/
Opinion
A farewell to the internet built for tired people
Agents that never tire threaten the revenue built on people who do. Whether that's good news depends on who the agent really works for.
A lot of the consumer internet makes its money from good intentions that never get acted on. You meant to cancel the free trial before it converted. You meant to compare insurance quotes before the renewal went through. You searched for one thing, a sponsored listing caught your eye, and you bought a cheaper, low quality equivalent.
For twenty-five years those lapses have been a dependable source of revenue, because there was always a person on the other end of the screen, and people get distracted, put things off and give up at the third "are you sure?" screen. Personal AI agents (software that shops, books and cancels on your behalf) have none of those habits, which is a problem for any business whose margin depends on them.
While we’ve seen this coming for some time, things got real, real fast, in September. Meta launched Muse, a personal agent that browses, compares and buys for its users, and it soon topped Apple's US chart of free apps. Within weeks Walmart, Sephora and Expedia had signed up as partners. Amazon on the other hand started blocking Muse from its store, arguing that the agent doesn't identify itself and that Amazon never agreed to let it in. Shares in Airbnb, Expedia and Tripadvisor fell around 5% the week Meta announced its updates.
Muse wasn't alone for long. OpenAI, Grok and newcomer Instinct (which raised $1 billion at a $10 billion valuation barely a month after opening on an invite-only basis) followed suit with their own always-on agents that could organize your professional or personal life.
While the signal is there, none of this means agents run anyone’s shopping yet. Adoption is early and uneven, trust takes years to build, and incumbents have time to adapt. The direction is clear enough, though, to ask which parts of the internet economy depend on human weakness and what happens to them as more of the buying is done by software that doesn't share it. Two distinct revenue streams are exposed, friction and attention, and what replaces them depends on a question nobody can answer yet.
Friction was a business model
Start with subscriptions. In a 2022 survey by C+R Research, Americans guessed they spent $86 a month on subscriptions, but when they went through them category by category the figure came to $219. And 42% said they had forgotten about a subscription they were still paying for. Bad memory accounts for part of that gap, and product design for the rest: free trials that convert without a nudge, cancellation flows with six screens of save offers, default enrollment in add-ons, prices that only reveal their full amount at checkout.
None of this makes a product better, only harder to leave.
But an agent doesn’t tire. It will sit through the retention script, decline every offer, keep the confirmation, and do the same for every idle subscription at once. It will also send you a reminder to cancel and repeat it until you do it.
The second-order effect matters more. Once cancelling costs nothing, switching costs nothing either, and a large group of businesses rely on customers who never switch: car and home insurance renewed without a second look, broadband contracts that drift onto higher tariffs, mobile plans nobody has compared since signing. The Clouded Judgement newsletter recently described the subscription economy as running on "churn that should happen but doesn't". Agents are what makes it happen.
Friction also kills good revenue
Friction protects a lot of bad revenue, but it destroys good revenue too, and this half of the story gets less attention. The Baymard Institute puts online cart abandonment at about 70%. Some of those shoppers were only browsing. Many others wanted the product and gave up at a surprise delivery charge, a forced account creation or one form too many.
Agents turn some of that dead intent into orders. The recipe you saved becomes a grocery basket, the weekend away you never got round to booking takes one sentence, and the plumber you kept meaning to call gets booked. The gain concentrates in frequent, elastic demand (travel, restaurants, reorders, local services), where people buy more when buying gets easier. It does little for infrequent, need-driven purchases, since nobody replaces a boiler twice because ordering one became simpler.
So the dividing line for a business is whether people want what it sells and whether it is the one delivering it. If both are true, lower friction grows its market. If its margin came from the friction itself, it was a toll booth, and agents drive around toll booths.
Substance beats size
Lower friction is usually read as "AI levels the playing field", and that reading isn’t necessarily true.
On the one hand, agents strip out the advantages that worked on human psychology: brand recall, habit, default placement, the ad funnel that put a familiar name in front of you at the right moment.
On the other hand, what they weigh instead can be checked, such as price, availability, delivery speed, returns and track record. Those largely favour scale and time in the market, which usually belongs to the incumbent. Trust takes years to earn, and a smaller company cannot match a large one's purchasing costs, logistics, or customer reviews just by being easier for an agent to read.
In other words, agents work like a solvent on fake moats and a magnifier on real ones.
Challengers win where the leader's position was soft: interchangeable products, markets won on marketing spend, customers who stayed with subpar products out of inertia. Brand still counts where it is part of what's being bought (taste, identity, the trust you place in a luxury house), and much less where it was only how you got noticed. Elsewhere the incumbent often ends up stronger at selling, even as it loses the revenue that relied on human attention.
Advertising doesn't die, it gets sorted
Most of the free internet is paid for by advertisers wagering that a person will stay on the page. Web advertising relies on three things only a person supplies: attention, the impulse gap between browsing and deciding, and a tolerance for interruption. An agent supplies none of them. It never scrolls past a sponsored listing, because it never sees one.
Comparison sites already show what happens to advertising when people stop looking, even though agents have barely arrived. Tripadvisor earns most of its core revenue from click-based hotel advertising: booking sites pay for each traveller who clicks through to them. Its hotels revenue fell 23% year on year in both the first and second quarters of 2026, which the company put down mainly to fewer visitors arriving from search, and so fewer paid clicks. Its chief executive has pointed to "the rise of AI overviews", Google's AI summaries answering hotel questions directly. If a summary can take that much out of an ad business, an agent that does the whole search will take more.
The damage won't be spread evenly, because attention comes in two kinds. Some of it is transactional: you look in order to compare, buy or renew. That looking can be delegated, and the ads attached to it lose their audience. The rest is attention for its own sake. You watch a series, scroll short videos or follow a match because that is the point, and nobody sends an agent to watch their show for them.
Advertising budgets don't vanish when commerce stops reaching people. They have to go somewhere, and the obvious place is where humans are still looking: streaming, social video, podcasts, live sport. Scarcer human attention chased by the same or larger budgets should push prices up. That is a prediction rather than an observed trend, but it follows from plain supply and demand.
The businesses that lose out are those whose audience came to shop: search engines, marketplaces and comparison sites, which sell ads to people on their way to a purchase. Once agents do the shopping, those visitors stop arriving, and since nobody opens a comparison site to pass an evening, the money follows people to the platforms where they spend their free time.
Amazon sits on the fault line
Amazon shows both forces inside one company. It sold $68.6 billion of advertising in 2025, about a tenth of its revenue, and the largest part was Sponsored Products, the paid listings that appear when you search its store. Amazon doesn't report advertising profit separately, but selling space on pages it already runs costs little to deliver, so the margin is probably among the highest in the group.
That is the part agents threaten. An agent searching Amazon on your behalf ignores the sponsored slot and buys what you asked for. When Amazon blocked Muse, it cited security and the fact that it was never asked, and added that agents can't offer its personalised recommendations. A Bloomberg Intelligence analyst put the commercial logic more plainly, pointing out that people browsing Amazon see other things they end up buying. Blocking the agent protects the browsing session, which is where the advertising lives.
The rest of Amazon looks safer. Its logistics network is what an agent needs when something has to arrive tomorrow, so agents may route more orders through it. Prime Video and Twitch sit on the entertainment side of the split, where human attention stays and should become more valuable. The store doesn't die, but it risks being demoted from store and ad network to store and delivery service, essential and paid accordingly. That wouldn't sink the company. It would mean losing a large slice of its most profitable retail business.
The most interesting question is whether ad money pays for Amazon's generosity. My working answer, which is reasoning and not something I can prove from the accounts, is that the logistics advantage comes mostly from scale and density and was built before advertising got big, so it would survive. Ad margin is plausibly the slack that funds the extras, though: free returns, subsidised Prime delivery speeds, loss-leading prices. Take that slack away and the service could tighten at the edges, which over time would wear down the habit of routing everything through Amazon. That would be a slow erosion rather than a collapse.
Who the agent works for
Commerce advertising won't disappear. Sellers have paid to be chosen for as long as shops have had shelves at eye level. But what changes is the method. In a world where agents shop, instead of interrupting a person, sellers will compete to be picked by software: publishing prices, verifiable stock and delivery terms, paying for a place on the agent's shortlist, becoming the default in a user's standing instructions, and paying whatever toll the agent platforms decide to charge.
That last item is the one to watch. A banner announces that it is an ad. An agent that ranks a paying seller first can look like genuine advice. And this matters because the companies building the most widely used agents are also largely advertising businesses: Meta, Google, Grok (via X), Amazon and arguably soon OpenAI. Nobody yet requires an agent to disclose when a recommendation was paid for, and it isn't obvious who would enforce such a rule, especially if the agent companies and ad platforms are one and the same.
Every optimistic claim in this piece (friction gone, good businesses rewarded, better products winning) assumes the agent works for the person who asked it, and nothing guarantees that yet. What ends here is a particular trick, interrupting a distracted person at the right moment. Advertising itself survives, moving from persuasion to placement. An honest agent is the best consumer deal in twenty-five years. A paid-off one is the same manipulation with the lights off.
