Table of Contents
In March 2023, Goldman Sachs published a research note estimating that generative AI could expose the equivalent of 300 million full-time jobs to automation. Two years on, the more useful story isn’t the headline number. It’s which jobs are quietly changing shape, and how differently each one is responding.
Some roles are losing their entry-level rung. Others are gaining a productivity layer that senior staff now refuse to work without. Three kinds of change keep showing up:
- Task compression. The boring 60% of a job gets faster, so the remaining 40% takes up more of the day.
- Junior work evaporates first. The tasks that used to train newcomers are the ones models handle best.
- Judgment becomes the product. Anyone whose value was mostly throughput is being repriced.
Here are the top 7 jobs that artificial intelligence could transform, ranked roughly by how visible the shift already is.
1. Customer Support and Service Agents
Klarna made the clearest public bet. In February 2024 the company announced its AI assistant had handled 2.3 million conversations in its first month, about two-thirds of all customer service chats, doing work equivalent to 700 full-time agents. Average resolution time fell from 11 minutes to under two.
That is one company and one product, and the numbers deserve scepticism. But the direction is hard to argue with. Password resets, order tracking, refund eligibility, subscription cancellations: all of it is pattern-matching against a policy document, which is exactly the shape of problem large language models solve cheaply.
What survives the cut
Escalations. Anything involving an upset customer, a borderline refund, or a situation the policy manual never anticipated. Companies are learning that a confident wrong answer costs more than a slow correct one, and the humans who clean up after automated mistakes are now the senior people on the floor. Surveys suggest customers are more open to this than managers expected, as the data on how much autonomy people will trust AI with shows.
2. Radiologists and Medical Imaging
This one has been predicted, wrongly, for over a decade. Geoffrey Hinton said in 2016 that we should stop training radiologists. Radiology is still hiring. What changed is the reading itself: the FDA has cleared well over 700 AI-enabled medical imaging devices, and tools that flag lung nodules or prioritise mammograms are now routine in large hospital systems.
The practical effect is a shift in what a radiologist’s day looks like. Less time hunting for a lesion on a 400-slice CT scan. More time on ambiguous cases, recommendations, and the conversation with an oncologist at 6pm. Liability still sits with the human who signs the report, and that signature is not going anywhere soon.
3. Software Developers
Coding assistants are the most widely adopted AI tool in any profession. GitHub’s own research found developers completed a defined task about 55% faster with Copilot; a later field study at Accenture found a smaller but still real gain of roughly 26%. Those two numbers tell you something useful about lab results versus office reality.
Where the job is moving
Boilerplate is going. So is the small stuff that used to fill a junior developer’s first year: writing CRUD endpoints, wiring up forms, chasing a typo in a config file. What’s growing is review, system design, and the unglamorous work of deciding whether the generated code is actually correct. Teams that shipped faster in 2025 mostly shipped more bugs too, and debugging AI output is now a named skill on job listings.
4. Financial Analysts and Accountants
Accountants spend an enormous share of their week on reconciliation, variance checks, and pulling numbers out of documents. All three are automation targets, and the big four have poured billions into tools that do them. Entry-level audit work in particular is thinning out.
The interesting shift is upward, into interpretation. A model can produce a variance table in seconds; explaining why margin moved 300 basis points in Q3 still requires someone who knows the business. That translation layer, from raw numbers to a decision a board can act on, is where the value is migrating. It’s the same instinct described in Yong Wang’s approach to turning visualization into insight, where the chart is never the point.
5. Graphic Designers and Illustrators
Image generation hit this field harder and faster than almost any other. Stock illustration marketplaces saw submissions and freelance demand drop sharply through 2023 and 2024, and the freelance platforms noticed it first. If your paid work was generic: a hero image, a blog header, a simple icon set, the market for it compressed within about eighteen months.
Designers with a distinctive style, client relationships, and the ability to art-direct a brand are in a different position. They use generative tools for mood boards and roughs, then do the part the model can’t: making choices that hold together. The unresolved argument over credit, consent, and what counts as authorship is worth reading up on, and this look at the problem with AI artists gets at why the debate refuses to settle.
6. Paralegals and Legal Researchers
Document review was already being outsourced to e-discovery software before large language models arrived. What’s new is drafting and research: summarising a 200-page deposition, comparing contract clauses across a portfolio, assembling a first-pass memo. Tools built on top of legal databases now cite the specific paragraph rather than hallucinating a case name.
That distinction matters because of what happened in Mata v. Avianca. In 2023 a New York judge sanctioned lawyers who filed a brief citing six fabricated cases that a chatbot had invented. The lesson wasn’t that AI can’t do legal work. It was that verification is not optional, and the person who verifies is still a human being with a licence on the line.
7. Teachers and Tutors
Classrooms are where the productivity gains are least visible and the stakes are highest. Teachers use AI to draft differentiated worksheets, generate practice problems at three difficulty levels, and turn a rubric into feedback on 30 essays at 11pm. Tools like Khanmigo put a Socratic tutor in front of students who’d never have access to a private one.
None of that replaces the job. A model can explain long division forty different ways. It cannot notice that a quiet kid stopped eating lunch, or decide that the lesson plan needs to be abandoned because the room is restless. Teaching is one of the few roles on this list where the human part is not the residue left over after automation, it’s the entire reason the job exists. Anyone building a career in a field like this should be thinking about which parts of their work are genuinely irreplaceable, an idea worth sitting with if you’re weighing whether to chase a role in AI itself or build somewhere the human element stays load-bearing.
The teachers who come out ahead of this aren’t the ones who ignore the tools or the ones who hand over their judgment to them. They’re the ones who use the saved hour to do something a machine still can’t: look a student in the eye and figure out what’s actually going on.
