What Happens When AI Can Do Most of the Work?

There's a big difference between AI being able to perform a task and AI being able to replace the person who used to do it. Recruitment is a good example. AI can draft job adverts, summarise applications, suggest interview questions and keep candidate information organised, which together make up a substantial share of a recruiter's week. Someone still has to judge whether a candidate is genuinely suitable, challenge a shortlist that looks wrong, handle difficult conversations and take responsibility for the final decision. The job may have changed considerably, but it hasn't gone away.

The research reflects this distinction. In 2025 the International Labour Organization estimated that around one in four workers worldwide is in an occupation with some exposure to generative AI.^[1] That figure is often quoted as though a quarter of jobs are about to disappear, but the ILO was clear that exposure measures the potential for tasks within a job to be affected. It isn't a forecast of job losses, and the researchers expected jobs to be transformed far more often than they are replaced.

This leads to a point I think is badly underappreciated. A company doesn't need to cut a single job to gain an economic advantage from AI. If the same staff can serve more customers, finish projects sooner or spend less time on administration, the business becomes more productive with exactly the same headcount. So the more revealing question is what happens to all that extra capacity, and who ends up with the value it creates.

You might expect the most experienced employees to get the biggest boost from AI, since they know their jobs best. One of the most widely cited workplace studies found the opposite. Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed more than 5,000 customer-support agents after their employer introduced an AI assistant. Productivity, measured as customer issues resolved per hour, rose by about 14% on average. For newer and less-skilled agents the improvement was around 34%, while the most experienced and highly skilled agents saw very little gain.^[2]

The researchers' explanation is the most interesting part. The assistant appeared to spread the know-how of the company's best performers to colleagues who hadn't yet picked it up for themselves. Knowledge that used to take years on the job to acquire became available to someone in their first few months.

It's one study of one tool in one customer-support operation, so it doesn't mean every business will see similar results. It also didn't show that the agents were paid any more for their higher output. What the research, published in The Quarterly Journal of Economics as Generative AI at Work, does show is that AI can raise productivity and even narrow the gap between stronger and weaker performers. Who receives the financial benefit of that improvement is decided somewhere else entirely.

A second large study shows why individual time savings don't automatically add up to anything bigger. Researchers gave a randomly selected group among 7,137 knowledge workers in 66 organisations access to a generative AI tool built into the email, meeting and document software they already used. Workers who used the tool spent about two fewer hours a week on email, which is a meaningful saving given how much office life revolves around messages.^[3]

Yet the researchers found no shift in the quantity or mix of tasks people were doing as a result of being given the tool individually. People saved time on particular activities, but the overall shape of their jobs stayed much the same. The study, Shifting Work Patterns with Generative AI, doesn't prove that wider change is impossible. What it suggests is that handing employees a tool is very different from redesigning how an organisation works. Without that second step, saved hours tend to be quietly absorbed by more email, more meetings and more tasks, and people can end up feeling as busy as ever despite having technology that makes each individual activity easier.

Suppose an AI system completes in five minutes a task that used to take someone two hours. On paper that's an enormous saving, but the calculation changes quickly once you include everything around it. If someone spends 45 minutes checking the output, the saving shrinks considerably. If the system occasionally invents information, or produces a recommendation that turns out to be wrong, putting it right can cost more than the time saved in the first place.

Businesses aren't simply swapping wages for a software subscription. Depending on the system, they may also be paying for integration with existing applications, data preparation, security, training, monitoring and human oversight. The cost of an error varies enormously too. An inaccurate meeting summary is an inconvenience, whereas a flawed employment decision or financial recommendation can have serious consequences for the people affected and for the organisation.

There's a subtler effect as well. When something becomes cheaper to produce, organisations tend to produce more of it. A marketing team that used to develop three campaign ideas might now generate thirty, but someone still has to read and judge all thirty. The genuine saving is the difference between the total cost of the old process and the total cost of the new one, including the checking, the rework and the extra volume. Speed is easy to measure, but value is the figure that matters.

Imagine a company genuinely saves £100,000 a year through AI. (That's an illustrative figure, not a research finding.) The money could increase profits, fund pay rises or new jobs, allow lower prices or be reinvested in better systems. There's no universal rule, and a great deal depends on the market the company operates in.

Picture two businesses using exactly the same technology. The first works in a fiercely competitive market where customers can switch easily, so much of its saving is likely to be passed on through lower prices or better service, because if it isn't, a competitor will do it instead. The second faces little competition and has customers who find it hard to leave, so it can keep far more of the gain for itself. The tool is identical, but the benefits end up in very different places.

Then there's the company supplying the AI. If businesses come to depend on a small number of technology providers, a share of the value created by automation will flow to those providers through subscriptions, usage charges and infrastructure fees. The organisation adopting AI doesn't necessarily capture all the gains, and its employees may capture very little of them.

If you become twice as productive, should you earn twice as much? It sounds fair, but that isn't how pay is usually set. Salaries reflect skills, labour-market conditions, employer policies, competition and bargaining power. Productivity matters, but it isn't a salary calculator.

We don't need to wait for AI to see this in action. The OECD Employment Outlook 2025 found that wage growth across OECD economies has been weaker than productivity growth, especially among lower-paid workers.^[4] In other words, economies have become more productive without workers receiving equivalent increases in pay. That's a finding about past trends rather than proof that AI will widen the gap, but it should make us wary of assuming that the gains AI promises will show up in ordinary pay packets.

The same logic applies within a single team. If a team that produced ten reports a week can now produce twenty, its employer might reward that with higher pay, invest in training or use the extra capacity to grow. Or twenty reports might simply become the new expectation. And if AI makes certain skills easier to reproduce, the market value of those skills may fall. People who combine deep specialist knowledge with AI may become more valuable, while those whose work is easiest to replicate may face more competition. This is why headline estimates of how much AI will add to the economy tell us very little about how much will reach the people doing the work.