AI Is Changing the Job. Who Is Changing the Deal?

The sentence that hides the decision

“We have decided to move forward with another candidate.” Sometimes that means a role was narrowly filled, a budget changed or a process simply ended. But a larger question now sits behind many hiring decisions: if AI absorbs the routine work through which people become experienced, what will a junior employee learn by doing?

The clearest answer is that AI is changing many jobs by rearranging tasks. Drafting, searching, summarising, checking and routine analysis may be assisted or partly automated while judgement, accountability, relationships and exception-handling remain important. That can raise productivity without eliminating an occupation. It can also reduce the number of junior openings through which people acquire those capabilities.

Exposure is not a redundancy notice

The International Labour Organization’s current guidance makes the essential distinction: exposure measures what AI could do to tasks under a relatively static view. It does not show whether a company can automate profitably, whether it will choose to do so, how customers respond or whether workers move into new work. The ILO’s point is deliberately less dramatic than a headline: technological susceptibility is not the same thing as a labour-market outcome.

The ILO–NASK global index, published on May 20, 2025, estimated that 25% of global employment sits in occupations with some potential exposure to generative AI, rising to 34% in high-income countries. Clerical work is the most exposed broad category, while newer measures also identify substantial exposure in some cognitive, analytical, administrative, managerial, business, finance, computing, mathematics and education roles. These are task-level estimates across more than 140 countries, not counts of jobs already lost.

A useful working distinction for this article is three-layered, although it is not an official ILO or OECD taxonomy. An assistant supports a human task, such as drafting a response. An agentic workflow is designed to execute a sequence with less continuous human intervention, perhaps retrieving information, making a recommendation and routing the result. A whole-job replacement claim is larger still: it requires evidence about the occupation’s full task bundle, organisational adoption and actual employment effects. Moving from “the model can do this task” to “the role has disappeared” skips the difficult middle.

Two forecasts, and the evidence between them

In his October 2024 essay Machines of Loving Grace, Dario Amodei describes a hopeful AI future but also a difficult economic transition. He expects human work to remain useful in the nearer term; in the longer run, he argues that sufficiently capable, cheap AI could require a different economic arrangement. It is a scenario about future capabilities, not a count of jobs already lost.

Sam Altman’s September 23, 2024 essay takes a more optimistic line. He predicted significant labour-market change, good and bad, but argued that most jobs would change more slowly than many people expect and that society would not run out of useful things to do. That outlook depends on assumptions about new demand, falling costs, entrepreneurship and the creation of work that is not yet visible. It is a strategic forecast, not a labour-market estimate.

Neither a hopeful essay nor a darker scenario establishes a universal percentage of jobs already lost to AI. Forecasts about AGI or superintelligence are not observed labour-market categories. The nearer question is more concrete: which tasks are being removed, who approves the output, and what happens to the people who used to perform the first version of the work?

There are early signals, but not a clean verdict

The best-known productivity evidence is not a mass-layoff study. The April 2023 NBER working-paper version, involving 5,179 customer-support agents found that an AI assistant increased issues resolved per hour by 14% on average, with a 34% improvement for novice and low-skilled workers. That supports augmentation in a particular workflow. It does not show that employers hired more people, shared the gains through pay or retained the same headcount across the wider economy.

Stanford’s August 12, 2026 update found no widespread economy-wide displacement in its ADP sample. But employment among US workers aged 22–25 in highly AI-exposed occupations was about 19% below the path implied by keeping pace with less-exposed peers, as of June 2026. The adjustment appeared mainly in reduced hiring. These are descriptive patterns, not causal estimates; the authors caution that the sample does not represent the whole US labour market.

An earlier Stanford analysis also examined interest rates and timing rather than simply attributing every change to AI. The broader lesson is methodological: restructuring, demand, previous hiring decisions and changing business strategy can matter too. A pattern worth investigating is not a licence to announce that causation is settled.

The junior ladder is part of the deal

A job can survive while its career ladder thins. If software produces the first draft, summarises the meeting and prepares the basic analysis, a junior employee may receive more responsibility sooner—or lose the supervised repetition that builds judgement. Anthropic’s December 2, 2025 internal study of 132 engineers and researchers, including 53 interviews, reported self-described productivity and output gains alongside concerns about technical depth, supervising AI outputs, reduced collaboration and mentorship. Its sample came from an unusually AI-exposed company, so it is a warning about learning rather than proof of a universal effect.

This is why “learn AI” is an incomplete employment policy. Training cannot create demand, paid time, progression, protection from bad automated decisions or a first job in which new skills can be practised. OpenAI’s September 2025 proposal is notable because it pairs AI access with paid training time, workflow-linked credentials, job matching, portability, human control of pivotal decisions and cooperation with organised labour. Those are proposals from a model developer, not settled obligations or proof that the measures work at scale.

Who gets the productivity dividend?

The sources reviewed here can show that a workflow becomes faster. They do not settle whether the value appears as higher pay, shorter hours, lower prices, more hiring, larger margins or investor returns. That distribution is a bargaining question. A company that introduces an assistant may improve service and reduce drudgery; it may also raise output targets, reduce entry-level recruitment or make one employee responsible for checking ten machine-generated decisions. The technology does not decide that allocation on its own.

A credible employment deal therefore needs more than a software licence. It should make clear which decisions remain human, who is accountable when an AI-assisted decision fails, how training time is paid, whether workers can challenge or audit deployment, how skills travel between employers and how junior staff will gain supervised experience. Governments and employers also need better evidence separating completed AI-driven task changes from announced plans, and AI effects from ordinary cost-cutting or restructuring.

For a founder or executive, the practical test is simple: when a task disappears, what replaces the learning pathway, the accountability and the income? For a worker, the corresponding question is whether “efficiency” arrives with paid development, a voice in implementation and a route to better work. Until those answers are negotiated, AI may be changing the job while leaving the deal—who bears risk and who captures gain—untouched.