replace: Automation Exposure vs Pay Across 20 Careers

The short answer

Nothing we track is safe because AI cannot do the tasks. Things are safe because someone has to be accountable for the answer. That is the pattern across all twenty careers on this site, and it is why the actuary - whose work is literally statistical modelling, the thing machines are best at - scores low exposure while the data analyst writing SQL scores medium.

An actuary's reserve opinion has to be signed by a credentialed human under state insurance law. A CPA's attestation opinion has the same property. A financial planner's value was never portfolio construction, which robo-advisors commoditised a decade ago; it is sitting with a widow and reorganising her finances. Accountability, liability and the relationship are the moat. Task difficulty is not.

Of the 20 careers, 8 score low exposure and 12 score medium. None scores high - which is a statement about our sample, not about the economy. We only cover careers worth switching into, and a career already being automated away is not one.

The table below puts each career's exposure next to its mid-career pay, ten-year growth and annual openings, so you can see which safe jobs are also worth having. Several are not.

Automation exposure vs pay: all 20 careers

Source: bls.gov/ooh/

Read this table by column, not by row. Exposure on its own tells you almost nothing useful: private equity associate and actuary both score low, and one pays $450,000 at mid-career while the other pays $105,000 with 1,500 openings a year nationwide.

The interesting structure appears when you cross exposure with pay. The median mid-career figure across the twenty is $135,015, which splits the set exactly ten and ten. Combine that split with the low/medium exposure score and you get four quadrants, each with a different problem.

One caution before the quadrants: the openings column comes from BLS proxy occupation codes, and several careers share a code. Software Developers (15-1252, about 106,100 annual openings) stands behind software engineer, machine learning engineer and AI engineer; project management specialists (about 76,500) stands behind both product manager and project manager. Do not add the openings column. Those are the same seats counted more than once.

The four quadrants

Here is the summary before the detail. Growth and pay figures below are medians within each quadrant, computed from the twenty careers in the table above.

Source: bls.gov/ooh/

Quadrant 1: the compounding seats

Five careers: private equity associate ($450,000), quantitative analyst ($400,000), product manager ($230,000), machine learning engineer ($200,000) and AI engineer ($155,000).

These score low because the work is deciding under uncertainty with money or a product on the line, and because AI tooling makes each of these people more productive rather than redundant. Quants are the sharpest case: they automate other people's jobs for a living, and better tooling raises their output. The catch is that it also shortens the half-life of any given trading signal, so the treadmill runs faster.

The reason this quadrant is not advice is supply. Private equity associate seats number roughly 3,000 a year in the US and quant seats a couple of thousand, and both recruit almost exclusively from investment banking, top-tier engineering and quantitative PhD pipelines. These are not careers you switch into at 34 by taking a certificate.

The two that are genuinely enterable are ML engineer and AI engineer, both of which sit behind the Software Developers code with its 106,100 annual openings. Note the honest caveat on AI engineer: low automation exposure, high *obsolescence* exposure. The role is three years old, definitionally unsettled, and dependent on a model-provider ecosystem that could consolidate. Low automation risk is not the same as low risk.

Product manager is the odd one out. Its core - deciding what to build and getting a cross-functional group to agree - is low exposure, but its clerical layer (status reports, ticket grooming, first-draft specs, competitor summaries) is among the most automatable white-collar work there is. Expect fewer PMs per engineer, doing more of the deciding.

Quadrant 2: the exposed premium

Five careers: investment banking analyst ($375,000), tech sales account executive ($180,000), management consultant ($145,000), data scientist ($140,000) and software engineer ($135,980).

These pay above the median and score medium, and in every one of them the exposure is concentrated in the first two years of the career - the exact rung an outsider would enter on.

Investment banking is the clearest example. Comps pulls, formatting, first-draft memos and data-room administration are precisely what AI tooling is being pointed at, and several banks have publicly discussed smaller analyst classes. Structuring and negotiation are not close to automated; the question is how many juniors a bank needs to get to the seat where you do those.

Management consulting has the same shape. Secondary research, first-draft slides, benchmark gathering and transcript synthesis used to absorb an analyst's first two years and now take an afternoon. In tech sales, AI SDR and prospecting tools are compressing top-of-funnel headcount, which is where beginners enter, while complex multi-stakeholder closing remains hard to automate. BLS projects essentially no employment change for the occupation across 1.57 million jobs, with about 123,400 annual openings driven almost entirely by turnover.

Software engineer deserves its own sentence, because it is the career people ask about most. BLS projects +10% growth to 2035 with 106,100 annual openings. Indeed's software-development postings index stood at 76.1 on 4 September 2026 against a February 2020 baseline of 100, while all US postings stood at 102.4. Both facts are true. Software hiring is roughly a quarter below pre-pandemic levels, the gap falls hardest on people with no professional experience, and the ten-year projection still looks good.

Data scientist is the growth outlier of the whole table at +35% to 2035 (BLS, 15-2051, 24,800 annual openings). Do not read that as an easy door. The code absorbs senior analysts and applied scientists, so the growth is concentrated well above the entry line, and entry-level postings across the economy were down 7.5% year over year as of May 2026 (Indeed Hiring Lab).

Quadrant 3: the licensed moat

Three careers: cybersecurity analyst ($124,000), financial planner ($115,000) and actuary ($105,000).

This is the quadrant with the clearest logic and the smallest doors. Each one is defended by something an AI system cannot supply: an incident someone has to own, a fiduciary relationship, or a signature required by statute.

The actuary is the purest case. Machine learning expanded the job - predictive analytics in pricing and underwriting is now core, which is why the SOA added Predictive Analytics and Advanced Topics in Predictive Analytics to the ASA pathway - without touching the credentialed signature. Statements of actuarial opinion on loss reserves must be signed by a qualified actuary under state insurance law. The price of that protection: about 1,500 openings a year nationally on a base of 31,200 jobs, and an exam sequence measured in years. See the SOA exams P and FM for what the first two cost you.

Cybersecurity analyst has the best ten-year outlook of any large computing occupation at +21% to 2035, and tooling that automates alert triage raises rather than lowers demand for people who can investigate and respond. But the growth sits above the entry line, tier-1 SOC roles attract enormous applicant pools, and the "skills gap" everyone cites is real at three-plus years of experience and much softer at zero.

Financial planner inverts the usual reading. BLS projects only +1% growth to 2035 - and about 17,100 openings a year, almost all replacement demand, because the advisory workforce is old and retiring. The opportunity is real; it comes from people leaving, not from the profession expanding. The CFP is the entry ticket.

Quadrant 4: the squeezed middle

Seven careers: cloud / DevOps engineer ($134,050), FP&A analyst ($110,000), UX / product designer ($104,000), project manager ($102,320), CPA ($94,750), data analyst ($88,000) and real estate agent ($52,830).

The biggest quadrant, the most openings, the lowest median growth at +7%, and the weakest defence. It is also where most career changers are aiming, which is the uncomfortable point of this whole exercise.

The mechanism is identical in six of the seven: the administrative half of the job is being absorbed, and the administrative half is what juniors used to do. Status reports and RAID logs for project managers. Data consolidation, variance commentary and reconciliation for FP&A, now inside Anaplan, Workday Adaptive and Pigment. Manual provisioning for cloud engineers, eliminated by infrastructure-as-code. Bookkeeping, transaction matching and first-pass audit sampling for accountants. Query writing for data analysts - which is why our honest read on the Google Data Analytics Certificate says the certificate teaches precisely the part of the job that is compressing fastest.

CPA is the one to separate out, because the exposure score hides a widening split. Unlicensed accounting clerk roles are shrinking while licensed, judgment-heavy roles grow, and the occupation still turns over about 115,300 openings a year on a base of 1.6 million. The CPA licence is the difference between being on the shrinking side and the growing side of the same medium score.

UX / product designer is the most damaged entry market on the list. Generative tools now produce passable screens, junior postings attract 500-800 applicants, and CareerFoundry - one of the largest UX bootcamps - entered insolvency in late 2025 and ceased operating in early 2026.

Real estate agent is the worst cell in the table on every axis: lowest mid-career pay at $52,830, +2% growth, and 40,400 annual openings that exist because people keep quitting. Portals and iBuyers already absorbed search and valuation. We wrote the full reality check on agent income separately, because the first-year number deserves its own page.

How we scored automation exposure

The exposure score is ours, and it is a judgment, not a model. We are saying so plainly because most "AI risk by occupation" tables on the internet trace back to one of two academic papers from 2013 and 2023 and do not say which.

Each career is scored low or medium on one question: *what share of the work would survive if the routine production of the output were free?* Three things push a career toward low:

1. A legal or fiduciary signature. An attestation opinion, an actuarial reserve opinion, a fiduciary recommendation. Only a licensed human can supply it.
2. Accountability for a live outcome. Someone owns the incident, the number in front of the board, the decision to ship.
3. A relationship the client will not delegate. Negotiation, multi-stakeholder closing, behavioural coaching.

Two things push it toward medium: the entry rung consists of routine production (comps pulls, status reports, reconciliations, first-draft SQL), or the output is already available for free from a consumer product.

The pay, growth and openings columns are not judgments. They come from the career data files on this site: salary.mid_median, outlook.growth_pct_10yr and outlook.openings_per_year, each sourced on its own career page, with BLS Occupational Outlook Handbook 2025-2035 projections as of September 2026 wherever a clean occupation code exists.

What this table cannot tell you

Four caveats, and they matter more than the table.

The openings column double-counts. Three careers share the Software Developers code and two share project management specialists. The column is useful per row and meaningless in aggregate.

Several careers have no BLS code at all. Product manager, AI engineer, ML engineer, quantitative analyst, tech sales AE and private equity associate are all mapped to proxy occupations that are broader than the job. The private equity openings figure of 3,000 is our order-of-magnitude estimate of US associate seats, not a BLS number, and we label it as such on the career page.

Exposure is scored for the occupation, not for you. Within any medium-exposure career, the person who owns a system, a client or a signature is in a different position from the person who processes tickets. The score describes the average seat.

Ten-year projections were made before the last eighteen months of tooling. BLS projections to 2035 are a structural forecast, not a nowcast. Where a near-term indicator contradicts them - the software postings index at 76.1, entry-level postings down 7.5% - we publish both and let them disagree.

Finally, nothing here scores high. If your question is "which jobs are actually disappearing", this table does not answer it, because we only track careers we think are worth switching into. Data entry, basic bookkeeping, tier-1 support scripting and template copywriting are not on this list for a reason.

What to do if you are already in a squeezed-middle career

Move toward the accountability, not away from the tools.

The practical version of that for each of the three most common cases on this site: if you are a data analyst, stop competing on query writing and start owning a metric a director is judged on - the Google Data Analytics Certificate plus a Power BI credential gets you to the table, but the ownership is what keeps you there. If you are an unlicensed accountant, get the CPA; it is the single clearest example on this table of a credential that moves you from the shrinking half of an occupation to the growing half. If you are a coordinator-level project manager, the administrative half of your role is the half being absorbed, so move toward scope, budget and the conversation where someone has to be told what is getting cut.

And be sceptical of anyone selling you an "AI-proof career" for $199. The two most defended careers on this table - actuary and CFP - take years of exams and have 1,500 and 17,100 annual openings respectively. Durability is priced. That is what makes it durable.