The friction was the business

There is a number that quietly governs most of the prices you pay, and it is not in any catalogue. It is the price of bothering. The price of reading the renewal letter, of phoning to cancel, of comparing the four tariffs, of moving the money, of reading clause 14(b). Standard economics, in its confident textbook form, sets this number to zero. It imagines a buyer with complete information and infinite patience, who will desert any seller charging a penny too much. Markets, on this account, are disciplined by the constant threat of departure.

The buyer in question has never existed. Real people are busy, tired, loss-averse, and attached to whatever they chose last time. They leave money on the table by the hundreds of billions, knowingly, because retrieving it is tedious. And a great deal of what we call corporate profit is simply the harvest of that tedium, the rent collected from the gap between the deal a customer would take if checking were free and the deal they actually accept because it is not. Whole industries are built on the reliable fact that the customer will not bother.

That fact is about to change, not because people will become more diligent, but because diligence is being decanted into software that feels none of the friction a human feels. The interesting consequence is not that a few intermediaries lose. It is that a foundational assumption of market economics, false since Adam Smith, may be about to become true for the first time, and the redistribution that follows runs from corporate margins to consumer surplus on a scale that the aggregate profit statistics are not yet braced for.

I · The behavioural foundationMarkets do not run on preferences. They run on whatever people can be bothered to act on.

Begin where the modern understanding begins. In 1961 George Stigler pointed out that prices for identical goods differ, persistently, and asked why anyone ever pays the higher one. His answer founded a field: searching for the better price has a cost, so people stop searching before they have found the best deal. Price dispersion, he wrote, is the measure of ignorance in the market, and ignorance is rational when knowledge is expensive.¹ The seller charging more is not defrauding anyone. The seller is being paid for the customer's reasonable decision to stop looking.

Behavioural economics then made the picture far worse for the textbook. It is not merely that search costs money. People systematically prefer to do nothing. In 1988 Samuelson and Zeckhauser ran a battery of experiments and named the effect status quo bias: faced with options, subjects disproportionately stuck with whatever was already in place, even when a switch was plainly better and the cost of switching was trivial.² The bias is held together by loss aversion (the downside of a bad switch looms larger than the upside of a good one), by regret avoidance, by the sense that the incumbent choice has already been vetted. Their field evidence was damning: in university health-plan data, new employees flocked to a plan with better premiums and deductibles while long-standing employees, free to switch at any time, simply did not.

The most vivid demonstration of the same force is one of policy's favourite findings. Johnson and Goldstein showed that whether a country makes organ donation opt-in or opt-out, a difference of a single default setting, swings registered-donor rates enormously: opt-out countries average roughly six times the registration of opt-in ones.³ Nobody's underlying willingness to donate changed. What changed was which outcome required the effort of acting. The default did the work, because the default always does the work. People accept what is placed in front of them.

Whatever requires an action will be under-chosen, and whatever requires no action will be over-chosen, regardless of which one people would prefer on reflection.— the operating principle of every default

Read those three findings together and a general law emerges. The outcome a market produces is not the outcome people prefer; it is the outcome that needs the least action. And since charging a passive customer more requires no action from that customer, the natural drift of almost every ongoing commercial relationship is upward, away from the deal the customer would have chosen and toward the deal they will tolerate. The economists call the harvested difference a rent. The behavioural scientists can tell you exactly which biases grow it.

II · The size of the harvestInertia is not a rounding error. It is a line item, measured in the billions.

Because regulators occasionally measure it, the rent of inattention is one of the few behavioural quantities we can actually price. In 2018 the UK consumer body Citizens Advice filed a super-complaint alleging that loyal customers were being systematically overcharged. The Competition and Markets Authority investigated and confirmed it: across just five markets (mobile, broadband, home insurance, savings, and mortgages) British consumers were losing about £4.1 billion a year to what became known as the loyalty penalty, with roughly 8 in 10 people overpaying on at least one service simply by staying put.⁴ The behavioural mechanism was explicit in the regulator's own reasoning: stealthy annual price rises, onerous cancellation, and the calculated exploitation of inertia.⁵

The pattern recurs wherever someone has looked. In US health insurance, one study found that eliminating customer inertia would have cut premiums by 13.2%, yet 87% of people renewed with the same provider anyway.⁶ The asymmetry is the whole game: firms make signing up take seconds and make leaving take an afternoon of hold music, an architecture the behavioural literature now calls sludge. The endowment effect makes people overvalue the plan they already hold; status quo bias makes change feel like risk; present bias makes the afternoon of effort loom larger than a year of savings. Each bias is individually small. Aggregated across every recurring bill a household pays, they sustain an enormous and entirely legal transfer.

And this is only the part that gets measured because a regulator forced a price comparison into the open. The unmeasured rent is larger and stranger. Consider the most placid pool of capital in the economy: bank deposits. Bank of America held average deposits of roughly $1.99 trillion in the third quarter of 2025, earning a net interest yield around 2.07% on its assets.⁷ A great deal of that base sits in accounts paying close to nothing while comparable money-market funds pay several points more. Not one dollar of it is being optimised, not because the depositors are irrational, but because moving it sits forever just below the threshold of what is worth doing today. The Federal Reserve's own research shows banks price deposits with exactly this in mind, offering better rates where customers are mobile and worse rates where they are stuck.⁸ The interest you are not paid is the measure of how trapped the bank correctly judges you to be.

III · What the agent actually changesIt does not make people diligent. It removes the cost of diligence.

Every rent described above rests on the same load-bearing assumption: that acting is costly enough to deter the customer. Knock that assumption out and the whole structure has nothing to stand on. This is precisely what a capable software agent does. It supplies vigilance at a marginal cost approaching zero. It does not get bored reading the forty-page policy. It feels no loss aversion, holds no endowment over the incumbent brand, suffers no present bias that makes today's small effort outweigh next year's saving. It will, uncomplaining, run the twelve-way comparison you would never run for a utility tariff, and run it again next month.

In behavioural terms the agent is a near-perfect debiasing instrument, because it simply does not host the biases. It also collapses Stigler's search cost toward its theoretical floor. The market the textbook always assumed (full information, costless comparison, a credible threat to leave) does not arrive because humans improve. It arrives because the searching is delegated to something that does it for nothing.

This is not a forecast about a distant decade. By early 2026, between 30% and 45% of US consumers were already using generative AI for product research and comparison.⁹ Three-quarters of retailers told Salesforce they expected AI agents to be essential to their business within the year.¹⁰ OpenAI shipped an Instant Checkout protocol in September 2025; Amazon built a tool to buy from rival sites without leaving its app, read widely as a defence of its roughly $56 billion advertising business, which exists only because people currently browse rather than delegate.¹¹ Even the search-advertising fortune at the centre of the web is, on inspection, a pure inattention rent: advertisers pay to sit at the top of the results because being found is worth more to them than the comparison is worth to a human who will not scroll. An agent scrolls every result, on every engine, for free.

The consulting estimates of the damage to incumbents have grown specific. Kearney put the figure at up to 500 basis points of EBIT erosion for retailers unprepared for agent-mediated commerce, from margin compression, traffic dispersion, and commoditisation together.¹² Bain warns that the businesses most exposed are the multi-brand intermediaries whose entire value was aggregating choices a customer could not be bothered to aggregate themselves.¹³ When comparison is free, the premium for not comparing has nowhere left to sit.

IV · The better-working marketReframe the loss. What looks like margin compression is the market finally doing its job.

It is tempting to read all this as damage, a wave of disintermediation crashing through one sector after another. That framing is true but shallow. Step up a level and something more fundamental is visible. A market in which prices are disciplined by inertia rather than by value is a broken market, one that has been quietly misallocating capital for as long as anyone has measured it. The loyal customer subsidising the switcher, the idle trillion earning nothing, the renewal that creeps up every year: these are not signs of a market working. They are the standing evidence that it is not.

Agentic vigilance fixes the defect at its root. When switching becomes effortless, price has to track value again, because there is no longer an inertia premium to hide behind. Capital that sat idle gets moved to where it earns. The low-cost producer, long protected against by customer laziness, finally gets the share its efficiency deserves. Three percent points of misallocated rent, redistributed, is not a calamity for the economy. It is an efficiency gain, of exactly the kind economists have spent a century saying competition is supposed to deliver and quietly knowing it mostly did not.

A price held up by inertia is a price that has stopped carrying information. Removing the inertia does not break the market. It switches the signal back on.— the case for the upgrade

The redistribution is the point, and it should be named honestly. Rent that used to appear as profit on corporate income statements reappears as surplus in consumer pockets: better rates, lower premiums, fewer creeping renewals, idle cash put to work. For a household this is unambiguously good. For the index it is a transfer away from capital. The two have been conflated for years because rising margins were read as rising prosperity. They were partly rising extraction, and the thing now coming for the extraction is, from the customer's chair, a long-overdue correction.

V · The arithmetic that should worry shareholdersA bigger economy with thinner margins can leave aggregate profit flat, or lower.

Here is the uncomfortable interaction, and it deserves to be stated precisely because it is so easily missed. US corporate margins are not merely healthy; they are at a multi-decade peak. The S&P 500 net profit margin reached its highest level in more than fifteen years in late 2025, with analysts forecasting around 14.2% for 2026, against a five-year average nearer 11.5%.¹⁴ Over the longer arc the climb is steeper still: De Loecker, Eeckhout and Unger estimate that the average US markup rose from about 18% over marginal cost in 1980 to roughly 67% in recent years, with the great majority of the increase happening within industries rather than through any change in the industry mix.¹⁵ A meaningful slice of that ascent is the compounding of inattention rent across a digitising economy.

Now run the numbers the way an investor should. Suppose AI lifts the level of GDP by something like 5% to 10% over a decade, the middle of most serious estimates. Suppose, too, that agentic vigilance compresses aggregate corporate margins by 100 to 200 basis points from a base around 12%. That is a haircut of one-twelfth to one-sixth on profitability. It is entirely possible for total corporate profit to be flat or lower at the end of the decade even as output is materially higher and consumers are unambiguously better off, because the share of a larger pie accruing to firms has shrunk by more than the pie has grown. A bigger economy can be a worse one to own and a better one to live in, at the same time.

There is a twist that makes the bear case partly self-financing, in a way good for nearly everyone except the incumbent. If agents genuinely hand consumers a better deal, they become more popular, adoption accelerates, and the technology diffuses faster. The very mechanism that thins corporate margins is the mechanism that makes the tools irresistible to the people holding them. Lower rents are not a drag on the AI transition. They are its consumer subsidy, paid out of margins that were partly extraction to begin with.

VI · Three reasons it might not happenThe honest objections, each of which has teeth.

A thesis worth holding has to survive its best counterarguments. There are three, and a quarter-century of evidence supports each.

One: transparency has been tried, and margins rose anyway.

The internet was supposed to do all of this. In 2000, a Harvard Business Review essay warned that the web's flood of price information was the gravest threat yet to a firm's ability to hold a margin.¹⁶ The logic was identical to the case for agents: make comparison free and the rent collapses. It did not collapse. Margins climbed through the entire era of total price transparency. The deepest reason is behavioural and damning: even when the information was free, people did not act on it. Work on consumer goods by Döpper, MacKay, Miller and Stiebale found that across 2006 to 2019 real prices stayed roughly flat while marginal costs fell, widening markups, with the econometrics pointing to declining consumer price sensitivity rather than collusion.¹⁷ A price-comparison website only helps the customer who bothers to use it and then bothers to switch. Most did neither. The open question, the one the whole thesis turns on, is whether an agent that does the bothering for you breaks the behavioural barrier where a website merely lowered it. There is reason to think delegation is categorically different from information. But it is a hypothesis, not yet a result.

Two: the gains may be reallocated, not erased.

The superstar-firm research, led by Autor and colleagues, found that cheap information technology did not flatten margins across the board last time. It concentrated them. The most capable firms used cheap IT to build logistics and software advantages, took share, and earned higher markups, while the median firm saw little change.¹⁸ If agents behave like the last wave of cheap information, they may not lower the average so much as widen the gap between the firm an agent recommends and the firm it never surfaces. Being the default answer an agent returns could become the most valuable position in commerce, a new rent for whoever wins it.

Three: rents relocate to whoever owns the new road, including the agent.

This is the most important objection, and it has a precedent and a present danger. The precedent is the travel industry, the clearest case on record of a vigilance machine dropped into a market built on friction. Booking a flight once required an intermediary with access to a reservation system ordinary people could not see; airline tickets generated up to 80% of agency revenue on an uncapped commission.¹⁹ When Expedia unlocked the reservation system in 1996 and the airlines began selling direct, the toll-takers were largely deleted: US accredited agency locations fell from nearly 47,000 in 1998 to around 12,000 today.¹⁹ Yet the rent did not vanish. It moved. Expedia's fastest-growing segment in 2024 was advertising, at $954 million, up 16%, a fresh toll-booth grown by the machine that killed the commission.¹⁹ The internet, in general, did not abolish the booth between buyer and seller; it relocated it to Google and Amazon, more concentrated and more profitable than the layer they replaced.

The present danger is that the agent layer becomes the next such booth. The whole optimistic case requires the agent to stay cheap and contestable, so that the rent it strips from your bank flows to you rather than to whoever sells the agent. Here the evidence cuts hopefully but not decisively. The cost of machine intelligence is collapsing faster than any comparable technology curve: GPT-4-class performance fell from around $20 per million tokens in late 2022 to roughly $0.40 by late 2025, a tenfold annual decline, and open-weight models such as DeepSeek-R1 reportedly reached frontier-class reasoning at a training cost under $6 million, against projections of billions.²⁰ If anyone can run a near-frontier model for the price of electricity, no one can charge a frontier rent for the thinking. But the same research shows the cost of genuinely frontier capability rising threefold to eighteenfold per year, and Gartner is blunt that cheap commodity tokens will not democratise the scarce, expensive reasoning, nor will the savings be fully passed to customers.²¹ Commodity intelligence has no rent. Frontier intelligence might have plenty. Which layer your vigilance agent needs to run on decides who keeps the surplus.

And one behavioural caveat applies even to the rosy case. Agents are not the frictionless rational optimisers the textbook imagined either. Early research evaluating what AI shopping agents actually buy finds them to be a new species of biased decision-maker, with their own quirks that sellers can learn to exploit, through the ordering of options, the phrasing of product data, the gaming of whatever the agent treats as a quality signal.²² The booth, if it returns, may not even need to own the agent. It may only need to learn how the agent can be nudged, which is the oldest move in the behavioural playbook, turned on the machine instead of the human.

VII · Who survives the upgradeThe defensible businesses were never selling friction.

Sort the economy by what its profit actually rests on, and the future falls out cleanly. A business earning on genuine, non-friction advantage keeps its pricing power and gets more efficient on top of it: a real network effect, a regulatory licence, a brand an agent will still recommend because it is the correct answer, a product no rival matches. A genuine low-cost producer gains, because effortless switching finally rewards the efficiency that inertia used to nullify. And businesses selling the things an agent cannot supply (judgement where the stakes are personal, trust where the downside is severe, taste where there is no objective ranking) may find that stripping away the commodity layer leaves their scarce part worth more, not less. Richer consumers, freed from a thousand small overpayments, have more to spend on the genuinely good.

What does not survive is the rent of inattention itself, the portion of any price that exists only because acting was too much trouble. That portion has been quietly enormous: in two trillion dollars of idle deposits, in the four-billion-pound loyalty penalty, in the renewal nobody opens, in the search result that costs money to sit atop. It was always a tax on the predictable limits of human attention, and a machine that never tires of attending is, in the end, a machine for finding that tax and competing it away.

The deep uncertainty is not whether the machine finds the rent. It is whether the machine becomes the next thing standing in the road with its hand out. The internet promised to abolish the toll-booth and built the richest ones in history. Whether agents break that pattern or merely relocate it one layer upward is the only forecast worth arguing about, and anyone certain of the answer is selling something. Quite possibly a subscription. Meanwhile, for the first time since Stigler named the problem, the customer who could never be bothered has hired someone who is never anything else, and a market that ran for two hundred years on the gap between what people would do and what they actually did is about to find out what it looks like with the gap closed.