Jobs AI will replace: what 200,000 real work conversations actually show.

The number everyone quotes — and the half that gets dropped

The World Economic Forum surveys employers directly — over 1,000 companies across 22 industries and 55 economies — and asks what they expect to do with their workforce. The 2025 edition projects that by 2030, 92 million roles will be displaced while 170 million new ones are created: a net increase of 78 million. It also projects that 22 percent of jobs will be structurally transformed and that roughly 39 percent of the skills people use today will be out of date within five years.

A net gain of 78 million sounds reassuring, and it is exactly as misleading as the 92 million on its own. Net figures hide the fact that the person losing a role and the person gaining one are rarely the same person, in the same place, with the same skills. A net-positive decade can still be a brutal one individually.

What the number does not tell you at all is which part of your work is in play. For that you need to look at what people actually hand to a machine when nobody is watching — which is precisely what the next study did.

What people actually give an AI at work

Instead of asking experts what AI might do, Microsoft Research took 200,000 anonymised conversations people had with Bing Copilot and classified, for each one, which work activities from the O*NET occupational database were being assisted or performed. It is one of the few studies in this field built on observed behaviour rather than opinion.

The finding, in their words, is that the most common and most successful AI-assisted activities involve information work — the creation, processing and communication of information. And because most occupations contain some information work, the applicability spreads far wider than the usual “tech jobs” framing suggests.

There is a second, subtler result in the same data: what users ask for and what the AI does are not the same thing. People most often come to gather information and to write. The model most often ends up providing information, writing, teaching and advising. In other words, it slides towards the explaining end of the work, not the deciding end.

Which occupations sit closest to it

Ranked by how much of their work overlaps with what the models are actually doing, the occupations at the top are the ones built almost entirely from language: interpreters and translators, historians, writers and authors, sales representatives of services, customer service representatives. At the bottom are jobs built from physical acts — surgery, moving objects, operating equipment on site.

Now the part the headlines cut. The researchers state plainly that the score is not a displacement forecast. Their own wording: “A job is far more than the collection of tasks that make it up” — and, more directly still, that the study “does not draw any conclusions about jobs being eliminated”. The reason is structural: a task database lists what gets done, not the interpersonal judgement, domain expertise and ethical weight that make someone accountable for doing it.

So when you next see a list titled “the 40 jobs AI will replace”, you are looking at a list of occupations whose tasks overlap heavily with current model ability — published by researchers who explicitly said it is not that list.

Which of your tasks are actually exposed?

Averages across occupations tell you nothing about your Tuesday. So here is the same logic applied to a week rather than a job title: tick the tasks that make up most of your working time. The result is not a verdict on your career — it is the share of your week that current models already handle well.

Most AI use is still not replacement

There is a second question hiding underneath “what can AI do”: when people use it, are they handing the task over completely, or working alongside it? Anthropic publishes an index that classifies exactly this across real usage. In the data from November 2025, augmentation accounted for 52 percent of conversations and automation 45 percent — augmentation up five points, automation down four.

The longer trend is more interesting than the snapshot. A year earlier the split was 55 to 41 in favour of augmentation; by March it was 55 to 42. In other words automation has been slowly gaining ground across the period, then gave some back. This is not a straight line to full automation, and anyone selling you one is describing a mood, not a measurement.

Read those two studies together and the picture sharpens considerably. The tasks most exposed are informational. The dominant mode of use is still one where a human stays in the loop. What is being sold off is not the job — it is the first draft.

What has already gone — from my own trade

I produce video and brand material, and over 1,500 of those videos were made with AI in the chain, with roughly 500 million views behind them. That gives me no authority over the labour market in general, but it does give me an exact view of one trade. Four things genuinely disappeared, and they are all tasks, not roles:

1. Stock photography. Buying a generic image of a generic person in a generic office is over. Generated imagery does it in series, with consistent lighting, and matched to the actual brief instead of approximately.
2. Booking a voice for a second language. A dubbed version used to mean a studio, a talent and a schedule. It is now made from the same audio track, including lip movement.
3. The variant loop. Twenty versions of an advert used to take an agency weeks. It now takes an afternoon — which moved the entire bottleneck from producing to choosing.
4. The rough first cut of a script. Not the script. The blank page in front of it.

Notice what is missing from that list: the client conversation, the decision about what the film is for, the judgement about which of the twenty versions represents the company, the responsibility when something is wrong. Those did not get cheaper. If anything, they became the entire job.

Which jobs are safe from AI

“Safe” is the wrong word — nothing is untouched, because every job has some information work inside it. The useful question is which properties make work resistant, and the studies agree on three:

This is why the “safe careers” lists you find online are misleading in a specific way: they name industries, when the property is what matters. A nurse is not safe because healthcare is safe — a nurse is safe because of presence and accountability. And a marketing manager is not doomed because marketing is exposed; the exposure sits on the drafting part of the week, not the deciding part.

Which leads to the only career advice in this text worth the words: move your week toward the three properties above. Not by learning a new profession — by changing the proportion of your existing one.

If you sell something, this is the part that concerns you

There is a commercial consequence that gets less attention than the jobs question, and it hits every business that produces anything visible. If a competent first version of nearly anything is now close to free, then anything that looks like a competent first version has lost its market value. Average is no longer a price point.

You can see this in output already. Models are trained on averages, so unguided material lands exactly on the mean of its category — polite, correct, interchangeable. That is the actual tell of AI-made work in 2026: not artefacts or six fingers, but the absence of edges. And an audience does not need to identify the cause to feel the effect.

So the differentiator moves to the one thing the tools do not supply: knowing which of the twenty options belongs to this company and no other. That is a judgement, it is learnable, and it is the reason the work I get paid for shifted over four years from producing footage to deciding what should exist.

What to do on Monday

Not a five-year plan. Three steps, all completable this week:

1. Write down where your week actually goes. Not what your job title says — the hours. Most people have never done this, and it is the only input that matters. The tool in section 04 does it in two minutes.
2. Take the single most exposed task and hand it over for four weeks. Measure the time it took before, and after. Not to save money — to find out where the output gets worse, because that boundary is your actual value.
3. Spend the recovered hours on the three properties. Presence, accountability, decisions. In practice this usually means more contact with clients and fewer hours in a document.

If the honest result is that your week is almost entirely exposed tasks, that is worth knowing now rather than in two years. It is also the most fixable problem in this whole article — because the fix is a reallocation of hours, not a new career.