The first question in the room is almost always the wrong one: what can we cut?
It's a natural instinct. A powerful new tool arrives that can do work people used to do, and the mind goes straight to subtraction — fewer hands, lower cost, a leaner line on the budget. It feels like prudence. It's actually a ceiling.
Here's the arithmetic nobody runs out loud. When you aim a powerful new capability at cutting cost, you are optimizing for the cheapest output your customer will still tolerate. That is the target you've set, so that is the target you'll hit. And it works — for a while. The savings arrive first, clean and easy to celebrate. The erosion arrives later, quietly, in the places that don't show up on a dashboard until a customer has already decided to leave.
The experiment that ran in public
Klarna ran this experiment where everyone could watch. In early 2024 the company announced its AI assistant had handled 2.3 million conversations in its first month — two-thirds of its customer-service chats — doing work it equated to 700 full-time people. It was the headline every operator wanted to be able to write: the technology arrives, the cost base collapses, the market applauds.
A year later the same CEO walked it back in plain terms. Cost, he said, had become the predominant factor in the decision, and what you end up with that way is lower quality. The company started rehiring. The detail worth sitting with is that the technology had not failed. The assistant could do what it was built to do. What had failed was the aim. They pointed the most capable customer-facing tool they had ever deployed at the single question of how much labor it could remove, hit that target exactly, and discovered that the cheapest tolerable service is a brittle thing to have optimized for — because "tolerable" is measured by customers, and customers stop tolerating without filing a notice first.
That is not an argument that Klarna was foolish. They were early, they were public, and they corrected. It is an argument that the instinct they followed first — the subtraction instinct — is the wrong one to follow first, and that the correction they made is the lesson the rest of us get to learn without paying for it in our own customer base.
A coupon and a printing press
This is the difference between a coupon and a printing press.
A coupon saves you a little money, once. You clip it, you spend it, and the transaction is over — the best possible outcome is that you paid slightly less for the same thing you were already buying. The value is real, and it is bounded, and it is finished the moment you redeem it. Nobody ever built an enterprise on coupons.
A printing press doesn't save you anything. It lets you make something you could never make before — books, by the thousand, faster than every scribe in Europe combined — again and again, at a scale no amount of hand-copying could ever reach. It didn't make the existing business of copying manuscripts cheaper. It made a new world possible: cheap books, mass literacy, newspapers, the spread of ideas at a speed the old method could not have produced at any price. One is subtraction. The other is capability. And the distinction is not poetic — it is arithmetic.
Subtraction has a floor. You can only cut to zero. Every dollar of cost you remove is a dollar you can never remove again, and when you have removed the last one there is nothing underneath it. The savings are banked once and gone. Capability has no ceiling at all. There is no upper limit on what becomes possible, because each new thing you can do becomes the ground the next new thing stands on.
Why one compounds and the other doesn't
That last point is the one that decides which kind of company comes out of this era stronger, so it's worth being precise about it.
A cost cut is a one-time event with a one-time result. You reduce a line on the budget, the reduction shows up once, and the following year that line is simply lower — it does not keep falling on its own, and you cannot cut it again without cutting into something you needed. The gain is a step down to a new floor, and floors don't compound.
Capability is different in kind, not degree. When AI lets you serve a customer you previously couldn't afford to serve, that customer is not a one-time saving — it's a new relationship that can grow, refer, renew. When it lets you answer a question that was never worth the labor of answering, you haven't trimmed a cost; you've added something the business could not previously offer. When it makes a market viable that never justified the cost of entry, you haven't subtracted — you've opened a door, and behind that door is another door you couldn't see until you were through the first one. Each of those gains becomes the platform for the next. That is what compounding means, and it is the reason the two strategies diverge so violently over time even when they look comparable in the first quarter.
Run the two side by side over three years and the shape is unmistakable. The cost-cutter banks a savings in year one and spends years two and three defending it as quality quietly slips at the edges. The capability-builder posts a smaller, less impressive number in year one and then watches it grow, because the thing they built in year one made the thing in year two possible. The headline belonged to the cost-cutter. The compounding belonged to the other one.
The quiet bill the dashboard doesn't show
There is also a cost to aiming at subtraction that never appears in the savings calculation, and Klarna named it: lower quality. When the target is the cheapest output the customer will tolerate, quality is not a goal — it is a constraint you are pressing against from above, lowering until something pushes back. The trouble is that the thing which pushes back is the customer, and the customer pushes back by leaving, not by complaining. By the time the erosion is visible in the numbers, the decision it caused has already been made in the customer's head, weeks or months earlier, in a service interaction nobody flagged.
This is why the savings look so clean and the damage looks like nothing for so long. The two effects are on different clocks. The savings land this quarter, in a line you can point to. The erosion lands later, diffuse, in renewals that don't happen and referrals that never get made — and it is almost impossible to attribute, which means it is almost impossible to defend against once it has started. Aiming at cost doesn't just cap your upside. It quietly mortgages a downside you won't see until it has already been collected.
It was never an argument against efficiency
None of this is an argument against efficiency. Efficiency is good. Waste is bad. If AI lets you do something you already do with less friction and lower cost, take it — that is real, and it is worth having.
It's an argument about where you aim, and about which question you ask first, because the first question sets the ceiling for everything that follows. Point the most capable tool you have ever had at making things cheaper, and the very best you can do is a cheaper version of what you already were. The ceiling is the company you are today, minus some cost. Point it at making things possible, and you change what the company is. There is no ceiling, because you are no longer optimizing the thing you already are — you are building the thing you couldn't be before.
So before the next conversation turns to what you can cut, ask the harder question first, and ask it out loud, because it's the one that sets the ceiling for everyone who speaks after you.
What would you finally attempt if this made the impossible affordable?