The complement, not the cut

The dominant story about AI and work is a story of subtraction. The most prominent voices in AI have warned of sweeping white-collar job losses within a few years — some suggesting half of such roles could go, with entry-level jobs first in line — and plenty of investors have put it more bluntly still: sell anything that consists of people sitting at a desk looking at a screen.^[1] The logic underneath is simple: AI can do some of what your people do, so you need fewer of them. Cut, and the savings fall to the bottom line.

I’ll be straight about where I come at this from. I’m a founder, and founders are builders by disposition — handed a powerful new capability, the instinct is to do more with it, not to do the same with less. So discount my view accordingly. But this isn’t only temperament. The economics back the instinct, and that’s what the rest of this is about: the businesses cutting are, I think, making a bigger mistake than the one they believe they’re avoiding.

I’ve spent my working life building and transforming companies, and the hardest, highest-leverage part of that work is never the technology. It’s finding, hiring and keeping people who are genuinely excellent — and I’ve had the rare luck of working alongside some of the best operators I expect to meet. If you’ve ever done that, the instinct to greet a powerful new tool by getting rid of those people should strike you as strange. You don’t finally get your hands on your scarcest resource, brilliant people, and decide to buy less of it.

The premise underneath the cuts

The case for cutting rests on a premise worth stating plainly: that AI is now cheaper than a person for the same work — so swapping one in for the other drops straight to the bottom line. That premise has two parts, and they deserve to be pulled apart. The first is whether AI is really getting cheaper. The second is whether “the same work” is the right comparison at all. Take the cost first; the rest of this piece is about the comparison.

The price of a single token isn’t the right number, and neither is the sticker price of the latest frontier model — both can rise. The number that matters is the cost of a given amount of capability: what it costs to get a fixed quality of work done. That has collapsed. By the Stanford AI Index’s measure, the cost of running a model at the level of 2022’s best fell more than 280-fold in about eighteen months.^[2]

There’s an honest wrinkle. The newest reasoning models can cost more per query than their predecessors, because they do far more work to answer, and total spending on AI is climbing, not falling. But that isn’t the unit cost of capability going up. It’s demand rising to meet a unit cost that is falling fast: when a useful thing gets cheaper, people don’t buy less of it, they buy dramatically more. Hold onto that, because it’s the whole argument in miniature.

Substitutes and complements

The job-loss thesis treats AI and human labour as substitutes: two ways of doing the same thing, where more of one means less of the other. Some tools really are like that. But the more important relationship is the other one — the complement. A complement is a tool that makes the scarce thing more productive rather than replacing it. And here is the part the layoffs miss: when you make a complement cheaper, demand for the scarce input it complements tends to go up, not down.

The cleanest illustration is nearly fifty years old. When the cash machine arrived, the obvious prediction was that it would gut the ranks of bank tellers — it was, after all, an automated teller. The opposite happened. The number of tellers a branch needed fell, from around twenty to about thirteen. But that made a branch cheaper to run, so banks opened far more of them — branch numbers rose by something like forty per cent — and total teller employment went up, not down.^[3] The work changed: freed from counting cash, tellers moved to the things a machine couldn’t do, selling and advising and handling the awkward cases. Cheaper routine work didn’t destroy the job. It raised the value of the part that only a person could do.

The fork — and the honest half of the story

But that example has a second half. The cash machine was a complement. Mobile banking was not. Once a phone could do the whole interaction, teller numbers genuinely fell — because the phone didn’t automate a task within the job, it replaced the job. That is the real fork, and it’s worth being precise about. When a tool takes over some tasks within a role, it tends to make the person doing it more valuable, and you want more of them, pointed at harder work. When a tool takes over the entire role, it replaces them.

The broadest recent synthesis of firm-level studies points in the same direction: so far, AI adoption has been associated with firm growth and rising employment, with job-displacing effects concentrated in particular sectors and tasks.^[4] You can see both halves of the fork in it — audit firms that adopted AI trimmed their audit headcount, while AI assistants in customer support raised output and the case for more of it. So the honest position is not that AI never costs a job. Some work is pure substitution, and for that work, headcount will fall — and anyone arguing otherwise is selling something.

The question for any specific role is which kind you’re looking at. And for the work that actually drives a business — judgment, taste, relationships, the willingness to own an outcome — AI is overwhelmingly a complement. It takes over the routine, repeatable parts and lets the genuinely scarce thing operate at far higher leverage. Make that complement cheaper and you have made your best people more valuable, not less.

And it’s open to everyone

There’s a part of this that the headcount framing misses entirely. These tools aren’t gated behind a specialism, the way most powerful technologies have been. You don’t need to be an engineer, an analyst, or a designer to do work that used to require one. It broadens what a motivated person in almost any role can attempt — and the evidence shows it helps the least experienced the most. In one field study of more than five thousand customer-support agents, an AI assistant raised productivity by about 14% on average — but the gains were concentrated among novices, who improved by around a third, while the most experienced barely moved.^[5] The tool pulls people up toward the standard of the best, rather than the reverse.

Which quietly changes what separates people. When access to capability is cheap and general, the dividing line stops being about who holds the credential and starts being about who is willing to pick up the tools and use them well — curiosity, judgment, the appetite to learn. That is a matter of attitude far more than pedigree, and it is the opposite of a reason to thin the ranks.

What the person is actually for

There’s a deeper reason the substitution sum doesn’t add up, and it took me a while in business to see it clearly. AI is extraordinary at the portable part of a job — the part that would look the same in any company. What it cannot be is up to speed on your organisation. One thing I learned well as a consultant: however expert you are in your field, it takes something like three months to get genuinely up to speed on any given organisation — how it actually works, who to call, what was tried before and why it failed. That knowledge is slow to build and it doesn’t transfer.

Which makes the arithmetic of replacement worse than it looks on the slide. The people you’d replace are already loaded with that context. Let one go, and you haven’t just lost someone who does a task — you’ve reset the context clock to zero and have to spend months rebuilding it in whoever comes next. Whereas the person already up to speed who needs, say, two months to learn a new skill leaves you net ahead. The portable task is the cheap, fast thing; the organisational context it sits inside is the expensive, slow thing. AI now hands you the cheap part for almost nothing. The error is to discard the expensive part to save on the cheap one.

A cut in anticipation — and the easy answer

Here is what should give the subtraction story pause. A good deal of the current cutting is not a response to AI actually doing the work. It is a response to the prediction that it will. In a survey of more than a thousand executives at the turn of the year, only about two per cent reported large headcount reductions tied to AI they had actually implemented; far more were cutting, or slowing hiring, because of what they expected AI to do later.^[6] And across the whole economy, AI was cited in only around 55,000 of last year’s announced job cuts — a small share of the total — with analysts cautioning that even that figure may be overstated, a tidier story to tell investors than weak demand or overhiring.^[7]

Part of the appeal, I suspect, is simply that the cut is the easy answer. “Replace role X with tool Y” is a clean line on a slide and a tidy story to tell investors — tidier, often, than weak demand or overhiring or a strategy that didn’t land. The harder thinking — working out what your people are now freed to do, reshaping roles around it, redeploying talent toward growth — is slower and less legible, and it doesn’t announce itself as decisive action. So the lazy answer and the flattering answer happen to be the same answer. That is exactly the condition under which businesses get a big decision wrong.

Cost-cutting was never the strategy

Say AI does let you do this year’s work with fewer people. That is a saving. It is not an advantage. If the only thing you have bought is “we did the same work with a smaller team,” every competitor buys the identical tool next quarter and matches you. You have cut to parity, not to an edge. Real advantage comes from using the same tool to do what rivals can’t — to serve customers better, move faster, take on the harder problem — and that takes the people, armed and aimed at something worth doing, not removed from the building. Businesses don’t win by being cheaper to run. They win by being the best at something that matters to whoever is paying them.

Where the constraint actually sits

Strip the layoff instinct back, and it’s a claim about where a business’s binding constraint sits. It says: the thing holding us back is the cost of our people, and AI has just relaxed it. For a handful of businesses, that’s true. For most, it’s a misreading. The constraint was never the cost of good people. It was finding enough genuinely excellent ones, and giving them the room and the tools to do their best work. AI doesn’t loosen that constraint by letting you fire. It loosens it by letting the people you already fought to hire do more than they could before.

Read the constraint as cost, and you cut. Read it correctly, and you do the opposite: you compete harder for the best people, and you arm them better than anyone else does.

I’ll declare my own bias here, because it’s the honest place to. This is the part of the shift that makes me most optimistic about what comes next. Not because I think a leaner team does more with less — but because when I come to build a team out, every excellent person I bring in will carry more leverage than they could have a few years ago. Building has rarely been more possible than it is now, and that makes the prospect more exciting, not less.

Three questions before you treat AI as a reason to shrink

1. For this particular role, is AI automating tasks within the job, or the whole job? The first makes the person more valuable; only the second replaces them — and most roles are the first.
2. If we cut here, what walks out of the door with the person — and what advantage have we actually bought? Something a competitor can’t copy by buying the same tool, or just a lower cost they’ll have matched by next quarter?
3. And the one I keep coming back to: is the thing really holding this business back the cost of our people, or our ability to find great ones and let them do their best work? Because if it’s the second, AI is the strongest argument for hiring I’ve seen — not against it.

The expensive mistakes I keep noticing share a shape. They misread where the structure of a thing really sits — who holds the power in a decision, where a cost is going to land, which input is genuinely scarce — and then dress the misreading up as a saving. This is that mistake, pointed squarely at the people who were always going to be the reason a business won.