There is a specific morning, a few weeks after a system goes live, when a team realises it has its week back.
The work that used to eat three days now runs unattended overnight. At a large London accounting firm, a billing automation I built saves them somewhere around 800 hours of staff time every month. Those hours were real. People used to spend them keying data, chasing approvals, and reconciling the same figures twice.
So here is the question that actually matters. Not "can we automate this", which is nearly always yes. The question is what those people do on Monday, because that single decision is what separates a company that used AI to get cheaper from one that used it to get bigger.
Most firms get it backwards.
The reflex, when you free up a skilled team's week, is to read it as a line on a budget. Fewer hours needed means fewer people needed. Cut them, book the saving, tell the board you are running lean. It is the lazy option, and it is usually the wrong one.
A layoff banks a one-off number and lowers your ceiling for good. The capability you just cut, the people who understood a process well enough to automate it, is the exact capability that could have built the next thing you sell. You do not grow revenue by subtraction.
I want to be fair about this. Layoffs are sometimes the right call, when a business is genuinely over-built or the market has moved under it. That is not what most of these cuts are. Most are companies trimming from a position of new strength, mistaking a windfall of time for a reason to have fewer hands.
Two companies show both roads.
Klarna took the first one. In 2024 the payments firm shrank from roughly 5,500 staff to about 3,400, and said in public that an AI assistant was doing the work of 700 agents. Efficiency, cashed out as headcount. By 2025 the story had turned: the chief executive admitted the company had gone too far, that leaning on cost-cutting had dented the quality customers were actually paying for, and Klarna started hiring people back under a promise that there would always be a human available for anyone who wanted one. The lesson is not that the AI failed. It worked. They spent the savings in the one place guaranteed to erode the product, and then had to buy it back.
Moderna took the other road. The drugmaker rolled ChatGPT Enterprise out across its entire workforce, and its own staff built more than 750 internal AI assistants tuned to their specific jobs. The framing, from the top, was augmentation. Give every person a tool, not a P45. Where that points is the interesting part. Moderna has said it wants to launch around 15 new products over five years with roughly the 6,000 people it already employs, work that on the old operating model would have needed a headcount many times that size. Treat the figure as stated ambition rather than booked fact, because that is what it is. The direction, though, is unmistakable. The freed capacity is aimed at new output, not at the door.
Same technology as Klarna. Opposite instinct. Opposite outcome.
This is not just two anecdotes pointed in convenient directions. A field study of around 5,000 customer support agents, run by the economist Erik Brynjolfsson and colleagues, found that generative AI lifted their productivity by about 14 per cent, and that the largest gains went to the least experienced staff. That is AI working as a leveller, making the people you already have measurably better, rather than a reason to replace them.
Even Jensen Huang, who has more reason than almost anyone alive to talk AI up, keeps making the same argument in blunter terms. You will not lose your job to an AI, he said at the Milken conference, but you might lose it to someone who uses one. Blaming AI for layoffs, he added more recently, is lazy, and ambitious companies will end up hiring more people, not fewer.
None of this happens on its own. Human in the loop is a decision you make before you build, not a rail you bolt on after something has already gone wrong.
In practice it comes down to a few habits. Automate the task, but keep the judgement with a person, because judgement was rarely the bottleneck in the first place. Measure the hours you free up as capacity to redeploy, and put that number in front of the people who decide where it goes. And decide where it goes before you flip the switch. Time with no plan on it does not get reinvested. It quietly evaporates into slightly shorter days, and then someone senior looks at the same output and the same payroll and reaches for the obvious, lazy lever.
AI does not save you money. It buys you time. What you build with that time is the whole game.