Will AI take my job?

Almost every alarming forecast about AI and employment shares one flaw: it treats a job as an indivisible unit that either survives or does not. Real jobs are bundles of tasks, and automation has historically taken tasks rather than whole occupations.

Consider what happened to accountants after spreadsheets. The tedious arithmetic disappeared almost entirely. The profession did not — it grew, and shifted towards analysis and advice. The task changed; the job re-formed around what was left.

That framing is more useful than a yes or no. Ask which tasks in your week are routine transformations of information, because those are the exposed ones, and what remains once they are gone.

The abstract question is unanswerable; the concrete one is not. Log your work for a week in half-day blocks and label each block by what it really consisted of — producing a routine output, deciding something, persuading someone, or handling a situation nobody had specified.

Then ask of each block whether the inputs were fully written down. Tasks where everything needed is in a document are the exposed ones. Tasks that depend on knowing who to ask, what happened last quarter, or what the client actually meant are not, because the necessary context is not available to be automated.

Most people discover their week is more mixed than they assumed, and that the exposed portion is real but smaller than the headlines imply. That is a more useful starting point than a percentage from a forecast about your job title.

Being specific about capability is more useful than speculation. Today’s systems are strong at producing plausible drafts, summarising and restructuring text, writing routine code, translating, and pattern-matching across large volumes of documents.

They are weak wherever the cost of being confidently wrong is high, where context lives in people’s heads rather than documents, where physical presence is required, and where the work is fundamentally about persuading, negotiating or holding responsibility.

Notice that this cuts across seniority rather than along it. Parts of a senior job can be highly exposed and parts of a junior job barely exposed at all, which is why occupation-level predictions are so unreliable.

• Most exposed: routine content production, first-line support, basic data entry and reconciliation, template-driven documentation
• Partly exposed: junior software work, paralegal research, entry-level analysis, first-draft design
• Least exposed: skilled trades, healthcare delivery, work requiring accountability or negotiation, roles built on relationships and trust

The most credible concern is not mass unemployment. It is that AI is best at precisely the tasks juniors were historically hired to do — the first draft, the research summary, the simple ticket, the boilerplate.

That breaks the traditional apprenticeship ladder. If nobody pays a junior to do the routine work, the profession loses the mechanism by which people become senior. This is a structural problem the industry has not solved, and if you are early in your career it is worth taking seriously rather than dismissing.

The practical response is to get closer to the parts of the work that were never routine — ownership of an outcome, direct contact with users or customers, judgement under uncertainty — earlier than previous generations had to.

Not "learning to code" or any other single skill. What holds up is a combination that is hard to unbundle: domain knowledge, judgement about what is worth doing, and the ability to take responsibility for a result.

Notice that the person who uses these tools well tends to displace the person who refuses to, far more reliably than the tools displace either. That has been the pattern with every previous productivity technology, and it is the most actionable thing in this whole discussion.

There is a quieter advantage too. Being the person in a team who can tell when a generated answer is subtly wrong is increasingly valuable, and that ability comes from domain depth rather than from tooling familiarity — it cannot be acquired quickly by someone outside the field.

• Deepen domain expertise — AI is general, and generality is cheap
• Move towards owning outcomes rather than producing outputs
• Get fluent with the tools in your own field rather than avoiding them
• Build the relationships and reputation that no model has access to
• Keep a written record of what you have actually delivered

Anyone giving you a confident percentage for how many jobs will vanish by a specific year is guessing. The historical record on technology forecasts is poor in both directions: predicted catastrophes that did not arrive, and disruptions nobody saw coming.

What is reasonably well supported is narrower and more useful. Task composition within jobs is shifting quickly. Demand is rising for people who can direct these tools competently. And the transition is uneven — some people and regions absorb it far more easily than others.

That unevenness is the part most worth planning around. Aggregate outcomes being fine is little comfort if your particular role, employer or region is on the wrong side of the distribution, and that is a risk you can actually act on.