The Prosperous Society

The Prosperous Society is a three-hour, four-part series that explores the political economy of artificial intelligence, grounded in the liberal tradition of the Western canon.

Framed as an extended response to John Kenneth Galbraith’s The Affluent Society, which serves as the intellectual foil for the series, The Prosperous Society makes the case that AI is not principally a story about intelligence, but about economics: it represents a progressive technological revolution that shifts the binding constraint on economic growth from production to distribution. In doing so, it redefines the role of digital advertising as the coordination infrastructure of an increasingly personalized economy.

Last summer, when I began outlining The Prosperous Society, I set out to write an economic defense of artificial intelligence to serve as ballast for more pessimistic narratives. It was originally intended to be a short book.

I pivoted to the podcast format when those narratives reached fever pitch earlier this year because I wanted to respond quickly and release new episodes on a regular cadence. But The Prosperous Society evolved into something broader: a defense of liberal political economy in the age of artificial intelligence.

Over the course of three hours, I develop that thesis through four linked but distinct arguments:

• Part 1: Why AI makes distribution more important than production.

• Part 2: Why autonomous commerce is a false ideal.

• Part 3: Why personalization changes the economics of the long tail—and of identity itself.

• Part 4: Where the moral boundary lies between AI that expands freedom and AI that robs us of it.

If the series succeeds, I hope it provides a rigorous intellectual case for why AI, if properly constrained and thoughtfully applied, can reaffirm the liberal tradition of Western political economy while helping to usher in one of the great periods of economic expansion in human history.

Transcript

How does one justify a progressive revolution? Not a revolution against tyranny, or oppression, or unjust treatment, or structural inequality and subjugation, but against stagnation and the will to sit idle when faced with the opportunity to embrace radical, transformative change. It’s an intriguing question. On the occasion of the United States’ 250th anniversary, we might find some guidance from the architects of the American Revolution.

This current moment presents society with another progressive revolution, an AI revolution that promises to reshape our economy, redefine our relationship with technology, and indeed, reconfigure the visceral sense of human identity. The core thesis of this series, the prosperous society, is that artificial intelligence will serve as the economic nervous system of modern society.

It represents a progressive technological revolution that will realign commerce utterly, transforming the central constraint on economic growth from production to distribution. Artificial intelligence presents a profound structural reorganization of the mechanics of the economy. And it will deliver astonishing growth and prosperity as applied at the distribution layer of consumption through the economy’s increasingly intelligent coordination layer, of which digital advertising is today the most important manifestation.

My motivation in writing the prosperous society was to anchor this thesis to the solid intellectual bedrock of the Western liberal canon, to make the case that insofar as these ideas are good, worthy, and aspirational, then so is the prospect of AI’s influence on our political economy. This undertaking is necessary and important because at this moment, the discourse around AI is dominated by two viewpoints from the extremes. Both camps assume that AI is primarily a story about intelligence. One fears that intelligence will eclipse humanity. The other believes intelligence alone will generate unlimited wealth.

I think they’re both pursuing a misguided interrogation. AI is first and foremost an economic technology. As such, it demands an economic interlocutor. Throughout this series, that interlocutor is John Kenneth Galbraith, and specifically, his influential and celebrated 1958 book, The Affluent Society. I position Galbraith as the principal foil for my argument in the prosperous society, not because I believe he was wrong, but because his work serves as one of the intellectual undercurrents to modern critiques of capitalism, advertising, technological progress, and now, artificial intelligence, often without being explicitly cited.

His theory of the dependence effect, that advertising manufactures demand in service of ever expanding production, still echoes through contemporary debates over consumerism and the social value of technological innovation. My argument in the prosperous society is that the economy Galbraith described no longer exists. Artificial intelligence fundamentally inverts the structure of his world. Production is becoming increasingly less bounded rather than scarce, with distribution emerging as the binding economic constraint.

In this world, advertising, but especially personalized digital advertising, ceases to function principally as an engine of persuasion and instead becomes coordination infrastructure. Much of what follows is therefore an extended conversation with Galbraith across nearly 70 years of economic history. There are those who dismiss AI as nothing more than a novelty that has yet to produce any measurable impact on economic productivity and has thus contributed to a financial bubble that will inevitably and likely very soon burst. On that same side, another group of opponents presents AI as imminently attaining self-sustaining agency to a degree that threatens the existence of humanity or at least of human labor.

These narratives talk past each other and fail to recognize AI as a powerful coordination mechanism that makes commerce broadly more efficient. Yet even fewer coherent, convincing arguments exist from the opposite perspective. Most are prosaic, self-aggrandizing, and promotional. I believe the applicability of AI to areas like commerce in ways that are neither viable nor desirable has likely contaminated and vitiated public opinion of AI. It’s hard to get excited about the soaring prospects of artificial intelligence when the most aspirational use cases put forth by its proponents are automatically delivered shipments of dog food or AI generated ad creatives featuring bicycles with two sets of handlebars.

AI can be economically expansive and beneficial to society even as corrosive agency-eroding use cases exist. We can celebrate the technology while rejecting frivolous use cases and products that supplant the kind of individual choice that is morally and expressively meaningful. Over the next three hours, I will attempt to navigate these tensions with four linked but distinct claims. In part one, why AI makes distribution more important than production. In part two, why autonomous commerce is a false ideal. In part three, why personalization changes the economics of the long tail and of identity itself. And in part four, where the moral boundary lies between AI that expands freedom and AI that robs us of it.

Throughout this series, I anchor the discussion to the works of thinkers like Smith, Locke, Hayek, Tocqueville, and Mill because they represent the mainstream tradition of the Western worldview. If one rejects the principles that this tradition rests upon—individual liberty, voluntary exchange, decentralized coordination, private property—then we’re no longer arguing about AI. And that’s fine, but I wanted this series to force that distinction. If AI really represents a civilizational transition, then it deserves to be analyzed with the same intellectual tools we use to understand every previous civilizational transition.

Neither market commentary nor product marketing provide an adequate framework for understanding a civilizational transition. This discussion is weighty, and it demands input from the writers and thinkers with the gravitas to justify a progressive revolution. But similarly, undertaking a work of this scale requires the discipline to reject the impulse to operate at the other extreme: to extol and espouse the merits of anarcho-capitalism and unshackled, uninhibited free markets. My purpose here is not to defend any particular school of economic thought; it is to defend the liberal political economy that empowers the free, open, democratic Western society.

Last summer, when I began outlining the prosperous society, I set out to write an economic defense of artificial intelligence to serve as ballast for more pessimistic narratives. It was originally intended to be a short book. I pivoted to the podcast format when those narratives reached fever pitch earlier this year because I wanted to respond quickly and to release new episodes on a regular cadence. But the prosperous society ended up becoming a defense of liberal political economy in the age of artificial intelligence. I think that’s a more appropriate body of work in the current context.

If this series is successful, and by that I mean it presents a cogent and compelling case for reaffirming the liberal tradition of Western political economy at the dawn of a new AI age, then it will offer not merely an optimistic perspective on artificial intelligence, but a rigorous intellectual justification for embracing one of the great progressive revolutions in human history. I’ve been struck by the response to the prosperous society, from investors to company executives to fans of the podcast. And I’m proud to release the series in its entirety on July 4, 2026. To those listening, whoever you are, wherever you are, I’m grateful for your time and I wish you the personal and economic dividends of a truly prosperous society.

Part one: The primacy of distribution.

Thomas Robert Malthus was an English economist who, along with Adam Smith and David Ricardo, is regarded as one of the central figures of classical economics. But he’s principally known for one idea, which is the Malthusian trap. Although he made other enduring contributions to economics, mainstream economics generally rejects the inevitability of Malthus’s trap. But the more structural idea that growth, and especially growth that experiences accelerating rates, eventually collides with constraints is taken for granted in modern growth theory. John Galbraith, in his seminal book, The Affluent Society, notes that Malthus viewed poverty as the default state of the human condition and therefore wasn’t concerned with alleviating it, and neither was he concerned with how the productive output of society was allocated.

There are two relevant extensions of this line of inquiry that apply to the current question of AI’s role in society and the potential displacement of all white-collar work. One: what happens when the demand for certain types of labor collapses because that labor can be competently replaced with AI tools? And two: can software command any price point at all if it can not only be conjured instantaneously from a prompt, but it can meet the prompter’s specific idiosyncratic needs perfectly?

There is a natural and obvious tension in taking these extensions to their logical limits, which is if the demand for white-collar work is mostly eliminated by software that can write software, what happens to the underlying demand for that software? Who is buying what any company produces? If the entire digital economy is absorbed into a handful of frontier model labs with this inverse Malthusian trap of infinite digital production leading to widespread famine or war as the global workforce adjusts to structural displacement in the knowledge economy, then the global economy must shrink dramatically as a result.

Who is doing the consuming? And if frontier labs at that point are merely optimizing to accumulate more and more of the shrinking global GDP, then they’ll look to economize too, and they’ll displace the human effort that writes the software that writes the software that writes the software until we’re left with one CEO of one frontier lab lording over the remainder of humanity. The collective action problem likely erodes far before that endpoint. In fact, there’s probably some natural equilibrium where the AI tool producers make more money by not encroaching on some aspect of the economy simply to preserve a customer base than by eliminating it completely.

But this is actually the aspect of the AI discussion that I find least interesting or clarifying. Meandering, tedious thought experiments and fantastical projections to logical extremes tend to be jejune and intellectually immature. But similarly, it’s unhelpful to be shortsighted about things like current limitations around the capabilities of frontier models, especially given the rate of improvements that we’ve seen in tasks like code production and other adjacent competencies. Yes, it’s true that the complexity with developing software may mostly be concentrated not in de novo code generation but in software maintenance, but I find very little reason to believe that agents won’t be capable of this in the future, and possibly the near-term future. So I don’t dismiss the software writing software premise on its face.

What we’re left with is a question of where competition, demand, and supply intersect in this new world order. Each era’s economic anxiety reflects its dominant constraint. Food in Malthus’s time, allocation in Galbraith’s, and distribution in the contemporary environment. The central premise of Galbraith’s The Affluent Society is that the so-called conventional wisdom with respect to economic growth had grown outdated in the context of that time. And that in advanced industrial societies, production had become a self-reinforcing system, accelerated in part by advertising, which created consumer demand. Galbraith argued that in an affluent society, the primary constraint was not one of food production, as with the Malthusian trap, but of the optimal allocation of resources between private endeavors and public institutions, resulting in what he characterized as private affluence and public squalor.

This podcast episode is the first installment in a series that I call the prosperous society. In this series, I’ll make the case that the efficiency benefits posed by AI enhanced development tools shift constraints from production to distribution, whereby human attention becomes the scarce resource in competition. This constraint serves as a natural limit on the number of products that can achieve commercial traction, channeling investments into the systems that best match consumer demand with the products that best satisfy consumer tastes and needs, which are personalized advertising platforms.

These systems will capture an increasing share of the digital economy, and in fact, I believe they will become the critical infrastructure of the overall economy going forward given their increasingly critical role in efficiently routing products to consumers. But these advertising platforms will also establish a natural ceiling on the number of products that can operate successfully in a category, given the need to advertise through those systems in aggregating attention. This will force software developers to deliver increasing amounts of value to users to justify ever larger levels of advertising spend, with much of that cost being defrayed for the consumer by their own adoption of advertising monetization mechanics, which becomes an irresistible opportunity given increased demand for human attention from software developers. All of this converges to a dynamic whereby consumer surplus expands in absolute terms even as platforms capture a larger share of monetized value. In this first episode of the series, I’ll argue that distribution becomes the principal concern of software developers in an environment of decreasing production costs.

Part one: The millionaire’s mall. In The Affluent Society, Galbraith introduces what he calls the dependence effect. The concept is simple and it is destabilizing. In advanced industrial societies, wants are not exogenous to production but are rather shaped by it such that production creates goods and advertising creates the desires for those goods. In a society characterized by the abundance of consumer goods, the bottleneck is not production capacity, but the ability to stimulate sufficient consumption to clear it. This was a profound insight in 1958. Galbraith observed a post-war industrial economy in which factories had mastered scale, logistics had improved, suburbanization had reorganized consumption patterns, and mass media like radio, print, and especially television provided a national megaphone for persuasion. The television set became the conduit through which desire was shaped and synchronized. The automobile, the washing machine, and the refrigerator were mass market goods marketed to mass audiences in a mass society through media that emerged that could reach large numbers of people simultaneously.

But advertising in that era operated at scale and with limited granularity. If everyone saw the same detergent advertisement and everyone saw it repeatedly, then preference formation could plausibly be attributed, at least in part, to the repetition itself. The dependence effect describes a world in which production precedes desire and persuasion fills the gap. It presumes a relatively small universe of goods, widely observable consumption, and a media environment that is national rather than personal, as is the modern day social media feed. It also presumes that wants can be synchronized.

That presumption weakens considerably in the digital economy. The post-war consumer could plausibly peruse the marketplace. Goods were visible, retail stores were finite, department stores curated their shelves from a limited set of commercial options. Consumption was legible. But the digital marketplace is not legible. Amazon hosts hundreds of millions of SKUs, the App Store contains millions of apps, Shopify powers storefronts that number in the millions. No individual can meaningfully browse these environments in their entirety. No individual can discover the long tail of digital products through casual observation either. It may be impossible to ascertain the brand of shoe a fellow rider on the subway is wearing if it’s not instantly recognizable without being advertised to.

And digital goods in particular are also qualitatively and fundamentally different from the goods of Galbraith’s era. They are niche and specialized. They target subcultures, micro-communities, and highly specific use cases. They are not laundry detergents or automobiles. They are goods that cannot be marketed efficiently through broad-based, large-audience advertising. The expected conversion rate is too low and the monetization window is too narrow. The willingness to pay is too heterogeneous. They can only be profitably exposed to consumers for whom they are specifically relevant. And this distinction matters. In a digital ecosystem characterized by extreme product heterogeneity and extreme audience heterogeneity, the economic viability of a product often depends on its ability to be matched with precisely the right subset of users through advertising. The salient question is not whether advertising can create demand for an arbitrary product; it is whether advertising can efficiently route existing demand to the product variant most capable of satisfying it.

In a piece I published in 2020, ahead of Apple’s ATT privacy policy, “Does digital advertising create demand?”, I described this dynamic explicitly. The core claim there is that advertising in the digital context does not function as a demand factory; it functions as a demand routing mechanism. Personalized digital advertising matches users with products on the basis of observable signals: historical behavior, contextual clues, demographic features, inferred intent, and increasingly, probabilistic estimates of discretionary spending capacity.

These systems operate through auctions in which advertisers bid against one another for access to users predicted to generate profitable outcomes. The mechanism is not persuasion at scale but selection at scale, and it’s optimized at the granularity of a specific user and not, as in the era of the affluent society, at large geographic regions or sweeping demographic profiles. If a user has exhibited behavior consistent with a preference for mid-core mobile strategy games and a history of in-app purchase activity above a certain threshold, then exposing that user to a new strategy game with a similar monetization profile may be economically rational.

The advertising platform evaluates the probability of conversion, the advertiser’s bid, and the derived expected value to the platform of the advertiser’s ad filling that impression. The ad is served if the expected value exceeds the threshold required to clear the auction. The ad platform is not fabricating desire ex nihilo. It is interpreting signals to indicate a predisposition toward a category of consumption and quantifying that into an expected value. That does not mean persuasion is absent. Creative matters, messaging matters, positioning matters, branding can matter. But the economic viability of most digital products, and of products that are predominantly sold through digital channels like D2C goods, depends less on manufacturing preference and more on discovering it.

This distinction explains why the deterioration of personalization reduces efficiency without annihilating demand. When Apple deprecated the IDFA with ATT, my argument was not that demand would evaporate; it was that the routing would become less efficient. In a world of constrained production capacity and synchronized mass media, Galbraith’s dependence effect had explanatory power. But in the current world of effectively infinite digital shelf space and algorithmic targeting, the bottleneck is different. It is not the creation of wants; it is the efficient alignment of heterogeneous wants with heterogeneous goods. What’s the probability of a user discovering, through entirely random organic diligence, any given product on an infinite digital retail shelf? It’s zero.

To understand how personalized advertising achieves this alignment, it is useful to revisit what I called the millionaire’s mall. In a piece I published in 2024, “Digital advertising, demand routing, and the millionaire’s mall,” I argue that digital advertising economics are shaped by fat-tailed value distributions. The digital advertising ecosystem is even more extreme. The economics of an entire cohort of users could be defined by just a few of them. This is the millionaire’s mall. The distribution of conversion value won’t be normal, but fat-tailed. And achieving those conversions dictates the profitability of the advertising campaign.

The thought experiment I propose in that piece is simple. Imagine you are standing in some nondescript, non-coastal shopping mall and are told that the average net worth of the shoppers in the mall is $50 million. Two plausible interpretations of the situation are that, one, everyone is extraordinarily wealthy, or two, most people are typical and a single billionaire is present. The distribution matters. Digital advertising operates in a similar environment. The vast majority of ad impressions do not result in conversion. Even fewer result in high-value conversion. The economic viability of a campaign can depend on a small subset of users who generate disproportionate revenue.

Targeting does not need to produce a uniform uplift across all users. It needs to shift the distribution enough that the tail contains sufficient value to justify the spend. The millionaire’s mall only requires the presence of one billionaire. Digital advertising is not broadly an exercise in persuasion. Digital ads don’t attempt to convince the median consumer to purchase something they never previously considered buying. It is about identifying the rare consumer whose latent willingness to spend makes the exposure economically rational. The challenge comes in identifying useful relationships and representations from that latent space. The most sophisticated advertising platforms are spending vast sums of money on doing just that.

When conversion optimization is layered on top of targeting, the platform effectively tells the advertiser: specify your objective and your value per objective, and we will attempt to deliver those outcomes at or below your bid price. The advertiser bids based on expected lifetime value, and the platform assumes the risk of wasted impressions and seeks to minimize it through better prediction. An advertiser can ensure that their margin targets are satisfied with conversion optimization by submitting bids against conversion objectives that are discounted against their actual economic value. If an advertiser pays $1 for a conversion, such as a purchase, that it expects to be worth $2, the difference in those values accrues to the advertiser as profit. The platform’s incentive is to refine its predictions continuously. The more accurately it can identify high-value users, the more budget it can capture. Budget flows towards absolute performance based on the advertiser’s ROAS requirements.

This feedback loop is economically expansionary. Better targeting leads to more conversions. More conversions produce more revenue. More revenue supports greater reinvestment into advertising. Greater reinvestment produces more data. More data improves targeting. This is not a machine for manufacturing arbitrary wants. It is a machine for compressing the search cost associated with matching a user to the product most capable of satisfying their existing preferences.

Galbraith’s dependence effect presumes that advertising manufactures demand in order to absorb output. The digital advertising ecosystem presumes that demand is heterogeneous, partially observable, and most importantly, extant and discoverable through data. The digital storefront is too vast for persuasion alone to compress it. No one watches an advertisement for a niche organic dog food D2C brand and decides, despite not owning a dog, to buy it. No one encounters an ad for a hyper-specific subscription box and develops an entirely novel taste as a consequence of a single exposure.

These categories emerged because some cohort of users already possessed a latent demand for them that personalized advertising could route. Personalized advertising makes those categories economically viable. If targeting degrades, demand does not collapse. It is routed through less efficient mechanisms like organic search, word of mouth, editorial curation, and total monetization declines relative to the alternative. The product category persists. This distinction weakens support for Galbraith’s dependence effect in the digital advertising domain. Production is not creating wants and then fabricating demand to satisfy them. Production is responding to the heterogeneous demand signals, and advertising is optimizing the alignment. What’s more, categories emerge because they are only viable because of the distribution capacity of various digital advertising channels.

There is support in the academic literature for the idea that increases in digital advertising spend are consistent with more product varieties being offered. This is because the demand routing value of digital advertising creates commercial viability for those products. This has consequences for the consumer experience. First, ads become more relevant. A relevant ad is less intrusive. It aligns with existing interests. It reduces the cognitive friction associated with irrelevant exposure. In a world where ad inventory is finite and user attention is scarce, relevance reduces annoyance and product distraction.

Second, personalization should improve monetization efficiency. If a platform can reliably deliver conversions at or below a profitable threshold, advertisers are willing to scale spend. That spend supports product development. It supports experimentation. It supports distribution at zero marginal price to the consumer. Many of the digital products that dominate consumer attention today are nominally free. They’re subsidized through advertising. The more efficiently advertising matches demand to products, the more viable that subsidy becomes. The consumer does not pay a direct price, but they exchange value by making their attention available for targeting.

As targeting improves, the expected value per impression should increase. The advertiser can justify higher bids. The platform can extract revenue while still delivering positive return on ad spend. The product developer can invest in features, performance, and user experience. The mechanism here is not coercion, and it achieves alignment across all three parties: the ad platform, the advertiser, and the consumer. If Galbraith described a society in which private production generated artificial wants, the digital ecosystem reflects a society in which private production attempts to identify and satisfy idiosyncratic wants at scale. Those are two very different things.

The dependence effect implied a kind of asymmetry: producers shaping consumers. Personalized digital advertising implies a different asymmetry: data-rich platforms optimizing the allocation of attention among competing producers. In the next episode of the series, I will explore the competitive implications of that asymmetry. But for now, the key point is this: when production costs decline and production heterogeneity explodes, the binding constraint shifts. It is no longer the stimulation of demand in aggregate. It is the efficient routing of heterogeneous demand across an effectively infinite supply landscape. And personalized digital advertising is the infrastructure that performs that routing.

In performing it well, it does not erode consumer welfare. It enhances it by reducing friction, increasing relevance, subsidizing access, and allowing niche products to find the users for whom they are most valuable. This is not an AI doom loop eroding the value of software broadly and subsuming the entire economy into a handful of frontier model labs. It is the foundation of a prosperous society.

Part two: The primacy of distribution. If AI is deflationary for production, it is inflationary for distribution. That framing can sound paradoxical at first. When we talk about generative AI, we tend to focus on the reduction in marginal production cost. Code generation becomes cheaper. Ad creative production becomes cheaper. Iteration becomes cheaper. Entire product surfaces can be scaffolded and deployed with dramatically less capital than even a few years ago. But trivially, when production becomes cheaper, more things get produced. When more things get produced, more firms compete for the same pool of human attention. And when more firms compete for a resource that does not scale, which human attention doesn’t, the price of accessing that resource rises.

In “The Inflationary Impact of AI-Generated Ad Creative,” I try to express this in straightforward economic terms. Generative AI is deflationary for content production but is inflationary for distribution. Generative AI will see the production costs of increasingly complex forms of content like video approach zero. These tools will instigate an immense expansion in the volume of each content format that they perfect. As content proliferates through generative AI tools, the challenge of capturing potential customer attention becomes more acute, necessitating an increased reliance on advertising. This is inflationary. The corpus of content will grow at a much more rapid pace than the human birth rate. Organic discovery becomes ineffective as content mushrooms. This dynamic gave birth to the search ads mechanism in the first place. Generative AI will similarly create competitive friction for the discovery of all forms of content.

That immense expansion is the critical part. Yes, creative becomes incrementally cheaper for existing advertisers. But critically, participation expands because more businesses can run ads. And if more products are trying to reach customers, and if customers still only have 24 hours in a day, then distribution becomes the locus of competition. That is the inflationary dynamic. In auction terms, this is a marginal story. The auction clears at the willingness to pay of the marginal bidder, so as more advertisers join the auction, the clearing price shifts towards whatever the new marginal bidder will pay. That movement raises average customer acquisition costs and progressively prices lower LTV products out of scalable paid distribution.

When there are too many products to discover organically, the system clears through paid distribution. And in digital markets, paid distribution clears through auctions. Auctions are not metaphors; they are concrete mechanisms. And when more bidders show up to an auction for a fixed inventory of impressions, clearing prices may rise. This is where the popular narrative about AI and product creation becomes misleading. There’s a tendency to think that if anyone can build software with AI, then barriers to success disappear. But that assumes production is the primary barrier. In an attention-constrained environment, production is not the primary barrier. Distribution is.

If AI reduces the cost of building a product from $2 million to $200,000, that delta does not necessarily translate into higher profit margins. In a competitive market, it often translates into more budget allocated to customer acquisition. The savings migrate and potentially are competed away in distribution. And this is not conjecture. We’ve already seen this dynamic in mobile gaming with the advent of mobile app stores, in subscription media, in the creator economy. As development tooling improves, more products enter the market. As more products enter the market, customer acquisition costs rise and the bottleneck shifts. AI will accelerate that migration across every form of content that it can produce, which increasingly is every form of content.

Now, this is where I want to strengthen the argument about platform rent capture, because it doesn’t merely rest on increased bid density and clearing prices. It is about what happens when AI expands advertiser participation. In “AI-Enabled Advertising and the Invisible Retail Consumer,” I make two specific claims about what AI enablement does to advertising markets. One: it will improve conversion rates to the extent that every ad performs at its theoretical potential. And two: it will increase participation by allowing any business that potentially could benefit from digital advertising to do so.

Those two effects compound. If conversion rates improve, the expected value of an impression rises for existing advertisers. That increases their willingness to pay in auction markets. And if participation expands, if more businesses are capable of advertising because AI reduces operational friction, then the number of bidders rises. Higher willingness to pay combined with more participants in the auction should result in higher clearing prices. But the invisible consumer concept adds an important nuance to this story. Today, a portion of consumers are effectively excluded altogether or undermonetized in digital advertising markets because their purchasing behavior is not legible through rich behavioral data. They transact offline, locally, and their digital footprints are sparse. That doesn’t mean they lack economic value. It means the current targeting apparatus struggles to value them precisely.

If AI reduces the friction for small and local businesses to advertise by automating creative production, targeting, and campaign management, then those businesses enter auction markets with their own valuation functions. They value consumers that e-commerce advertisers might undervalue. They bid on impressions that were previously priced too low to clear. This expands the bidder base not just quantitatively, but qualitatively. And when a platform intermediates a scarce resource like attention and simultaneously expands the set of buyers for that resource, it strengthens its position in the value chain. The platform is not simply taking a fee; it is operating the market in which scarcity is priced. As AI expands participation and improves performance, platforms don’t need to arbitrarily raise prices; the auction mechanism does that. More bidders, better conversion, and more efficient monetization—these push clearing prices higher. The platform captures a share of that increased value because it controls allocation.

Importantly, this does not contradict the existence of a long tail. AI will absolutely produce a proliferation of niche products, utilities, and small businesses. But the number of scaled winners in any given category remains constrained by distribution economics. And scale is gated by the cost of attention. This is why we won’t see every local restaurant build its own bespoke version of DoorDash for accepting delivery orders. How would its app get discovered? How would it recruit drivers and delivery people? Again, assume every restaurant can develop a perfect, entirely functional app and backend from a prompt. I don’t argue that we are not heading to that eventuality. We are. Assume every restaurant’s app can be maintained by an AI tool cheaply or costlessly. Assume customer support, fraud detection, and logistics and routing can be managed cheaply or costlessly. Provide this local restaurant’s app with every benefit of the doubt and you still confront the reality that if they can do it, so can everyone else. And the savings provided by AI tools in building and maintaining their app are eroded by the cost of getting their app in front of customers, given that every other restaurant on their street is attempting to do the exact same thing.

The ceiling here is the economy of scale or the network effects that explain the success of food delivery apps today. Unbundling every single restaurant into its own app depletes those economies of scale, and the fiercer distribution competition will almost certainly prohibit restaurants below some threshold from participating in independent, scaled distribution. Certainly, some will. But some won’t be able to clear the distribution hurdle and will be better positioned to remain on DoorDash. The ceiling in this context is not about whether software can be built, but about whether independent distribution can be sustained at scale when attention and network density are scarce.

Now, the obvious critique at this point is: if platforms capture increasing rent as attention scarcity intensifies, does that negate consumer surplus? Does all of this just enrich gatekeepers given a structural reorganization around allocation? That conclusion does not follow for a number of reasons. First, because the same mechanisms that increase clearing prices also improve matching efficiency. As more products enter the advertising ecosystem, platforms have a broader selection set when predicting relevance for a given user. This assumed better match quality could increase conversion probability, retention, and downstream monetization. Higher lifetime value supports higher acquisition spend, which sustains the auction. But the consumer experience can improve in the process. More relevant ads are less distracting. They align more closely with intent. And when products monetize efficiently through advertising, they can subsidize access by reducing or eliminating upfront price gates. Advertising revenue, when routed efficiently, lowers direct price barriers.

And there’s another feedback loop worth noting. As competition for attention intensifies, products that aggregate attention, like media platforms and social media networks, face a marginal decision. They can refine attention into engagement, extracting value through subscriptions or commerce, or they can sell attention as inventory into advertising markets. When advertising prices rise, the opportunity cost of not selling attention rises. Some products will choose to monetize more aggressively through ads, which expands ad inventory supply at the margin. It can moderate price inflation without eliminating scarcity. But again, attention remains finite. It is simply allocated through a more complex equilibrium, which brings the conversation back to the core thesis of this episode. When AI collapses production costs, the economic system does not dissolve into frictionless abundance. It reorganizes around the next binding constraint. Human attention does not scale with compute or with model parameters or with token throughput. There are no scaling laws to hours in the day. Distribution is the mechanism through which that finite resource is allocated.

And as AI expands production and participation, distribution becomes the principal concern of software developers. Engineering becomes cheaper relative to marketing, so feature velocity becomes less differentiating than acquisition efficiency, particularly when software can simply be cloned whole-cloth from a prompt. The ability to command attention becomes the principal determinant of success for software. And the firms that intermediate attention by operating the auctions, predicting relevance, and controlling the surfaces through which products are discovered are structurally positioned to capture a larger share of the surplus created by AI-driven efficiency.

That scarcity is not an accident. It is a structural feature of a digital economy organized around attention. And in an environment where attention is the binding constraint, the economics of distribution, not production, determine which products scale, which firms capture value, and how surplus is allocated across the system.

If there is a through line connecting Malthus, Galbraith, and the present moment, it is this: economic systems are organized around whatever constraint is binding. For Malthus, that constraint was food. He believed that population growth would eventually exceed agricultural output. For Galbraith, writing in 1958, production was no longer the binding constraint in advanced industrial economies. Factories had mastered scale, logistics had matured, and suburbanization had reorganized consumption. What concerned Galbraith was not the ability to produce goods, but the allocation of resources between private abundance and public need. In that environment, advertising appeared as a mechanism for absorbing output with production and persuasion interleaved. The constraint had shifted from sheer output to distribution across social priorities.

Today, we are witnessing another migration of constraint. AI is collapsing the marginal cost of digital production. Code, creative, design, analysis, iteration—all become cheaper and faster. The production frontier expands dramatically, but the existence of more supply does not eliminate scarcity; it relocates it. The constraint facing software developers is no longer the ability to build, but the ability to support discovery through sufficient monetization. Human attention is finite. And in a digital economy where distribution clears through auctions, the allocation of that finite resource determines commercial success.

As AI reduces production costs, more products enter the market. More advertisers enter auction systems. More creative variants compete for the same surfaces. The result is inflationary pressure in distribution markets; customer acquisition costs will rise. In that environment, production savings are not automatically retained as profit. They migrate. They are redeployed into distribution and competed away in customer acquisition. The bottleneck shifts from engineering bandwidth to monetization efficiency because monetization is what supports an advertiser’s bid. When distribution becomes the binding constraint, the entities that intermediate distribution—which are the platforms that aggregate and allocate human attention—occupy the pivotal position in the value chain. They do not need to manufacture demand to capture value because they operate the market in which scarcity is priced. As AI expands participation and improves performance, those markets intensify. More bidders, better conversion, broader advertiser sets—all of these increase the value of allocation.

This is not a dystopian claim. It is an equilibrium claim. And most importantly, it does not imply that consumers lose. As matching improves and product heterogeneity expands, consumers encounter products that align more precisely with their preferences. Ever more niche goods become viable. Advertising revenue subsidizes access. Direct price barriers can fall. Surplus is created not by coercion, but by more efficient alignment between heterogeneous demand and heterogeneous supply. But that supply is capped by the underlying monetization power of the product supporting the cost of distribution. Auctions are by definition mutually exclusive. Only one participant can win.

If Malthus worried that production would always lag population, and Galbraith worried that production would outrun socially optimal allocation, we are confronting a different imbalance: production outruns discoverability. In saturated markets, allocation systems matter more than production systems, which leads to the next installment in this series, which relates to the narrative that AI will eliminate the need for advertising altogether.

The idea is that instead of browsing, consumers will delegate purchasing decisions to agents. Those agents will query APIs to discover new products and decision product adoption based on price, with product discovery becoming programmatic, automated, and abstracted from the consumer’s cognizance. But even in a world of total agentic autonomy, discovery requires a catalog. The catalog is the central data structure in OpenAI’s agentic commerce protocol, for instance. It sources the options that can be exposed in the instant checkout viewport. But whether a product catalog takes the form of a traditional digital storefront, an API endpoint, or a machine-readable commerce protocol like MCP, ACP, or Google’s recently announced UCP, the economic function remains the same. Someone intermediates discovery. And when someone intermediates discovery, they control allocation. They decide what is included in the catalog, and that control is economically meaningful. Even if discovery becomes invisible to the human eye, even if it is entirely abstracted away from consumers—and I don’t think it will be—the scarcity problem does not disappear. There will still be more products than any system can prioritize equally. There will still be competition for inclusion, ranking, and prominence. There will still be mechanisms that determine which products are routed to which users. That’s an allocation problem that naturally leads to advertising. So the next installment in this series will examine this proposition more directly: that agentic commerce will not obviate advertising. It may transform its interface and make it less visible to consumers, but it will not eliminate the economic function of paying for distribution within a scarce discovery environment.

Part two: The human nexus of commerce.

One of the great promises of artificial intelligence is the potential for an intelligent system to replace human effort in tedium, to assume the rote and low-value tasks that humans may only marginally outperform computers at, such that more time and opportunity is afforded to more creative or fulfilling work. This is the generally accepted and most readily apparent benefit of artificial intelligence, and it contrasts cleanly with the doomsday scenarios that depict artificial intelligence leapfrogging human intelligence to the extent that humans are ultimately subjugated by the superior competencies of the machines.

The central thesis of this prosperous society series is simple: that AI will engender an era of unprecedented human creative fulfillment, civilizational cohesion, and material abundance, principally as a function of dramatic productive and organizational efficiency gains that accelerate economic growth through deep personalization. The first installment of this series makes the case that much of that growth will be delivered through an increase in output that necessitates more effective and revenue-performant demand routing through optimized digital advertising, which sits at the heart of the modern economy.

All of the arguments in this series, and indeed, in my view, the entire premise of the transformative power of artificial intelligence, sit downstream of that notion. Thus, AI’s impact on commercial output and the ability to pair that output with consumers most efficiently capture its true value proposition. In that sense, my own views on the value and impact of AI are firmly aligned with the former sentiment: that AI will unburden humans from the mundane and empower them to pursue the profound. But how do we distinguish between the mundane and the profound? Because the contours of that distinction color any interpretation of the impact of AI.

For millennia, humans have attached moral virtue to work and commerce. And a large preponderance of enlightenment canon is centered on the notion of the individual’s rights as a centerpiece of a functional, liberal constitutional order organized around private property and mutually beneficial trade. If we consider the ways in which AI may realign society, but especially to society’s benefit, we must consider how and whether those manifestations and implementations of AI comply with these fundamental principles.

In that framework, economic activity is not instrumental; it is expressive. The act of choosing what to produce, what to exchange, and what to consume is the mechanism through which individual preference is realized. It is the expression of personal freedom. And it becomes even more fundamental to the notion of individualism as AI contributes to a vast, dramatic expansion of productive output that can be effectively routed to individual preferences with conversion-optimized digital advertising, as I argue in part one.

And it’s critical to interrogate the consistency of these applications of AI with our fundamental conceptions of freedom and liberty because one can believe that AI will have a transformationally positive impact on society without believing that every possible use case for AI is so. Again, where do we draw the distinction between the mundane and the profound tasks that AI should undertake for us? Because any encroachment of AI on the human effort and activities that render us free, render us actualized, or render us prosperous should be viewed skeptically as consistent with the doomer narrative, no matter how loudly they are championed.

Which raises the question of agentic commerce. This is a nebulous term, often invoked without precision. My sense is that most people who use it don’t really know exactly what they mean by it. Taken literally, it should mean that a person’s AI-enabled agent makes purchases on their behalf. Less literal interpretations could mean that products are surfaced to a person through some AI-enriched sorting or classification logic. I’ve stated that agentic commerce—and by that I mean independent agents not affiliated with any given retail platform acting on behalf of a user—is a mirage. That it may superficially sound appealing, but that it runs counter to the economic incentives of retailers, retail platforms, and consumers.

And I’ve made the case that the lighter-weight implementation that amounts essentially to a recommendation system monetized through an affiliate model will trend toward surfacing the kind of broadly appealing, lower-cost items that consumers simply don’t need recommended to them, erodes the product differentiation and personalization capabilities of AI that I unpack in part one, and delivers less commercial value than a conversion-optimized digital advertising model. I’ll unpack those arguments in more depth later in this episode.

But the more substantive argument I’ll put forward is that agentic commerce, as conceived of with an agent autonomously making purchases on behalf of a consumer, is inconsistent with the fundamental principles of individual empowerment and liberty and, as an application of AI, is detrimental to a free society. In other words, in this episode I’ll make the case that true agentic commerce as manifested through AI-empowered autonomous purchasing agents is inimical to the Western ideals of individual freedom and the moral virtue of economic choice. And I’ll make the case that agentic commerce as manifested through independent agents surfacing product recommendations from retailers and monetizing them with affiliate commissions collapses under its own competing incentives.

And in this way, I propose that AI’s role in the prosperous society must be mediated by positive and negative applications, benign and malign incarnations. And as we delineate between them, we must invoke our fundamental conceptions of social value as rooted in the tradition and ideology of Western political economy. And ignoring that presupposes either that we embrace the negative applications and malign incarnations of AI, which is the doomer position, or that we reject the fundamental precepts of the enlightenment canon, which requires substantially more work.

At first blush, the prospect of truly agentic commerce, wherein a specialized commerce agent anticipates the needs of a user and fulfills them through proactive search and purchase functionality, seems natural, intuitively appealing. Why wouldn’t consumers want agents discovering useful products on their behalf? Any time saved on the chore of shopping can be dedicated to more fulfilling and stimulating tasks or leisure. But this mental model groups all shopping into the same bucket as a burden, where a consumer aimlessly wanders the aisles of a supermarket or home improvement store, ticking boxes off a shopping list. This mental model presumes that all goods are commodities, differentiated principally on price or some other entirely quantifiable factor that can be quickly and objectively evaluated with a cursory glance. This mental model reduces shopping to work, and exactly the kind of work that one would expect to be able to delegate to an AI.

Shopping isn’t recreation, and it’s only personal to the extent that a personal agent needs to express a simulacrum of a consumer’s tastes and preferences. Shopping is a rote, mechanical task, a chore that consumers would prefer to offload onto a machine. Anyone with that worldview must be unfamiliar with the phrase “retail therapy.” In part one, I described John Kenneth Galbraith’s conception of the dependence effect. My view of advertising, but particularly the modern variant of advertising that didn’t exist when The Affluent Society was published in 1958, is that it is not a demand catalyst or a demand fabricator, but rather a routing mechanism for demand. The auction system finds the clearing price for a consumer’s attention with that person’s behavioral history as an input and matches it with the ad that produces the greatest expected value from it.

In Galbraith’s view, agentic commerce should be the antidote to the perversion of demand by advertising and the production-consumption industrial complex. In that framework, consumption should be an optimization task: the efficient satisfaction of needs at the lowest cost. But this would ignore the role of consumption in the expression of the individual, as a means of asserting preferences, of pronouncing one’s character and personality through things like clothes, music, travel, furniture, art, food, devices, and the experiences we choose to pursue. These are outward enunciations of the self. These are representations of the individual.

Agentic commerce represents the opposite of this: a delegated coordination machine that strips the individual of their prerogative to apply their own idiosyncratic tastes in the ways they deem optimal. My point is simply that agentic commerce promises to impose on society exactly what it is deployed to avoid: to strip agency from the individual and render it onto an algorithm. Whether or not that algorithm is tuned perfectly to the individual’s tastes and preferences is ultimately beside the point. The essence of the individual is captured in the action, not the intent to act, or the probability of acting in some way.

Put another way, if we see the benefits of AI to labor as obviating the need for mechanical, tedious, banal, repetitive work, freeing time for more creatively or intellectually fulfilling endeavors, why would we then also apply it to the things we do with the fruits of our labor? Why bother making work more interesting or fulfilling if the use of the paycheck is surrendered? That is dystopian. Society’s adoption of AI should enable fulfillment not just in labor but across the spectrum of daily life. It should provide more choice, not subsume choice, and it should allow individuals to be more expressive, more latitude for people to further the aims for which they care, not less.

And if one argues that agentic commerce is not designed to fulfill big-ticket, high-consideration purchases like cars, or vacations, or even clothing, but rather to ensure that household staples never run out, then it adds marginal value at best. I brush my teeth twice a day. I don’t need to train a deep neural network to understand when I’ll run out of toothpaste. Grocery delivery apps already allow for scheduled purchases. An agent adds nothing to that consumer interaction. The value proposition of an agent is to conduct background research in light of a consumer’s stated and historical revealed preferences and to surface the options that most appropriately capture their tastes while also allowing them to make the final allocation decision. That’s an excellent use case, but it’s not the literal meaning of agentic commerce. And as I’ll argue next, it is unlikely to be fulfilled by independent agents.

My argument against the notion of agentic commerce as embodied by independent agents making purchases on behalf of consumers is normative and conceptual. I make the case that it violates the conception of the sovereign individual as rooted in the tradition of Western political economy. But my argument against the notion of agentic commerce as embodied by independent agents surfacing product suggestions that are sourced from a far-reaching web of retail platforms and monetized through affiliate links is more pedestrian and straightforward. I think the economics simply don’t work and are inferior to those presented by the advertising model.

To understand why this is the case, it is helpful to begin with the most basic distinction between affiliate monetization and advertising, which is not how recommendations are generated or how transactions are executed, but how value is measured and surfaced within a system. An affiliate model monetizes transactions ex post, applying a fixed commission to whatever purchase occurs regardless of the underlying economics of that purchase. An advertising model, by contrast, allocates demand ex ante, using a bid to express the advertiser’s private estimate of value for a given impression. This distinction determines not just how revenue is generated, but how products are ranked, how surplus is distributed, and whether the system can support heterogeneous outcomes across categories, price points, and margin structures.

In an affiliate system, the ranking function has no access to the advertiser’s internal valuation of a conversion and therefore cannot represent margin structure, customer lifetime value, inventory constraints, or strategic priorities that might influence how aggressively a retailer would pursue a given customer. It sees only surface-level signals such as price, historical conversion rates, and contextual relevance derived from the interaction itself. In the absence of a bid or any analogous mechanism for expressing private value, those signals necessarily dominate. And when they dominate, the system converges to a narrow equilibrium that is likely to favor goods that are inexpensive, broadly appealing, and easy to convert across a wide population.

If a model is optimizing for conversion probability without the benefit of a value signal, then price becomes the most reliable lever through which that probability can be increased, since lower-priced goods present less friction and less perceived risk and therefore convert more readily across a diverse set of users. In a system that cannot observe or incorporate differences in value conditional on conversion, those goods will systematically outcompete higher-priced alternatives even when those alternatives produce sustainably more surplus to the retailer when they do convert.

The result is not merely a change in ranking, but a compression of the product space itself, in which differentiation is vitiated and high-margin, niche, or premium products become underrepresented because the system lacks the information required to surface them appropriately. It can infer likelihood of purchase but not the magnitude of value associated with that purchase, and so it optimizes toward what it can observe rather than what actually matters. Furthermore, it prevents new products from having any meaningful opportunity to increase their representation in the surfaced results. This is precisely the problem that the auction mechanism in digital advertising was designed to solve.

By utilizing an advertiser’s bid to evaluate candidate ads based on expected value, an advertising platform optimizes its own revenue subject to quality and other idiosyncratic scoring constraints. It ranks ads according to the amount of money they are predicted to generate for the platform. And because the bid is submitted by the advertiser based on their own proprietary calculations of revenue, an ad can be seen as the content that best monetizes the user’s attention, not in terms of counts of clicks or purchases, but of total transaction volume. In an ad auction, each candidate’s rank encapsulates the advertiser’s private estimate of an impression’s value—its bid—so the winning ad tends to be the highest expected monetization of the impression, and not merely the highest conversion probability.

The significance of this mechanism is that it allows each advertiser to inject private information into the system in a way that is both scalable and incentive-aligned, transforming ranking from a probabilistic to a value-maximizing exercise that aligns platform revenue with total surplus. It ensures that impressions are allocated not to the product most likely to be purchased in isolation, but to the product that produces the greatest expected value when purchased, conditional on both the probability of conversion and the magnitude of the outcome.

An affiliate model cannot replicate this because it has no source of value truth and no way of incorporating advertiser-specific valuations into the ranking process. It cannot distinguish between a low-priced product with a high conversion rate and a high-priced product with a low conversion rate if the former produces a higher expected commission under a fixed fee structure, nor can it internalize differences in margin structure, lifetime value, or viral propensity that would lead a retailer to value one customer above another. It therefore defaults to optimizing on the signals that are available, and those signals are insufficient to support a high-dimensional, differentiated marketplace. The system converges, and in doing so, it narrows.

The second structural issue is more prosaic but no less consequential, which is that the cost structure imposed by affiliate-based commerce is incompatible with the unit economics of a large portion of the e-commerce ecosystem, particularly among mid-market and consumer packaged goods retailers that operate on thin margins and price customer acquisition on a lifetime value basis rather than a per-transaction basis. A single-digit transaction fee may appear modest in isolation, but it does not exist in isolation, and instead it stacks on top of payment processing fees, platform fees, logistics costs, and costs of goods sold, all of which collectively compress margins to the point where an additional fixed percentage fee applied at the transaction level is not incremental but existential for many retailers. Many merchants operate on net margins in the mid-single digits and therefore cannot absorb additional costs imposed uniformly across transactions without fundamentally altering their pricing or acquisition strategies.

The critical issue here is not simply that the fee is high, but that it is exogenous, meaning that it is imposed rather than chosen and does not adapt to the underlying economics of the advertiser. In an advertising system, the advertiser determines the price they are willing to pay for a conversion, calibrating that price to margin, lifetime value, inventory constraints, and strategic priorities, and adjusting it dynamically as those conditions change. In an affiliate system, the fee is fixed and invariant, applied uniformly across transactions regardless of whether those transactions are profitable for the retailer. This distinction between endogenous pricing and exogenous cost imposition explains why advertising can scale across categories with vastly different margin structures, while affiliate models tend to concentrate in high-margin verticals where such fees can be absorbed. It also explains why retailers can profitably acquire customers through advertising, where they control the price of acquisition, but may be unable to fulfill transactions under a fixed commission regime that does not account for their internal economics.

The third structural issue is incentive alignment, or more precisely, the lack thereof. Independent agentic commerce presupposes that retailers will willingly cede the consumer relationship to an intermediary that captures discovery, mediates transactions, and extracts a fee, while also limiting the retailer’s ability to communicate with the customer or utilize the data generated by that interaction in the ways that are most valuable to the business. This presupposition does not hold in a market where the largest retail platforms derive substantial value not just from transactions but from the data and relationships those transactions create.

Amazon derives as much, if not more, value from the user-level data artifacts created by a transaction than from the transaction itself. That data is used for advertising, for recommendations, and for pricing. User-level behavioral data is mission-critical for Amazon. Maintaining exclusive access to it is a strategic imperative. This observation is not specific to Amazon but generalizes to any platform operating at scale in digital commerce or advertising, where the transaction serves as a source of signals that inform targeting, personalization, pricing, and monetization across multiple surfaces and time horizons.

To interpose an independent agent between the retailer and the consumer is to redirect or attenuate those data flows, to limit the retailer’s ability to communicate with the customer, and to reduce the long-term value that can be extracted from the relationship. Even when nominal access to the customer is preserved, it is often constrained in precisely the ways that matter most, such that the retailer may receive transactional information sufficient to fulfill the order but be prohibited from using that information for marketing or retargeting purposes, effectively severing the economic substance of the relationship while maintaining its formal structure.

Under these conditions, retailer participation in independent agentic systems is not a default outcome, but an exception that occurs only when the incremental transactions delivered by the agent outweigh the value lost through disintermediation—a tradeoff that becomes increasingly unfavorable as retailers invest more heavily in their own first-party data and advertising capabilities.

The fourth issue is control, which is closely related to incentive alignment but manifests operationally in the way demand is generated and managed. Affiliate conversions are, by construction, outside of the retailer’s direct control, depending on factors such as search volumes, platform algorithms for surfacing content, and the internal logic of the agent that determines which products are recommended in a given context. A retailer cannot determine when or how often its products are surfaced, cannot scale exposure by increasing spend, and cannot guarantee a minimum level of traffic or conversions, making the channel inherently unpredictable and difficult to manage. Advertising provides a fundamentally different model of control, allowing the retailer to modulate demand through bids and budgets, to scale exposure proactively rather than waiting to be selected, and to align acquisition with business objectives through continuous optimization. This difference transforms customer acquisition from a stochastic outcome into a managed process in which the retailer can determine not only how much attention to activate, but at what cost and under what conditions.

The distinction becomes particularly salient when considering new or emerging products, which, in an affiliate-driven system, suffer from incumbency bias because they lack historical conversion data and therefore are less likely to be surfaced by models that rely on past performance. In an advertising system, the auction provides a mechanism through which new products can compete on value rather than history, enabling entry and experimentation in a way that is not possible when ranking is based solely on observed conversion rates.

The fifth issue is the degradation of data inputs, which is both subtle and compounding over time. If chatbots ingest review content and preempt the monetization of those sources through affiliate links, then those reviewers are starved of income and are unlikely to write more reviews. And the chatbot loses access to a critical input that its own personalization engine needs, creating a feedback loop in which the system consumes the resources it depends on without replenishing them. This dynamic extends beyond third-party review sites to the broader ecosystem of signals that inform recommendation quality, including user-generated reviews, ratings, and other forms of feedback that are typically collected at the point of transaction. If transactions are subsumed into chatbot environments that do not facilitate or incentivize such feedback, then those signals are not captured and the system’s ability to learn and improve over time is diminished.

At the same time, the data available to independent agents is inherently limited relative to that available to large advertising platforms, since it is derived primarily from contextual interactions within a chat interface rather than from a comprehensive view of the user’s behavior across platforms, devices, and categories. Advertising platforms integrate signals from purchase history, browsing behavior, and off-platform events through mechanisms such as pixels and server-side APIs like the CAPI, constructing high-dimensional representations of user preferences that extend far beyond what can be inferred from any single interaction. This difference in data scope has direct implications for personalization, as a system that relies primarily on in-chat context cannot match the granularity or accuracy of one that leverages longitudinal behavioral data, regardless of the sophistication of the underlying model.

The sixth issue is monetization surface area, which determines the total revenue potential of the system. Affiliate models tied to contextual recommendations can only monetize interactions with a clear commercial locus, meaning that a significant portion of user engagement, including informational and exploratory queries, remains unmonetized. Display advertising is not similarly constrained and can operate across a broader range of contexts provided that sufficient data exists to inform targeting, thereby increasing the total addressable monetization opportunity.

Taken together, these issues form a coherent and reinforcing set of constraints that limit the viability of independent agentic commerce as a standalone model. An affiliate-based system lacks a mechanism for expressing true value and therefore converges toward low-price, high-conversion goods. It imposes fixed costs that are incompatible with many retailer unit economics. It conflicts with the incentives of platforms that derive substantial value from owning the consumer relationship. It deprives retailers of control over distribution and reach. It degrades the data inputs required for effective personalization. It operates on a narrower and less informative data set. And it leaves a substantial portion of user engagement unmonetized.

A final constraint on the viability of independent agentic commerce emerges from the incentives of the largest retail platforms, which are not neutral participants in this ecosystem but profit-maximizing entities built around owning the consumer relationship, the transaction flow, and the data generated by both. Any model that attempts to intermediate those components must contend not only with the economic inefficiencies, but with the strategic resistance of platforms whose interests are directly undermined by such intermediation.

This is most clearly illustrated in the case of Amazon, which has no incentive to allow independent agents to route demand away from its own surfaces into an external interface that it does not control, particularly when doing so would diminish its ability to monetize that demand through advertising. Amazon’s advertising business generated $69 billion in 2025 with 22% year-over-year growth. It is deeply integrated with its retail operations, relying on product visibility within its own environment to capture high-intent demand at the moment of consideration and conversion. Any system that captures that moment externally and completes the transaction outside of Amazon’s native experience reduces the inventory against which Amazon can serve ads and weakens one of its most profitable business lines.

The strategic logic for platforms is not to enable independent agents, but to internalize their functionality, ensuring that any agentic layer sits within the platform’s own interface, aligned with existing monetization mechanisms and data flows. Amazon’s Rufus agent reflects this logic, extending the platform’s control over discovery and purchase while preserving its ability to surface sponsored products. This dynamic highlights an asymmetry in how the arguments for and against agentic commerce apply across contexts. The case for agentic commerce aligns with platform incentives when deployed natively, because it increases engagement, improves conversion, and expands advertising surface area within a controlled environment.

The case against independent agentic commerce arises precisely because it attempts to extract those benefits into an external layer that captures value the platform would otherwise retain. Independent agents must therefore operate in an environment where the most important platforms are either competing directly with them or engaging only on terms that limit their effectiveness, while simultaneously building their own agentic capabilities that replicate the functionality in a more economically coherent form.

Advertising simply addresses each of these deficiencies simultaneously, allowing value to be expressed through bids, enabling flexible pricing aligned with retailer economics, preserving data flows and consumer relationships, providing control over distribution and attention acquisition, leveraging rich behavioral data for personalization, monetizing a broader range of interactions, and critically, preserving and concentrating leverage. Ultimately, this undermines support for the independent agentic commerce use case. Advertising is simply the more effective and economically suitable model for driving product discovery.

A rigorous analysis of any nascent technological paradigm involves examining the tasks or objectives it isn’t suitable for. It’s possible to be optimistic about AI and confident that it will accrue significant efficiencies to many aspects of modern labor, that it will catalyze a wholesale evolution of the human-computer interaction model, and that it will reorganize society around more fulfilling, more intellectually stimulating work and leisure without committing oneself to the application of AI to every possible facet of life.

Galbraith’s Affluent Society saw consumption tethered to output, artificially amplified by advertising, and financed through consumer debt. Galbraith observed that the conventional wisdom of seeking full economic production capacity resulted in affluence being concentrated privately to a degree that crowded out expenditure on public goods, things like public infrastructure, to such an extent that consumer spending would ultimately buckle under the weight of unsustainable debt and an immunity to advertising. But that never happened. What we’ve seen instead is the evolution of advertising into a personalized product discovery mechanism that ever more efficiently pairs consumers with relevant products to them. Consumer debt never reached unsustainable levels. Rather, consumer credit card balances as a share of GDP have remained relatively stable in the 3% to 5% range for the past two decades. The advertising-based economy similarly never collapsed.

Galbraith likely couldn’t have foreseen the stunning degree of personalization that modern ad platforms have achieved. Rather than inuring consumers to it, advertising has penetrated every digital surface area and become one of the principal paths to sustaining what I’ve called a humanity-scale business. The promise of AI is vastly more powerful personalization, both in terms of the products made available to consumers, particularly digital products, but also the messaging used to make consumers aware of those products in the form of advertising. That personalization power is what I believe any AI-anchored growth narrative should embrace. In considering an application of AI to an existing consumer interaction or one that is being developed from whole cloth, AI is a pathway to personalization. It extracts and engages with the most robust expression of the individual, and it shouldn’t attempt to suppress or convolve it.

Part three: The collapse of the Pareto principle.

Galbraith describes what he denotes as a new class: a group of people that emerged to pursue work for its intellectual purpose, and not out of necessity, lack of optionality, or financial reward. Galbraith notes that this new class was a product of affluence. As a society became more wealthy as a function of increased productivity, not only could the young and the old drop out of the workforce, but those of prime working age could dedicate their efforts to more rewarding pursuits rather than merely limiting their working hours.

Some of these pursuits may be less oriented around production than they might have been during, for instance, the industrialization process that took root in England in the mid-19th century and was soon thereafter adopted in the United States. About the new class, Galbraith writes that nearly all societies and nearly all times have had a leisure class: a class of persons who are exempt from toil. In modern times, and especially in the United States, the leisure class, at least as an easily identifiable phenomenon, has disappeared. To be idle is no longer considered rewarding or even entirely respectable. But we’ve barely noticed that the leisure class has been replaced by another and much larger class, to which work has none of the older connotation of pain, fatigue, or other mental or physical discomfort.

Galbraith’s point was that the absorption of large swaths of society into this new class represented true social progress because it not only lifted that group into a new intellectual baseline as a function of its emphasis on education, but it also served as a sort of relaxation of labor requirements afforded by society’s affluence. While laborers weren’t granted the option of working less, they were given the luxury of attaching their efforts and energies to projects that held meaningful purpose to them, even though those efforts resulted in less value being created for society and, as a result, lower remuneration.

Galbraith argued that the new class threatened that model of social alignment, which he believed was supported through what he called the dependence effect, which was the use of advertising to fabricate demand. He notes that the new class regime jeopardized this model because if people chose work principally for fulfillment and not as a means of consumption, then the conventional wisdom would be abandoned as a modus operandi.

Galbraith seemed to believe this affluence-driven corrosion of the machinery built to enact the conventional wisdom was inevitable. He believed that the virtue of labor for its own sake, at least so far as it was enshrined in the belief system that characterized the modern economy, had already been repudiated, and that the real issue facing society was not too few jobs but too many people. If this sounds familiar, it’s because it’s essentially the proto-AI doomerism argument.

But what if Galbraith’s presupposition that the marginal utility of products necessarily falls over time as society becomes more affluent simply breaks in the face of technological progress? Galbraith is probably right in the context of the 1950s. Once a household has a car, a television, a washing machine and a dryer, a dishwasher, and other consumer staples, the marginal utility of a better version of those things may not motivate a person to continue to toil away in a factory if the alternative is pursuing a more stimulating but less lucrative career while not having those things.

But the consumption options available in the 2020s are vastly more numerous and varied than those of the 1950s and 1960s. And society sits on the precipice of a dramatic expansion of those consumption options delivered by artificial intelligence. I argue in part one that the binding constraint in an environment defined by a total abundance of choice is distribution: the ability of a firm to reach its potential customers. And I make the point that AI-enabled personalization, from the perspective of production but also the capacity of modern digital advertising platforms to match consumers with the products most relevant to them, ameliorates this challenge, enabling greater levels of commerce through more performant demand routing and allowing every individual to be exposed to the products most relevant to them.

But the positive effects of AI in this regard aren’t limited to discovery and matching. The benefits also pertain to production. Companies can produce ever more niche products because they can be sure that they can avail themselves of the sophistication of these advertising platforms to reach the consumers for whom they are most germane. This is an important idea. In Galbraith’s model of the world, demand is catalyzed by advertising in part because consumer bases had homogeneous needs, and so too were the products that served them. The ability to reach a total addressable market defined what was produced, and since the primary advertising channels like television, radio, newspapers, and magazines didn’t support small and niche total addressable markets, the products that might best service those specific groups of consumers weren’t pursued by industry.

We obviously don’t live in that same world. The media landscape is not restricted to a handful of outlets; it’s nearly infinite. Society has already effectively reached a point where media consumption is personalized, at least in terms of how it is curated. We are moving toward an eventuality in which it is entirely derived. I spoke of this idea in my book, Freemium Economics, published in 2014, in which I described a theoretical continuous monetization curve for freemium products. The theoretical basis of the continuous monetization curve is that a product catalog should be so complete that at any given point in their tenure with the product, users are presented with a diverse and relevant set of potential purchasable items from which to choose.

This catalog should be composed of not only static, predefined purchasable items, but also of dynamic purchasable items created specifically for the needs of a particular user. The size of the range the LTV metric can take is a function of the size of the product catalog. A small product catalog necessarily limits the breadth of values the LTV metric can assume, given that a small catalog of purchasable items doesn’t allow for a large number of combinations of purchases. Large product catalogs offer users choice. The larger the degree of choice afforded to a user, the more the user is given the opportunity to monetize. Engineering continuity in a product catalog generally requires the presence of dynamic products that are not strictly designed but materialize as collections of customized features existing across a large or infinite number of combinations.

This is consequential because it represents a direct repudiation of Galbraith’s thesis, which is that demand is mostly homogeneous, can be catalyzed through advertising, and can be satisfied to an extent that comes close to extinguishing it. That simply doesn’t apply when total addressable markets shrink, possibly even approaching one, and when those ever smaller markets can be reached efficiently through digital advertising. I argue in part two of this series that because consumption choices are expressive acts through which individuality is manifested, the commercial value of AI allows individuals to better invoke their true sense of self as a function of consumption because that consumption is more specifically tailored to their individual preferences and tastes.

In this part of the series, I’ll make the case that this dynamic allows for a feedback loop to take root, where ever more niche products are created in the first place because those products can be efficiently routed to the consumers for whom they are most appealing. Rather than stunt economic growth by directing society toward indolence or idleness, it instills a productive fervor because labor is rewarded across two vectors: satisfaction in addressing more specific and idiosyncratic desires of specific markets, and the ability to use the fruits of that labor to satisfy their own.

AI represents the erosion of the logic of the Pareto principle, the hard floor on the productivity calculus as defined by total addressable market. AI will create an environment where total addressable markets become ever smaller, ever more specific, ever more specialized and personalized. AI does not extinguish economic ambition through abundance; it multiplies ambition by making specificity profitable. The Pareto distribution is named after Italian economist Vilfredo Pareto, who observed in the late 19th century that wealth and property ownership tended toward conspicuous asymmetry. In one of the most frequently repeated examples attached to his work, Pareto noted that roughly 80% of the land in Italy was owned by 20% of the population. Pareto’s broader insight was that social and economic outcomes often cluster in ways that produce heavy concentration at the top and relative scarcity for the many beneath.

Industrial capitalism was organized around large plants, expensive machinery, lengthy production runs, national distribution systems, and mass audiences consuming relatively standardized goods. Under those conditions, concentration was not merely common but often rational. A small number of factories could satisfy broad national demand because the fixed costs of production were substantial and the marginal costs of replication declined with scale. The channels through which consumers learned of products were limited enough that only a narrow band of firms could efficiently command attention. If one wished to sell refrigerators, automobiles, cigarettes, or televisions across an entire country, one required manufacturing capacity, logistics competence, working capital, shelf access, and a presence in the handful of media properties through which demand could be cultivated.

The digital economy began to soften some of these constraints long before the present wave of artificial intelligence took root. Software products, and especially freemium software products, introduced an altogether different cost structure. Once code is written, the next unit can often be distributed at negligible marginal cost. A product can therefore accommodate heterogeneous willingness to pay through differentiated features, subscriptions, cosmetic purchases, virtual goods, or individualized upgrade paths. A sufficiently rich product catalog expands the monetization surface because users encounter a wider range of offers calibrated to their own preferences and tenure states. The catalog no longer resembles the static shelf, but a wholly personalized storefront where each user’s particular tastes are served.

Yet even in the digital environment, concentration persisted because discovery remained scarce. App stores rank only so many apps at the top of a category page. Search engines present only so many results above the fold. Social feeds surface only so many posts before attention dissipates. The long tail existed technically, but technical existence is not the same as commercial viability. Millions of products can reside in a database while only a tiny fraction are ever meaningfully encountered. Thus, the Pareto pattern survived the first digital era because distribution bottlenecks survived it. Infinite shelf space coexisted with finite discoverability.

Artificial intelligence changes the terms of that equilibrium because it acts directly on the problem of matching. Modern advertising systems are demand routing mechanisms that infer latent preferences from behavioral and contextual signals, then allocate impressions through auctions that reflect expected value. The commercial significance of this mechanism is not exhausted by higher conversion rates for incumbent advertisers. It also further reduces the penalty historically imposed on specificity. A niche product with a sharply defined customer profile no longer requires national awareness in order to flourish. It merely requires the capacity to locate the comparatively small set of users for whom it is unusually valuable.

When targeting precision improves, the viable size of a market can shrink dramatically while remaining economically attractive. Or it can increase participation in these digital channels from firms that were otherwise excluded from them as a technical limitation. This is the beginning of the erosion of the Pareto principle in commercial life. I don’t mean that skewed distributions vanish mathematically from the commercial realm, nor that every market suddenly becomes egalitarian. I mean that many concentrations, previously treated as natural laws, were in substantial measure artifacts of search costs, media scarcity, production rigidity, logistical challenges, and coordination frictions.

When those frictions weaken as a result of AI, a broader array of producers can clear the threshold of viability and revenue propagates across more categories. Consumer attention disperses across more offerings and tails thicken. Artificial intelligence intensifies this taste matching granularity because personalization need not stop at the level of recommendation. Every interface element can be tailored. Product descriptions can be rewritten to resonate with different motivations. Creative assets can be dynamically generated for distinct cohorts. Onboarding flows can adapt to prior behavior, including the ad a consumer clicked on to reach that point. Search results can be sequenced according to inferred preference structures. Prices, bundles, merchandising surfaces, support experiences, and educational prompts can all be modulated by systems that learn continuously from interaction data. The storefront becomes eminently malleable.

And when it does, the economics of variety change again. Historically, offering many variants imposed managerial complexity, inventory risk, design cost, merchandising confusion, and diluted messaging. Under software-mediated personalization, many of those burdens can be mitigated. A catalog may appear singular to the operator while presenting itself plurally to the market. Different users encounter different emphases, different combinations, different paths through the same underlying supply. This grants firms something close to individualized merchandising at population scale.

The consequences for the long tail are substantial. Products that once failed because they appealed too narrowly can succeed because narrowness is no longer synonymous with obscurity. Cultural goods that once required mainstream gatekeepers can sustain themselves through direct audience matching. Physical goods with eccentric use cases can aggregate dispersed demand globally. Services built for unusual schedules, rare preferences, or obscure hobbies can locate enough customers to matter. Markets composed of tiny islands become navigable once navigation improves.

Part two of this series argues that consumption choices are expressive acts through which individuality is manifested. The economic implications of that proposition are larger than they may first appear. If consumers derive utility from being understood in their specificity, then personalization is not merely a convenience; it is an act of expression. A society in which people can more readily discover the products, communities, aesthetics, experiences, and tools that correspond to their actual preferences is a society in which revealed preference becomes more accurate because the menu of choice is more commensurate with the person choosing.

This creates a recursive dynamic. Better matching raises monetization because consumers encounter offerings that genuinely fit them. And better unit economics, accounting for more efficient distribution, attract more producers willing to serve narrower cohorts. Greater producer participation enlarges the available set of goods and experiences. The enlarged set generates richer behavioral signals about preference heterogeneity. And those signals are aggregated and improve matching. A system once organized around average demand becomes progressively more adept at serving singular demand.

Galbraith worried that affluent societies saturated with standardized goods and buoyed by advertising would drift toward a kind of purposeless consumption while under-investing in public goods and higher aspirations. That critique had force in an era where abundance often meant another incrementally improved appliance marketed to a national audience through repetitive persuasion. But this emerging commercial environment marks a genuine departure. Abundance can now mean precision rather than repetition. It can mean the ability of a consumer to discover exactly the educational resource, creative tool, recreational community, or product configuration suited to their circumstances. It can mean a producer building something exquisitely useful for a population too small to have mattered previously.

The Pareto principle will continue to describe many phenomena because concentration can arise from any number of factors: talent differentials, network effects, cumulative reputation, capital intensity, and human attention itself. But its authority as a universal explanation is undermined when technology lowers the cost of specificity. We should expect some distributions to flatten, some monopolies of mindshare to fragment, some categories to proliferate, and some hierarchies to lose their monolithic stature.

What erodes is not asymmetry itself. What erodes is the assumption that asymmetry is destiny. And that distinction is central to the prosperous society. Prosperity in the age of artificial intelligence is not merely more output measured in the aggregate. It is a richer correspondence between what can be produced and what particular people, in their particular lives, actually value. Society becomes more prosperous when it not only better activates its diverse and diffuse preferences, but when it understands those preferences more fully through the iterative process of choice.

In the second part of this series, I make the case that commerce is an expression of the self in the sense that it provides an outlet to articulate preferences across a wide variety of opportunities. Note that this argument doesn’t distill down to something like: the purpose of life is to consume. Commerce is art, commerce is dining, commerce is travel, commerce is a hobby, a concert with friends, a trip to the movie theater with family. All of these things are reflections of the self and are important to retain in anchoring our identities. The point of part two is that AI’s role in expanding the breadth and variety of commerce provides more latitude in that expression.

But how do we discover our tastes, our preferences, indeed our personality, or our character? I would argue that it is done through an iterative process of trial and error: exposure to new experiences or opportunities that provide intrinsic personal resonance. Increasingly, that exposure happens on digital surfaces and is mediated by algorithms. The self is dynamic and disclosed through contact with the world. Human beings possess latent capacities and untested inclinations. But these remain indistinct until they are activated through action. We come to know ourselves less through private contemplation than through the cumulative evidence of what we pursue, what we reject, what we persist in, and what unexpectedly animates us once encountered.

That framework maps surprisingly well onto the modern digital environment. Recommendation systems, search engines, advertising platforms, and increasingly AI-mediated interfaces attempt to infer preference from observable behavior. They do not begin with perfect knowledge of the user. They begin with uncertainty, then refine their models through engagement data: searches, clicks, purchases, dwell time, subscriptions, dismissals, returns, repeated use, and countless other signals that function as partial disclosures of taste. Both self-knowledge and algorithmic personalization depend on a movement from latency to actuality.

The deeper connection to this section is economic. If artificial intelligence dramatically expands the supply of content, products, tools, services, and communities, then the value of systems that can match individuals to the portions of that abundance most relevant to them rises substantially. But matching does more than allocate goods efficiently. It can also widen the process through which individuals discover who they are. The algorithm becomes commercially significant not merely because it sells, but because it reveals.

The self becomes intelligible through manifestation. One does not inquire inward and emerge with a completed account of one’s character. Individuality requires movement, the transition from possibility to conduct. One does not know what one really is until he has made himself a reality through action. Our digital footprints over time form a pattern through which preference can be approximated. They place individuality into daylight.

An individual cannot know what he really is until he has made himself a reality through action. We often demand certainty in advance of evidence. Such certainty is unavailable because the evidence is generated through the action itself. Our digital lives increasingly participate in this loop, molded not as static representations that are inferred and held constant, but through sequences of encounters. And it is through these encounters, the majority of which result in no engagement on our end, that we discover our own tastes and preferences. This is sometimes described as an echo chamber, but it’s a mirror onto the self. And what’s more, those preferences evolve and mutate over time. Further, they are sourced from adjacencies that expose us to concepts and ideas that we might not otherwise have discovered of our own volition. Even ignored recommendations help shape preference boundaries by clarifying what does not resonate.

This becomes more consequential as AI expands supply. If artificial intelligence increases the production of media, software, educational resources, niche goods, and highly specialized services, then recommendation systems gain access to a vastly enlarged possibility set from which to test fit. The user who once chose from a narrow menu can now be presented with a broad frontier of options, many of which would never have existed under prior production constraints. Self-discovery becomes richer because the field of discoverable possibilities becomes richer. Identity is the conjunction of internal capacity with external opportunity. Talent alone remains dormant, absent action. And identity emerges when aptitude meets an object worthy of it, and when curiosity meets an avenue through which it can be pursued.

One discovers ability by applying it somewhere meaningful. One discovers meaning by testing ability against the world. Modern algorithms increasingly mediate that conjunction. They can expose an amateur musician to composition tools, collaborators, audiences, and genres previously inaccessible. They can present a person with unusual curiosities to others who share them, converting private eccentricity into social identity. Again, this was the initial promise of the internet, and it is amplified under the auspices of artificial intelligence across not just content creation, which is an obvious artifact, but through content recommendation and distribution.

There are, of course, risks in this process. Those risks have been well-documented and are real. But those risks don’t negate the broader structural point. Properly designed systems widen experiential surfaces by lowering the cost of production and improving efficiency in distribution. They help people test hypotheses about themselves against a larger world. Many preferences are neither fully formed nor wholly fabricated. They’re nascent and contingent, awaiting collision with the right opportunity. Recommendation systems often perform the matching function that allows those dormant inclinations to become explicit.

The result is a feedback loop between selfhood and system intelligence. We act, and our actions reveal something about us. Systems observe those revelations and surface new opportunities. We respond to those opportunities and, in responding, learn something further about ourselves. The profile becomes more accurate as the person becomes more defined. And identity is refined through that participation.

I’ve made two interleaving claims here. The first is that many of the concentrations historically described through the language of the Pareto principle were downstream of technological and institutional constraints—the economic necessity of targeting broad averages rather than particular individuals. The second is that preference itself is often clarified through encounter, experimentation, and visible acts of choice, increasingly mediated through algorithmic systems that learn from behavior and respond with new opportunities, but nonetheless explore a larger opportunity set than what we might encounter in their absence. When these claims are considered jointly, the role of AI is clarified as delivering personalization as an engine of self-identity.

The most immediate commercial consequence of better matching is higher levels of engagement and better unit economics. When a consumer is presented with a product, service, experience, or piece of content more closely aligned with their actual preferences, the probability of engagement rises. The probability of purchase rises. Retention often improves. And satisfaction can improve with it. This dynamic has already been demonstrated repeatedly in digital advertising markets, where superior targeting and relevance can justify higher bids, deliver stronger returns on ad spend, and result in deeper participation from advertisers. But the principle extends beyond advertising inventory. Any system that more accurately pairs differentiated demand with differentiated supply increases the value latent in both sides of the exchange.

Those improved returns then attract additional participants. Producers who might previously have judged a market too small, too diffuse, or too expensive to reach can now rationally enter it because distribution becomes more precise and measurable. Firms can build for narrow cohorts with confidence that those cohorts can be assembled economically. Incumbents can pursue specialized sub-brands, limited releases, or tailored product lines because demand can be located with greater certainty. The commercial threshold for viability declines. Markets once ignored because they were too specific become investable because specificity itself presents an opportunity.

This is where the distinction between advertising and recommendation systems begins to lose practical relevance. In a world of sufficiently advanced personalization systems, both functions converge around the same underlying task: identifying what a person is likely to value and placing it before them at the right moment, in the right context, with the right framing. Commerce becomes a continuous matching process. That process applies with particular force to physical goods, which have historically been constrained by forecasting error and uncertain demand. But AI-enabled personalization changes that calculus. The range of goods expands because the certainty of reaching relevant buyers expands alongside it, through the commercial benefit of better matching and more efficient distribution.

As that supply expands, the systems responsible for personalization improve in turn. A richer catalog of goods, services, communities, and content generates denser signals about what people actually value under conditions of genuine choice. When the menu is narrow, preference data is crude because choices are constrained. When the menu becomes broad, observed behavior becomes more revealing. The user who consistently chooses one niche aesthetic over another, one learning modality over another, one form of recreation over another, communicates something more precise than the user who simply selects from a handful of mass market defaults.

This is why algorithmic systems increasingly function as discovery tools rather than mere sorting mechanisms. As argued in part two, commerce often serves expressive ends. It helps individuals articulate values, tastes, aspirations, and affinities. In an environment where artificial intelligence expands productive capacity and recommendation systems navigate that enlarged field of options, discovery itself becomes a mirror. Individuals are surfaced products they did not know existed, communities they did not know they would value, creative works they would not have independently sought, tools that unlock dormant capabilities, and experiences that refine their understanding of themselves. Preference is inferred, then tested, then refined.

The macroeconomic implications of this process are substantial because improved coordination reduces waste throughout the system. Producers spend less capital broadcasting irrelevant messages to indifferent audiences, and consumers spend less time searching through unsuitable options. Inventory can be planned with greater granularity. Innovation becomes less dependent on appealing to the median buyer and more responsive to dispersed tastes that previously remained commercially invisible. This is a quieter form of productivity growth than the image of towering factories or dramatic automation, but it may prove no less consequential.

Artificial intelligence amplifies the flywheel further because it expands not only recommendation capacity, but supply itself. AI-generated creative assets can tailor messaging to narrower cohorts. AI-assisted design can accelerate the creation of specialized products. AI-authored media can serve previously neglected tastes. AI-enabled software can produce tools for small professional communities or hobbyist groups that would never have justified bespoke development under old cost structures. Every increase in productive flexibility enlarges the universe of matchable supply, which gives personalization systems more to work with, which increases returns, which attracts further participation.

We can therefore describe a recursive commercial dynamic: better matching increases monetization. Stronger monetization attracts producers and advertisers. Greater participation expands supply. Broader supply improves personalization inputs. Improved personalization deepens matching once again. The cycle compounds. And unlike Galbraith’s dependence effect, in which advertising was said to fabricate demand in order to absorb standardized output, this system is oriented towards discovering heterogeneous demand in order to support differentiated output. It does not rely upon the erosion of utility through repetitive persuasion; it relies upon the revelation of utility through relevance. This is the synthesis of the prosperous society. In a world of abundance, consumption can become a mechanism of self-discovery because the range of available choices is wider and the systems mediating discovery become increasingly adept at understanding lives in their particularity. Prosperity is not captured merely by more units produced or more dollars spent. It is reflected rather in the richer correspondence between human individuality and the goods, experiences, and opportunities through which that individuality is expressed.

Part four: Per commercium virtus.

The so-called conventional wisdom of maximal economic output as a matter of principle encouraged production for its own sake. That production wasn’t tethered naturally to demand, so demand had to be manufactured in what Galbraith described as the dependence effect. Advertising catalyzed consumer demand that absorbed the products being produced in accordance with the goal of maximal output. Because the consumer economy was insufficiently large to attain maximal output, the public sector and, in particular, the military, needed to be enlisted in service of that goal with the research undertaken there creating spillover effects that resulted in new consumer products being invented that could advance the economic frontier. But these research and development efforts aimed at building weapon systems capable of eliminating humanity created obvious existential risks.

As I’ve noted throughout this series, we no longer live in the world of Galbraith’s affluent society, which he characterized as overinvestment into private enterprise at the expense of public goods, or private affluence and public squalor. I don’t think Galbraith was wrong per se. The world was a different place. But his work is worth revisiting now because his arguments echo to this day. Galbraith’s ghost is still in the room. The arguments he advanced in his book 70 years ago are used now, sometimes verbatim, to protest against innovation in digital advertising and in artificial intelligence. But I don’t think they’re adequate in interrogating the current moment.

Galbraith observed a post-World War II industrial economy that focused on mass market good production for consumers that, in large part, were buying these products for the first time. New construction methods reduced the cost of development dramatically and, combined with backing from the Federal Housing Administration and policies like the GI Bill, which allowed military veterans to buy homes without a down payment and at low interest rates, contributed to the rapid suburbanization of the United States. Living in larger homes, in farther-flung suburbs, the American middle class required consumer goods like washing machines, dishwashers, televisions, and automobiles. These goods were promoted in mass market media that reached a middle class demographically clustered in spacious, affordable homes in family-friendly neighborhoods, and buoyed by meteoric wage growth and in the expanding professional economy of the world’s remaining superpower. This newfound financial stability precipitated a demographic explosion, the baby boom.

But we no longer live in the society of Galbraith. The country is aging. No natural built-in demographic dividend exists. There is no population tailwind producing de facto economic growth. If we desire economic growth—and we should; I’d argue that civilizationaly we must—then we are forced to engineer it. The economic structure Galbraith described does not apply to the society we occupy. The basket of tools and technologies characterized by artificial intelligence is the clearest encapsulation of that economic growth opportunity.

Unlike in Galbraith’s affluent society, our modern economy isn’t driven by mass market media and household staples. It is increasingly personalized and individuated, mediated by digital advertising targeted behaviorally to niche tastes and preferences. This is empowered by artificial intelligence. What’s more, the firms that operate the largest and most sophisticated of these digital advertising platforms are the ones investing most heavily into artificial intelligence research and compute resources. Galbraith’s economic constraints have been reversed. Our largest private firms sit at the innovation frontier and deliver technological breakthroughs with the military as a licensor. The economic promise of artificial intelligence lies not merely in increasing production, but in increasing the precision with which human preferences, talents, and ambitions can be matched with economic opportunity.

In part one, I described this as the primacy of distribution. The bulk of AI investment is currently being undertaken by the world’s largest advertising platforms. That’s no accident. The distribution layer of the internet economy, which is increasingly the distribution layer of the overall economy, is the point at which AI produces the most commercial impact. As these distribution systems get ever more sophisticated, precise, and effective, the economy grows not just through improved matching and targeting, but also through the compounded effects of digital advertising performance, lower barriers to entry in the digital advertising economy, and the expansion of production of ever more niche goods that are only viable as a result of improved distribution.

This is a flywheel across three axes: efficiency, participation, and product availability. More products for more advertisers with higher levels of adoption because they can be targeted more precisely. The application of AI to digital advertising provides for maximal commercial personalization at historically unparalleled precision. AI increasingly rich user representations can be computed efficiently and deployed at massive scale. The significance of this architecture is not merely technical sophistication. It is the increasing precision with which platforms can model consumer receptiveness and commercial intent. Behavioral targeting systems ingest sequences of historical actions and engagement patterns in order to produce richer representations of preference and intent. Request-centric ranking architecture allows these systems to evaluate commercial relevance with increasing granularity while maintaining the latency standards required for scaled advertising markets.

The practical consequence of this sophistication is that specificity becomes economically viable. And this is the key transition in the flywheel. Historically, mass market economics favored scale because scale reduced distribution costs. Broadly appealing products were economically advantaged because broad appeal facilitated efficient marketing through the small number of mass media channels that existed. So niche products struggled or were simply prima facie non-viable and thus not produced because discovering the consumers for whom they were relevant was too expensive relative to the size of the addressable market. Digital advertising has eroded that constraint, and AI enablement will accelerate that erosion. Better targeting allows increasingly specific products to find increasingly specific audiences profitably.

And when specificity becomes economically viable, the structure of the economy changes. Ads targeted to smaller, more niche, and better-defined audiences can generate net new revenue in targeting inventory that might otherwise have gone unsold. This insight extends beyond advertising auctions. Better matching infrastructure creates net new economic activity because products that previously could not sustain economically viable distribution suddenly can. This dynamic becomes increasingly important when considered alongside the broader structure of retail commerce. E-commerce accounted for just 16.6% of total US retail sales in the fourth quarter of 2025. That figure should provoke reflection because it implies that the overwhelming majority of retail spend still occurs in environments characterized by vastly less efficient discovery and merchandising infrastructure than what digital systems accommodate.

A substantial portion of future economic growth will come from the migration of commerce into environments where matching precision is higher. And this transition compounds recursively. Better targeting increases the viability of niche products. More niche products allow more firms to participate. More participating firms create more experimentation. More experimentation broadens preference satisfaction because consumers encounter products that more closely align with their actual tastes and identities. Broader preference satisfaction increases economic output because more commercial activity becomes viable. This is what I described in part three of this series as the collapse of the Pareto principle.

That phrase should not be interpreted to mean that inequality disappears or that market concentration evaporates entirely. Industrial economies naturally concentrated because industrial production and mass media distribution rewarded standardization and scale. And large firms enjoyed structural advantages because broad products distributed through broad channels were economically dominant. But AI-enabled targeting infrastructure weakens those concentration dynamics because specificity becomes commercially sustainable. In fact, specificity could become a competitive advantage. A product no longer needs universal appeal to support a meaningful business. That is a profound change. The industrial economy rewarded homogenization because homogenization simplified distribution. AI-enabled commerce rewards differentiation because differentiation improves relevance and because relevance improves commercial efficiency.

The result is not the elimination of large firms, but the weakening of the structural advantages historically attached to mass market standardization. The economy becomes more heterogeneous and, importantly, this heterogeneity is economically expansionary because these sophisticated, AI-enabled systems coordinate it. The prosperous society does not emerge from limitless production in the abstract. It emerges from increasingly precise coordination between highly specific and differentiated supply and demand. This coordination dynamic also explains why AI infrastructure investment can be economically rational even at extraordinary scale.

The economic value of AI infrastructure lies not merely in present monetization, but in the future coordination capacity it supports. This is the fourth stage of the flywheel: compounding economic participation. Performance advertising creates a recursive reinvestment dynamic in which commercial success funds additional advertising expenditure that, in turn, funds additional growth. The significance of performance marketing is not merely that advertising spend can be measured against commercial outcomes. It is that profitable customer acquisition produces a compounding cycle of reinvestment and expansion. Performance advertising is not static allocation. A business that reliably acquires customers profitably does not simply maintain a fixed level of advertising expenditure. It reinvests. More customers generate more revenue, and more revenue funds more advertising. This compounding loop is why performance marketing became such an enormously powerful force within the digital economy and why platforms optimized around measurable commercial outcomes became dominant.

Now, imagine the size of the absolute effect when that 16.6% inches upward because SMBs that were previously excluded from digital advertising are participating in that market. And when it inches upward again because new firms emerge to serve niche tastes that were previously impractical to satisfy through blunt marketing tools. Consider how much slack exists in just the consumer economy that can be addressed with AI enablement at the distribution layer. And critically, the flywheel of the prosperous society is not merely mechanical. It is expressive, personal, and socially productive.

As products become more specialized and discovery becomes more precise, individuals gain greater capacity to express preference and identity through commerce. The economy becomes increasingly capable of supporting differentiated tastes and ambitions and forms of self-definition that previously lacked sufficient scale to survive commercially. This is one reason why digital advertising should not be understood merely as a revenue mechanism for internet platforms. Digital advertising is coordination infrastructure for an increasingly differentiated economy. The prosperity generated by this system is not confined to aggregate GDP expansion, although that’s an important consideration, and it undoubtedly produces that. The prosperity emerges from increasing the fidelity and resolution of human creativity, of the human propensity for self-expression. And this is ultimately why the AI investment cycle currently underway is rational. There is immense value in unbridled personal expression.

If the application of AI merely automated existing workflows while producing no expansion in participation or coordination capacity, then many of the concerns around speculative excess would be justified. But the largest advertising platforms in the world are investing aggressively into AI precisely because they recognize that increasingly precise economic coordination unlocks enormous latent commercial activity, not just through more precise targeting, but expanded participation and more expressive commercial participation from consumers. If I buy things that make me happier, I’m not just an affirmative response in a customer survey. I’m a more contented member of society. This sounds nebulous and rhetorically saccharine, but it isn’t. Consumers are willing to spend more for goods that better meet their needs.

The future growth opportunity does not exist solely in improving monetization against existing economic activity. It exists in onboarding incrementally new firms and in enabling incrementally new forms of commerce that historically could not participate profitably. The prosperous society flywheel can therefore be stated relatively forthrightly: AI increases productive possibility and expands the universe of products for sale. Advertising increases matching precision within that expanded universe. Better matching individual tastes with the products that satisfy them. That specificity increases expressive individuality. That individuality expands economic participation. That participation compounds economic growth. That growth funds further infrastructure and innovation. That is the flywheel.

The argument presented throughout this series is ultimately not about advertising or artificial intelligence in isolation. It is an argument about coordination and individuality and the conditions under which human flourishing becomes economically sustainable within technologically advanced societies. And while I believe the application of AI to the distribution layer of the economy can unlock enormous productive and creative potential, I also believe that this outcome is not guaranteed. The same personalization infrastructure that increases expressive capacity can also produce fragmentation and passivity and alienation if applied without restraint or wisdom.

This is why I believe the contemporary debate around AI is so frequently unsatisfying. Much of the discourse oscillates between the extremes of apocalypticism and utopianism, not just from the perspective of human agency and self-determination, but also from the perspective of the economic sustainability of the investments being made into the compute infrastructure being built to support the application of AI. Both perspectives flatten the question into something deterministic. But technology does not possess moral direction independent of the institutional and commercial incentives through which it is deployed.

The application of AI to commerce can be expansive and humanizing, or corrosive and socially destructive. It can be wildly expansionary or a total deadweight loss. The distinction depends on whether these systems augment human agency or progressively displace it. This concern is not new. Tocqueville articulated how democratic societies contained within them an underlying tension between individuality and atomization. In Democracy in America, he observed that democratic citizens are simultaneously liberated and isolated, freed from inherited structures and aristocratic hierarchies, yet increasingly susceptible to a kind of anxious restlessness that emerges when identity must be continuously self-authored.

Democratic equality expands personal aspiration while also weakening the intermediary structures that historically grounded social cohesion. Individuals become more independent while simultaneously becoming more psychologically vulnerable. That observation feels remarkably contemporary. The digital economy increasingly organizes itself around personalization. Increasingly, AI systems mediate not merely what we buy, but what we see, and what we encounter, and what forms of information are and become salient to us all. And while this personalization infrastructure can produce enormous welfare gains by reducing friction and improving relevance, it also introduces the risk that individuals retreat progressively inward into algorithmically curated realities that reinforce preference rather than challenge it.

Individuality is not an unalloyed good. This is an important distinction because the argument presented throughout this series is not that all forms of personalization are inherently emancipatory. Human flourishing requires individuality, but individuality untethered from shared civic and cultural structures can devolve into fragmentation. A society in which every individual inhabits a unique and siloed digital consciousness may also become a society in which common experience deteriorates and social solidarity weakens. Hyper-personalization can increase expressive freedom while simultaneously eliminating the shared perspective that binds a society. And this is where the moral dimension of AI becomes unavoidable.

The flywheel described in the previous section is economically expansive because it increases the precision with which differentiated preferences can be matched with differentiated products and opportunities. But that process only remains socially constructive insofar as it expands agency and participation and creativity rather than replacing them. The distinction may sound abstract, but it becomes obvious when considered in the context of agentic commerce. Consumption is not merely transactional; it is expressive. The choices that individuals make about what to buy, and what to wear, and what communities to affiliate with, and what aesthetic identities to cultivate are intertwined with the process of self-authorship itself.

Commercial participation is not reducible to utility maximization because human beings do not experience life as optimization functions. Meaning emerges through aspiration and discovery. The act of choosing is itself meaningful. This is why I remain skeptical of visions of fully autonomous agentic commerce in which AI systems continuously purchase products and services on behalf of users with minimal human participation. Such systems may produce extraordinary efficiency gains. They may reduce cognitive overhead and compress transaction costs. But not every application of AI expands human flourishing merely because it increases efficiency. That distinction is central to the entire argument presented throughout this series.

There is an important difference between systems that augment human intentionality and systems that replace it. A recommendation system that exposes an individual to a niche product category that aligns with latent interests can expand self-discovery. A system that autonomously resolves all acts of consumption on behalf of the user risks undermining intentionality altogether. Convenience can quietly suppress expression. And this distinction extends beyond commerce itself. A healthy commercial society does not merely allocate resources efficiently; it creates conditions under which individuals can pursue differentiated ambitions and discover differentiated identities. Economic systems are ultimately social systems because they shape how individuals encounter one another, and how they perceive possibility, and how they orient themselves toward aspiration. Commercial abundance alone is insufficient if the mechanisms producing that abundance simultaneously erode the human capacities that give abundance meaning.

This is why I have repeatedly emphasized throughout this series that the most economically meaningful application of AI lies in expanding productive and expressive participation, rather than eliminating it. The fundamental distinction between the two is that rendering distribution more informed and capacious expands choice, and suppressing or short-circuiting the act of doing something suppresses it. This distinction isn’t exclusive to commerce and shopping. It’s universal across potential applications of AI. And a firm’s decision to pursue one strategic purpose or another dictates whether its investments in that infrastructure can be profitable.

Good AI expands agency; it doesn’t subvert it. It expands participation in commerce by reducing barriers to entry. It expands creativity by reducing operational friction around production and distribution. It expands discovery by allowing individuals to encounter products and communities and opportunities that would otherwise remain invisible. It expands self-authorship because increasingly precise coordination infrastructure permits more differentiated forms of economic participation to become viable. Bad AI progressively displaces agency. It narrows experience by optimizing excessively toward behavioral predictability. It collapses intentionality by automating decisions that constitute personal identity. It substitutes passive consumption for active exploration, and ultimately it risks reducing individuals to observers of the mechanisms that give their lives meaning.

These distinctions are subtle, but enormously important because they explain why I remain simultaneously optimistic and cautious about the trajectory of AI-enabled commerce. I believe deeply that personalization can be socially beneficial. In fact, I would argue that modern digital advertising systems have already generated enormous welfare gains by democratizing access to information and entertainment and entrepreneurship. Personalized advertising funds a substantial portion of the free digital infrastructure upon which contemporary society increasingly depends. It allows niche businesses to compete against incumbents. It allows consumers to discover products that more precisely satisfy their preferences. It reduces the brute inefficiency of mass market broadcasting and improves the precision with which commercial information is routed through the economy. It undermines many of the issues that Galbraith took with the economic structure that dominated in the middle of the last century.

But these systems become socially dangerous when optimization ceases to serve human intentionality and instead begins to replace it. One shortcoming around the contemporary commentary about AI is that it ignores how contingent these outcomes are upon institutional design and competitive incentives. The same foundational technologies can produce radically different social outcomes depending on how they are applied. Recommendation systems can facilitate discovery and curiosity, or they can facilitate social isolation and alienation. Personalization systems can expand expressive possibility or they can produce algorithmic solipsism. The moral choice therefore does not exist at the level of whether AI should advance; that is largely settled already. The moral choice exists at the level of what kinds of human capacities these systems are designed to amplify.

And this brings the discussion back to infrastructure investment. Throughout the current AI cycle, enormous attention has been devoted to the scale of hyperscaler capital expenditure and to whether the infrastructure currently being constructed can justify its implied valuation. This debate is frequently framed almost exclusively through the lens of near-term monetization: how many subscriptions can be sold, and how many inference requests can be monetized, and how quickly AI products can generate operating leverage. But that framing misses the deeper significance of what is being built. The infrastructure being constructed today is ultimately coordination infrastructure. Its value does not emerge solely from replacing labor or generating synthetic content or reducing customer support costs. Its value emerges from expanding the capacity of economic systems to coordinate increasingly differentiated forms of production and discovery.

AI infrastructure matters because it can expand the number of people capable of participating meaningfully in sophisticated commercial activity, and because it can increase the precision with which human creativity and ambition are translated into economically sustainable outcomes. And this is true across both digital and physical products. When AI is attached to the distribution and discovery layer of the economy, it impacts everything, not just chatbots or SaaS software licenses. But these investments are justified only insofar as they support the flywheel of human flourishing described throughout this series. The modern economy increasingly derives value from intangible and expressive forms of production. Software and media and entertainment and education and community formation and entrepreneurship are all becoming progressively more individualized and digitally mediated.

In such an environment, coordination infrastructure becomes extraordinarily important because the economy itself becomes more differentiated. The industrial era depended upon standardization because standardization simplified distribution. The AI era increasingly depends on personalization because personalization allows differentiation to scale economically. But again, differentiation alone is not sufficient. A civilization cannot sustain itself purely through hyper-personalized consumption loops detached from shared institutions and shared meaning. Open societies require intermediary structures capable of binding individuals together despite expanding personal autonomy. One danger posed by algorithmic systems is that they may progressively weaken those structures by mediating experience too individually and too frictionlessly. Individuals can become isolated, but more than that, they can be condemned to prescriptive roles from which they cannot escape because the systems that mediate discovery strip them of choice.

Here I’d refer to the work of Karl Popper in The Open Society and Its Enemies. Popper distinguished between closed and open societies on the basis of freedom of choice and social mobility. Popper’s point here is that the prescriptive rigidity of the closed society, based on tribal norms, taboos that can’t be broken, and in some cases explanatory magic, encapsulate the restraints of a closed society. And that the open society, even in its exaggerated abstract form, allowed for personal discretion, choice, rejection of taboos, and mobility. Popper made clear that while the rigidity of Plato’s concept of happiness was alluring, once a society had discovered the benefits of individualism and personal choice, it simply couldn’t abandon them.

This is why I believe restraint and intentionality matter profoundly in how AI systems are designed and deployed. There are applications of AI that would obviously be more reflective of Popper’s conception of a closed society: that would remove autonomy, erode personal choice, and place society into cohorts or segments that are permanent and non-malleable. That is a frightening outcome, even if it is packaged under the auspices of convenience. As I argue in part three of this series, some forms of friction are expressive and capture meaning. Discovery itself often requires uncertainty and exploration and experimentation. Relationships require negotiation and compromise. Artistic taste develops through exposure and reflection rather than instantaneous optimization. Human beings do not flourish merely because every system surrounding them minimizes inconvenience.

A society organized entirely around frictionless optimization may ultimately become spiritually inert, even while remaining materially prosperous. And yet rejecting AI altogether would represent a profound mistake as well, because the expansive potential of these systems is immense and multifaceted. AI-enabled distribution infrastructure will allow, and does now allow, individuals with highly specific talents and interests to participate in and contribute to society in ways that previously would have been impossible. It can lower barriers to entrepreneurship and creative production. It can increase access to education and information and entertainment for communities that previously were entirely excluded from those pursuits. It can allow more individuals to pursue meaningful forms of differentiated striving. That last point is perhaps the most important.

The prosperous society described throughout this series is not one in which technology eliminates striving; it is one in which technology expands the number of people capable of striving meaningfully and productively and specifically. A low-friction economy does not need to become a low-agency society. In fact, if these systems are designed thoughtfully and competitively, they may ultimately produce the opposite outcome: a society in which more individuals can discover highly specific communities and products and forms of work aligned with their particular interests and talents. That is the moral choice embedded within the current AI transition. It is the inversion of Galbraith’s affluent society. AI has the capacity to expand the number of people capable of striving meaningfully and productively with purpose and intention, as a technologically buoyed open society that embraces the value of personal discretion and individuality and doesn’t surrender choice to technology merely in the name of convenience. It is a prosperous society. I’m Eric Seufert. Thank you for listening.