Why We Keep Misreading Disruption
What Google’s ‘Shop with AI’ Reveals About Our Blind Spots Around Technological Change
Google just released Shop with AI mode,1 and we can expect a week of hot takes declaring the death of e-commerce as we know it. But if past hype cycles are any guide, most of these predictions will age poorly—not because they overestimate the technology, but because they misread how disruption works.
This essay steps back to ask a different kind of question. Rather than evaluating the product, I want to explore what this moment reveals about the deeper shifts underway in how value is created and coordinated. We’ll begin by examining why our visions of the future so often go awry, then revisit a model of globalisation that explains structural change—and finally propose a reframing that might help us see where the system is starting to bend.
One reason we misread technological change is that we often conflate two very different kinds of progress. Incremental change happens within a paradigm: it makes systems faster, cheaper, or more efficient without altering their core logic. Punctuation, by contrast, signals a shift between paradigms—it restructures the system itself.
Most innovation hype mistakes the former for the latter. It imagines that the next app or AI feature will overturn everything, when in fact it often just rides the long arc of incremental improvement. True disruption happens when an innovation resolves a deep constraint in how value is created and coordinated—triggering a structural reorganisation.
In this light, the question isn’t “What does Shop with AI do?” but “What kind of shift does it imply?” To answer that, we need to step back from the product and examine the broader system it’s embedded in.
We can only understand what is going on around us now by viewing it as the bridge between past and future. This is a bit like what historians will tell us: to understand any phenomenon, you need to trace its origins and preconditions. A historian studying the French Revolution, for example, would naturally want to examine the financial crisis, social tensions, and political structures of the Old Regime. This is obviously recursive, as to understand the Old Regime you need to look into its past, and so on. It’s all too easy to use today as a lens to understand the past, but this results in us forcing concepts and framings onto the past that make no sense. If we want to learn the lessons of the French Revolution, then we need to understand it on its own terms, which means looking into the confluence of social, demographic, and natural trends that shaped it.
In many ways this is true of the future too—we use today as a lens to view the future, leading us to incorrect (and often hilarious) conclusions about want the future will bring. One of my amusements is to explore paleofuture, images created in the past which represent the future that we now live in. A famous example of this is a series of cigarette cards2 commissioned by a French toy maker in Lyon for the 1900 World Exhibition in Paris,3 under the title “En l’an 2020”, depicting what life might be like in the year 2020.4
The trap we fall into when predicting the future is that we’re fairly good at predicting within a possible future, but terrible at predicting which possible future will eventuate, and we don’t distinguish between the two. In De la Terre à la Lune, Jules Verne imagined reaching the moon via an artillery shell—a vision that stretched the confidence of 19th-century engineering. He was wrong about the method, but prescient in placing his launch site in Florida, near the equator—anticipating orbital dynamics before the physics was formalised.
This is why futurists often seem both brilliant and wrong at the same time. Given a range of possible futures, they’re terrible at predicting which future will eventuate. But if a future does eventuate, they (and most people, actually) are quite good at extrapolating within that future. Consider how many 1950s predictions about the year 2000 accurately anticipated some aspects of modern life (widespread automation, powerful computers) while completely missing others (the internet, mobile devices). They could predict within their assumed technological trajectory but not the fundamental shifts that would reshape everything.
Kurzweil’s predictions offer a fascinating case study, as his observations on exponential trends within the digital and computing paradigm correctly anticipated the internet’s growth, mobile computing, speech recognition, and aspects of AI development. However, his predictions outside this possible world are looking shaky. His predictions about life extension, for instance, assume biotechnology will develop primarily through computational approaches (genomics, AI-driven drug discovery, nanotechnology) have been less successful. Similarly, his predictions about AI achieving human-level intelligence (Artificial General Intelligence, AGI) assume current machine learning approaches will scale up. As I pointed out in Let generative AI be itself, not an imitation human,5 there are limits to a computational approach that suggest that AGI is beyond us in the current paradigm.
If we’re to understand the impact of LLMs, and solutions such as Shop with AI, then we need to escape the tyranny of now, step back, and consider if the technology—and the disruption assumed to be associated with it—will emerge in the current world, within the current paradigm, or in a future world which we struggle to imagine.
Back in The Great Unraveling6 we explored how we can approach this kind of structural shift as a constraint cascade. Richard Baldwin developed an elegant model of globalisation based on cascading constraints,7 which we built on to understand digitalisation. He weaves a narrative that helps us see how the world evolved from a pre-globalised state, through a first unbundling where a dramatic drop in trade costs enabled production and consumption to split, and then a second unbundling which ushered us into the current globalisation phase as low coordination costs enabled the creation of today’s global supply chains. Each of these unbundlings, these punctuations, represent a transition from one possible world to another. Between two punctuations we see incremental development, the sort of development easily predicted by Kurzweil et al. There is, however, little—if any—disruption within this incremental phase as disruption emerges when we cross a punctuation, such as when the development of the global multi-modal container network revolutionised cargo handling, disrupting the entire logistics value chain and ushering in a new wave of globalisation.8
When we ask if LLMs are disruptive, we’re really asking whether they represent gradual progress within our current technological paradigm, or whether they mark a shift to a new paradigm. So far they have been deployed within existing systems, which makes them firmly part of the long arc of incremental development. Shop with AI is interesting though, as it suggest that something deeper might be afoot.
One of the benefits (or problems, if you prefer) with a narrative approach, like constraint cascades, is that we’re interpreting the events we see when we build the narrative. There are always multiple possible narratives, and the narrative we land on is just the one which we think fits past events the best. Baldwin’s globalisation narrative, for example, suggested that the next phase of globalisation—the globotics era—would emerge when the cost of in-person interaction dropped, possibly via virtual and augmented relativity, telerobotics, and so on.9 While these technologies have emerged, the anticipated third unbundling has not.
Events have separated from the narrative creating ambiguity, and it’s reasonable to assume that Baldwin’s third unbundling won’t eventuate. But then, Shop with AI also doesn’t fit this narrative. Nor does how the dropping cost of manufacturing—with short runs becoming shorter and cheaper—is enabling finally assembly to migrate toward demand, toward the final customer. These inconsistencies suggest a deeper shift is underway—one not well captured by the old logic of supply-side constraints. Perhaps there is another narrative, a new one, emerging?
Baldwin’s model framed globalisation as a progressive unbundling of production. What if we extend this logic—not to the supply side, but to the demand side? What if the key shift now is not how things are made, but how intent is captured and fulfilled?
Baldwin’s model helps us see how supply-side constraints shaped past shifts. But what if the constraints shaping this next shift lie not in production—but in how demand is expressed, captured, and resolved? Globalisation isn’t just about lowering costs via offshoring. It’s increasingly about decoupling where capabilities exist from where demand arises—and using platforms and tools to match them in real time. Shop with AI’s features—personalised AI recommendations, streamlined discovery, and integrated checkout—decouple the act of shopping (demand expression) from traditional fixed pathways of supply. The shopper doesn’t browse a shelf or marketplace—they interact with an AI layer that assembles a solution in real time from distributed supply. That’s a vivid expression of demand-side unbundling.
The first two unbundlings in our new narrative are similar to the old: in the first producer and consumer separate; in the second, producer and production separate to create value chains. The third unbundling, however, is different. Rather than globalising production by lowering coordination costs, this shift globalises customer interaction by collapsing the friction between intent and fulfilment.
Our reframing proposes that the next leap won’t be about coordination but about compression—compressing the gap between intent and solution. This marks a more profound reconfiguration of value flows: from supply chain optimisation to interface-centric orchestration. Where Baldwin’s third unbundling was about transmitting capabilities to where demand exists, this new unbundling is about transmitting solutions, instantaneously and adaptively, to where intent emerges.
Shop with AI embodies a new affordance: the ability to dynamically assemble value propositions around real-time consumer needs, using distributed capabilities. It’s not just e-commerce. It’s ambient commerce—shopping that occurs as a side-effect of engaging with a context-aware, intent-responsive interface. This is a much broader use of the term ‘ambient commerce’ than we saw pre-pandemic, which focused on using sensors to streamline purchasing,10 as it involves a ”demand-side unbundling” where value crystallises closer to the point of intent, and solutions are assembled proactively, rather than requiring the consumer to actively “shop.”
There have been many predictions that the current global environment implies that we’re seeing the end of globalisation—that we’re stuck where we are, or we might even revert to an earlier more feudal state. Our new narrative, on the other hand, suggests that the next wave of globalisation might lie in front of us—that we’re at the leading edge of a new punctuation. LLMs are not, nor is AI in general, responsible for this new wave. It’s due to a confluence of social, demographic, industry, and technological trends, not a particular technology. Nor do we have a clear understanding of what’s on the other side of the punctuation, but we can see what will shape it.
First is the idea that value crystallises at the point of customer interaction, not in supply chains. This flips the traditional economic understanding of where value is created. It’s not just about designing good products, producing goods efficiently or delivering them cheaply. Instead, the ultimate leverage point is the interface where intent is captured and resolved. This suggests a re-prioritisation for businesses: invest heavily in understanding and orchestrating the customer interaction layer, even if it means de-emphasising direct ownership of physical assets further down the supply chain. This is profound for how companies should allocate capital and focus their strategic efforts.
This implies that the interface becomes the “new bottleneck” (or leverage point) when distribution wanes. If AI helps us abstracts away the need for humans to actively “search” or navigate platforms, then the competitive advantage shifts from owning the search engine (Google’s historical moat) to owning or orchestrating the agent-to-agent or agent-to-capability interaction. The new bottleneck isn’t finding information, but being the trusted agent or the accessible capability that an agent selects. This suggests a future where API design, semantic understanding of capabilities, and trust frameworks for agents become far more critical than traditional SEO.
It’s also important to note that this is a subtle but powerful evolution of Baldwin’s unbundlings—not a repudiation. We assume that globalisation means goods move around the world. The “drift of final assembly toward the customer” isn’t just about reducing shipping costs or time; it’s a response to the complexity of capturing intent and regulation. In a world of hyper-personalisation and diverse regulatory environments, it becomes more efficient to “finish” the product as close to the demand as possible, leveraging local capabilities and accounting for local nuances—much like how buildings are designed not just for purpose, but for place. This has implications for the shift to factory-driven (off-site and pre-fab) construction. This shifts the strategic importance from global mega-factories to agile, localised micro-fulfilment and customisation hubs. Value chains will by organised in layers of “Proximity to Intent” rather than “Distribution from Geography”, as this model colonises all layers of the production stack.
Consider online recruitment (Seek, Indeed, et al). These platforms serve primarily as digital noticeboards—broadly distributing listings to geographies in which talent was presumed to reside. But that logic is shifting. Increasingly, recruitment is being reorganised around proximity to intent: not just where candidates live, but what they’re signalling through behaviours, preferences, and capabilities. Platforms now use AI to infer intent, dynamically match candidates to roles. They might even reshape the role itself to fit emergent demand. This reframes the recruitment process, not as distributing job ads to a geographic market, but as assembling opportunity around intent in real time.
What we’re seeing, then, is a broader pattern: across domains, the system is reorganising around the moment where need crystallises. Whether in hiring, shopping, or manufacturing, the point of value creation is drifting toward the edge—toward where intention is formed and expressed. Globalisation doesn’t vanish—it recedes into the background, becoming infrastructure for a more fluid, intent-driven economy. The third unbundling, in this light, is not a continuation of the past trajectory but a reorientation of its axis—from distance to desire.
We also assume that platforms like Google and Amazon are powerful and extract value. The concept of a “distribution tax” highlights the hidden cost imposed by dominant platforms on merchants and users. Open ecosystems, by shifting power from ownership to orchestration, have the potential to democratise access to capabilities and customers. This could lead to a more fragmented, yet potentially more efficient and equitable, global commerce landscape where smaller players can more easily connect supply with demand without paying significant tolls to large aggregators. This is a potential anti-monopolistic force—particularly when we consider how Deepseek has shown that assumptions around the intensity of capital required for LLMs are likely wrong.
The final point to make is that this framework emphasises that true systemic “punctuation” (a significant, irreversible shift) doesn’t just happen because AI (or some other technology) exists. It happens when a “bundle” effectively relieves a significant tension and constraint within the existing system. This provides a diagnostic filter: not every AI application will lead to punctuation, only those that address deep-seated inefficiencies and limitations in how value is currently created and exchanged. This helps to separate hype from true transformative potential.
I’ve deliberately glided over many of the practical implications raised here—each deserving its own treatment. If recruitment is reorganising around “proximity to intent” rather than geography, what does this mean for labour markets or regional development? If the interface becomes the new bottleneck, how do we prevent this from simply reconstituting platform power in a new guise? These are deep and nuanced questions. Rather than rush them here, I’ll explore them more fully in future essays.
Perhaps the most important insight from Shop with AI is not about the feature set, or even the technology itself—but what it reveals about the phase shift we may be living through. This isn’t disruption as we’ve come to mythologise: a single product, company, or innovation overturning the status quo. Instead, it’s a deeper and more distributed reconfiguration—a shift in where and how value is generated, coordinated, and captured.
We misread disruption because we look for it in the wrong place: in the heroics of singular technologies, or in the early hype cycles that surround their release. But disruption, properly understood, is a structural transformation—a punctuation that reshapes the logic of the system. That’s why it often appears in hindsight, not foresight. And why it’s easier to see when we stop asking what the technology does and start asking what tension it resolves.
The lesson, then, is not to chase the next platform or paradigm, but to attend more carefully to where intent is crystallising, where frictions are compressing, and where the interface is becoming the new locus of value. That’s where the next unbundling is likely to take root. And that’s where the real shape of tomorrow’s global economy is already starting to emerge—not in the abstractions of AI, but in the subtle rewiring of how needs are met and solutions assembled.
If we want to understand the future, we shouldn’t ask what comes next. We should ask: where is the system starting to bend?
Mehta, Ivan. “Google Introduces Shop with AI Mode with Price Tracking, Agentic Checkout, and Virtual Try-On.” TechCrunch (blog), May 20, 2025. https://techcrunch.com/2025/05/20/google-adds-ai-powered-shopping-features-for-discovery-and-easy-check-out/.
Cigarette cards were small collectible cards that tobacco companies inserted into cigarette packs, primarily from the 1870s through the 1940s. They started as a practical solution, as the cards helped stiffen the soft cigarette packs and protect the cigarettes from being crushed. What began as functional packaging quickly evolved into a marketing tool. Companies realised people enjoyed collecting the cards, so they began printing them with colourful images and information on various topics. The subjects were incredibly diverse: sports figures, movie stars, military uniforms, flowers, birds, flags of different countries, historical events, or educational topics like “Wonders of the World.” The cards typically measured about 2.5 by 1.5 inches. The front usually featured an illustration or photograph, while the back might have explanatory text, statistics, or be part of a series that told a story when collected together.
Originally commissioned by Armand Gervais, a French toy manufacturer in Lyon, for the 1900 World exhibition in Paris, the first fifty of these paper cards were produced by Jean-Marc Côté, designed to be enclosed in cigarette boxes and, later, sent as postcards. See Dukes, Hunter. “Lost Futures: A 19th-Century Vision of the Year 2000.” The Public Domain Review (blog). Accessed May 22, 2025. https://publicdomainreview.org/collection/a-19th-century-vision-of-the-year-2000/.
Côté, Jean-Marc (18-19 ) Dessinateur présumé. “En l’an 2000 : A l’école - Correspondance Cinéma-Phono-Télégraphique - Parlez Au Concierge - Dictant Son Courrier - Automobiles de Guerre - Audition Du Journal - Une Curiosité - Chauffage Au Radium - Un Tailleur Dernier Genre : [Estampe] (3e Série) / [Jean Marc Côté].” Image. Gallica, 1910. https://gallica.bnf.fr/ark:/12148/btv1b52512335z.
Evans-Greenwood, Peter. “Let Generative AI Be Itself, Not an Imitation Human.” Substack newsletter. The Puzzle and Its Pieces (blog), March 3, 2025. https://thepuzzleanditspieces.substack.com/p/let-generative-ai-be-itself-not-an.
Evans-Greenwood, Peter. “The Great Unraveling.” Substack newsletter. The Puzzle and Its Pieces (blog), March 20, 2025. https://thepuzzleanditspieces.substack.com/p/the-great-unraveling.
Baldwin, Richard E. The Great Convergence: Information Technology and the New Globalization. Cambridge, Massachusetts: The Belknap Press of Harvard University Press, 2016.
Levinson, Marc. The Box: How the Shipping Container Made the World Smaller and the World Economy Bigger. Second Edition. Princeton: Princeton University Press, 2016. https://press.princeton.edu/books/paperback/9780691170817/the-box.
Baldwin, Richard E. The Globotics Upheaval: Globalization, Robotics, and the Future of Work. Oxford University Press, 2019.
Gorey, Colm. “What Is ‘Ambient Commerce’ and Why Is It Valued at $5.3bn?” Silicon Republic (blog), October 26, 2018. https://www.siliconrepublic.com/machines/ambient-commerce-iot-amazon-go.