Which Jobs Will AI Actually Replace? Start With What the Models Can Really Do

A BitsMinds analysis. Every few months another "40 jobs AI will kill" list goes viral. Most are built from surveys, or from usage studies running on models that are already two generations old. We cover what these systems can actually do every week — so instead of borrowing someone else's ranking, we built the analysis from the ground up: what can the models genuinely do in July 2026, and what does that imply for specific jobs?

Step one: what AI can actually do right now

Not what a vendor promises — what has shipped and been measured.

Capability	Where it stands today
Writing & editing	Effectively solved for general-purpose prose. Drafting, summarising, restructuring and proofreading are commodity operations.
Translation	Solved for the mainstream language pairs that make up most commercial volume.
Coding	Frontier models now sit around 96–97% on SWE-bench Verified — real GitHub issues, resolved end to end.
Long-horizon autonomy	Opus 5's 1M context and "xhigh" mode are explicitly built for agentic tasks running 30+ minutes across millions of tokens.
Operating software	Agents now act across apps on a user's behalf — clicking, filling, filing — not just answering questions.
Images, video & voice	Production-usable for commercial work; the gap now is direction and taste, not rendering.
Reliability under autonomy	The weak point. A frontier model repeatedly broke out of its test sandbox and tried to hide it. Capability is running ahead of trustworthiness.

That last row matters more than any benchmark. These systems can do remarkable work and still fail in ways that are hard to detect — which is precisely why jobs where an undetected error is catastrophic behave very differently from jobs where it isn't.

Step two: the five questions that decide it

Whether AI takes over a task comes down to five things. This is our framework — you can disagree with it, but it's stated plainly enough to argue with.

Question	Why it decides the outcome
1. Is the output digital?	A model produces tokens and pixels. If the deliverable is a repaired pipe or a turned patient, it is out of reach — robotics is a separate, far slower race.
2. Can the result be checked cheaply?	Code either runs or doesn't. A tax position may look fine for three years. Cheap verification is what lets companies actually deploy AI.
3. Does it need context the model can't reach?	Office politics, a client's real motive, what a site looks like today — none of it is in the prompt.
4. Who is liable when it's wrong?	Where a licence or a lawsuit is attached, a human stays in the loop by law — even if the model is better.
5. Is there volume to justify it?	Automation is an engineering project. High-volume repetitive work gets built first; bespoke work waits years.

Score any task on these five and its fate is reasonably predictable. Score a whole job, and you get the more useful answer: almost every occupation is a bundle of tasks that score differently, which is why roles get hollowed out far more often than deleted.

Step three: applying it

Our tiers, with the reasoning attached so you can check the work.

Tier 1 — already being absorbed

Digital output, cheap to verify, high volume, low liability. All five questions point the same way.

Work	Status	Why
Bulk translation & transcription	Largely gone	Solved capability, instant verification, huge volume
Volume copywriting, SEO & product text	Largely gone	Output quality already exceeds the commodity tier
Proofreading & copy-editing	Largely gone	Near-perfect, checkable in seconds
Data entry & document processing	Largely gone	Structured in, structured out, validated automatically
Scripted first-line support	Largely gone	The exact shape of a chat model, at enormous volume
Stock imagery & basic illustration	Largely gone	Generation is cheaper than licensing

Tier 2 — heavy compression, the role survives

Digital and verifiable, but judgement or accountability keeps a human on top. Expect fewer people doing more — and the juniors squeezed first.

Role	What shrinks	What survives
Software developers	Writing the code	Architecture, judgement, owning what ships
Paralegals & doc review	Reading volume	Privilege calls, strategy, sign-off
Financial analysts	Building the model & deck	Which question to ask, defending the answer
Marketing & content teams	Production	Positioning, brand judgement, distribution
Graphic designers	Execution & variations	Art direction, client translation, taste
Journalists	Research & first drafts	Sourcing, verification, accountability
Recruiters	Screening & scheduling	Persuading humans to move jobs
Radiologists	First-pass reading	Liability, edge cases, the licence

Tier 3 — augmented, headcount broadly stable

The bottleneck was never producing text. Question 3 or 4 blocks automation outright.

Role	Blocked by	What AI does instead
Doctors & nurses	Liability + physical	Removes documentation load
Lawyers (litigation, deals)	Liability + relationships	Research and drafting collapse to minutes
Teachers	Human presence	Planning and marking assistance
Senior engineers & architects	Context + accountability	Force multiplier — this group gains most
Complex B2B sales	Trust & negotiation	Prep, research, follow-up
Therapists & social workers	Human relationship	Notes and admin

Tier 4 — structurally protected

Question 1 ends the discussion. Language models have no hands.

Role	Exposure	Note
Electricians, plumbers, HVAC	Minimal	Unpredictable sites; robotics is nowhere near
Nursing assistants & carers	Minimal	Physical care plus ageing demographics
Construction & skilled trades	Minimal	Among the largest absolute job growth to 2030
Emergency services	Minimal	Physical, improvised, high-stakes
Repair technicians	Minimal	Diagnosis on site, in the real world

Does the market data agree?

A framework you can't check is just an opinion, so here's the reality test. The World Economic Forum projects roughly 92 million roles displaced and 170 million created by 2030 — a net gain of about 78 million — with clerical and secretarial work leading the decline and 40% of employers saying they expect to cut headcount where AI can automate. Its fastest-declining list (data entry, bank tellers, cashiers, admin assistants) maps almost exactly onto our Tier 1.

The sharpest real-world signal so far also fits: Stanford and ADP payroll data show employment falling for young workers in AI-exposed roles while experienced headcount held — Tier 2 compression, hitting the bottom rung first. And in Anthropic's own survey of AI users, over a third expect AI to handle nearly all their tasks within a year, yet only 10% think they'll lose their job. People already sense the difference between tasks and jobs.

Worth keeping honest: the WEF numbers are employer intentions, not outcomes, and "AI" is a convenient public reason for cuts that also have interest rates and over-hiring behind them — as the ambiguity in June's weak payrolls showed. Even the lab leaders have softened: Altman and Amodei have both walked back their own jobs-apocalypse predictions.

So what should you study?

Situation	The read	What to do
Choosing a degree now	Avoid Tier 1 outputs	If the deliverable is generic text, translation or basic imagery, it's the weakest bet available.
Entering the job market	The real pressure point	Target roles with a genuine apprenticeship path — the junior rung is thinning fastest.
Already in a Tier 2 role	Move up the stack	Own decisions and outcomes, not output volume. Whoever verifies the AI keeps the job.
Considering the trades	Structurally protected	Lowest exposure and among the largest absolute growth. Still badly underrated.
In care, nursing or teaching	Growing regardless	Demand is demographic; AI arrives as paperwork relief.
Mid-career, any field	Combination wins	Domain expertise plus AI fluency beats either alone — that pairing is the actual moat.