But in consumer commerce, much of the buying journey has already been solved reasonably well by the browse-and-click paradigm. Comparing options was the last remaining bottleneck that ChatGPT now, often, solves better than the traditional listing search engine.

B2B procurement is different. Much of the buying journey has never been effectively digitized because the work cannot be reduced to browsing a catalogue and clicking through a predetermined route.

Agents are unusually well suited to overcoming the bottlenecks that have kept B2B procurement excessively manual, despite an abundance of (clunky) software.

It’s not just the buying journey that gets transformed, though. As agents get into comparing options and negotiating inside the workflow, the decision architecture across the firm changes as well.

Less obvious, but likely more consequential, are the many competitive games agents introduce between:

• The company deploying the agents and the providers providing them

• Adjacent workflows and decision makers across the firm rewiring its workflow

• Players ranging from corporate payments providers to systems of record as every player tries to go AI-enabled and launches its own procurement agents

• Negotiation games between buyer-side agents and seller-side agents

All of this opens a bunch of fascinating questions about how agents transform workflows, the nature of the firm, and the nature of competition in the ecosystem. This was very much the central argument of my book Reshuffle - that AI transforms not just how the task is performed but how entire economic systems are coordinated.

Today, I’m launching a new series of initiatives to track and map out how the ideas I first presented in Reshuffle are playing out across different domains.

I’m thrilled to announce a full redesign of the Reshuffle website, structured around the book and the new research to track how these ideas are playing out.

Explore the Agentic AI Observatory

Why I’m building the Agentic AI Observatory

Much of the current hype around agent autonomy looks for change in the wrong place.

We ask whether agents are autonomous, whether humans remain in the loop, or how many tasks have been automated. But these questions miss the more important changes that are already transforming the system.

An agent can determine which options are considered, how those options are negotiated to meet a workflow’s requirements (often the more consequential effect of autonomy), what information is eventually shown to an approver, and which actions are executed - while a person managing the role makes the final decision.

On paper, control has not moved. It’s time to celebrate the ‘human-in-the-loop’.

In practice, the path to the outcome may already have been redesigned.

This is what I hope to track with the Agentic Procurement Observatory - the first domain I examine in the Agentic AI Observatory. This will be followed by six other observatories - each on workflows that are likely to show unique structural effects owing to the deployment of agents.

Read the thesis on each domain here.

The case for agentic procurement

There are two factors that make B2B procurement a ripe candidate for agentic reshuffling.

First, B2B procurement is largely an exercise in constraint-based buying.

A purchase must satisfy a changing combination of requirements set by the buyer: technical specs, budget limits, policies, approved suppliers, contractual terms, delivery conditions, and other rules and preferences. These constraints often emerge gradually, conflict with one another, or vary from one transaction to the next.

Second, it is also an exercise in quote construction.

The buyer is not always choosing among finished products at posted prices. Buyers and suppliers frequently construct the offer together, engaging in a back-and-forth to configure scope, quantities, service levels, commercial terms, implementation obligations, and exceptions until a viable transaction emerges.

Traditional procurement software has struggled with both activities. It is generally designed to manage buying based on predefined forms and workflows. It works well when the options and rules are fully known in advance. It works much less well when the system must interpret an incomplete request, gather context, resolve competing constraints, coordinate several parties, or construct a transaction iteratively.

Activities have stayed manual not because there is something special about human involvement but only because they are contextual, negotiated, and constraint-heavy.

Agents are unusually well suited to this problem. They can work with unstructured information and discover missing requirements. They can also maintain context across interactions, reconcile constraints, generate alternatives, and coordinate the construction of a quote in real time. Instead of forcing every purchase into a fixed workflow, an agent can help assemble the required workflow around the purchase.

This makes procurement a domain in which agents may organize work in a way that the previous software paradigm simply could not. Agents may finally make previously unprogrammable parts of business buying programmable.

The Agentic Procurement Observatory

Agents can enter the buying journey at every stage, but their presence alone tells us very little about how power is changing.

The Observatory therefore looks at the second-order effects.

The Observatory follows agents through the buying system: where they have a structural advantage, where they are demonstrably acting, how far their actions reach, who makes the resulting outcome binding, whether organizations are redesigning around them, and what new competitive mechanisms may be forming.

Agents prepare, shape, constrain, execute, and sometimes enforce consequential parts of the workflow. A human may retain final approval while an agent determines which options are presented, which risks are surfaced, or which pre-existing authority is invoked during a negotiation. The final click may remain with a human, even as power over the choice increasingly shifts to the tool provider.

This dynamic is particularly visible in Ramp’s expansion from corporate cards into procurement.

Ramp enters the transaction from an adjacent control point: it already sees the request for spend, understands the company’s policies, controls access to payment, and records what was ultimately purchased.

From that position, it can move upstream across the procurement workflow and influence supplier selection before the payment ever occurs.

This points to a broader competitive shift that is also revealed through the data.

In an agentic market, companies do not need to attack an incumbent’s product directly. They can enter from a neighbouring workflow, use their existing data and control position to absorb adjacent decisions, and gradually redefine where the category begins and ends.

The visual below (interactive on the site) illustrates various entry points through which players enter the procurement worflow and expand their influence across it.

This lateral competition - competition between providers that originate in different categories but converge on the same consequential decision - is a hallmark of agentic competition.

A corporate-card company, procurement platform, expense-management provider, ERP, marketplace, or contract-management system may all approach the buying decision from different starting points. Once agents can carry context and coordinate actions across workflow boundaries, each can attempt to become the layer through which the transaction is interpreted, constructed, and controlled.

As a result, procurement software is no longer the well-defined Saas category that it used to be. It is about which provider can use its existing position to become the environment in which buying decisions are made.

The Observatory examines these dynamics through multiple connected lenses: agentic pressure across the buying journey, depth of work transformation, alternative workflow configurations, evidence of operating-model movement, emerging market mechanisms and scenarios for how power could shift.

These effects are analyzed at three levels:

The Workflow

The Firm

The Competitive Ecosystem

The Observatory distinguishes agents from configured rules and conventional automation. It separates observed actions from product claims. It records counterexamples and missing evidence alongside positive findings. And it traces each conclusion back to the concepts and theoretical foundations in the wider Reshuffle knowledge graph.

The Observatory is the first empirical expression of the Reshuffle knowledge graph, with many more to follow.

If you’re intrigued by this research, or would like to build custom research instruments like this one, feel free to write back.

And keep an eye out for more launches in the weeks ahead.

In this conversation, Rita McGrath and I - guided by the always-insightful Aidan McCullen - bring together complementary ideas from the books we have written and those we are currently working on.

It promises to be a treat: a wide-ranging exploration of what strategy becomes when the assumptions it rests on begin to shift.

Rethinking strategy

Our central argument is that viewing AI through the traditional mechanisation/automation lens traps us in looking for operational benefits (much of the discourse today), when the real advantage lies in understanding its impact on the rules of competition. It changes where scarcity lies, how industries are organized, and what constitutes a defensible advantage. These are ideas I’ve explored at length in Reshuffle.

One of the points I make, particularly in my next book Unfair Advantage (still in the works) is that we may also have over-indexed on modularity, mistaking a response to high coordination costs for an enduring law of business. As AI makes intelligence within individual modules abundant, advantage increasingly moves to those who can connect activities, learn across their boundaries, and improve the whole system together.

Vertical integration may therefore be returning, but in a new form. The prize is no longer necessarily ownership of every asset. It is ownership of the feedback loops that connect them.

Here are eight propositions that take the discussion beyond the obvious.

1. The strategic problem is no longer how to win the game, but whether the game still exists

Traditional strategy works something like this: identify an attractive position, build capabilities around it, and execute better than rivals.

AI introduces a more dangerous form of uncertainty. It does not merely change how efficiently firms perform existing activities; it changes which activities remain valuable, who can perform them, and where industry boundaries sit.

That means operational excellence can become strategically destructive.

A firm may become dramatically better at performing work whose economic value is disappearing. Winning the old game faster can simply accelerate movement toward irrelevance.

Traditional consulting firms illustrate the risk. Many are understandably investing in AI to make consultants faster at research, analysis, and deck creation. The deeper threat, however, is that clients can now perform portions of that work themselves while model providers absorb others. If consulting firms treat AI mainly as a productivity tool, they optimize the shrinking middle rather than asking what new work becomes valuable once production of knowledge work is no longer scarce.

2. Competitive advantage moves with scarcity, not with capability

Managers often identify advantage with something their company is particularly good at. But capabilities only generate rents while they remain scarce.

Once technology makes a capability widely accessible, excellence in that capability matters less economically even if the capability itself becomes more powerful.

Strategy should follow the constraint rather than celebrate the newly abundant capability.

Agencies, historically, charged partly because producing good-enough creative variations required expensive, skilled labor. Generative AI radically expands the supply of that output. An agency that responds only by producing more variants faster in the name of ‘innovation’ is competing harder in the activity whose scarcity is disappearing.

The only defensible position that remains is building a proprietary learning system that understands what makes a message work, integrating it into the customer’s commercial system, measuring outcomes, and understanding which components and choices influenced the output - and in what way.

3. AI-native advantage comes from redefining the unit of value, not adding AI to the old product

Incumbents commonly interpret a new technology as another delivery channel for an existing product. Native entrants instead ask whether the technology allows the product itself to be reorganised around a smaller or different unit.

That change in unit can reorganize workflows, interfaces, governance, and eventually the industry.

Adobe moved successfully from packaged software to subscriptions and cloud delivery, but much of its underlying logic remained file-centric. Figma treated individual design elements as reusable organizational objects. That sounds like a technical distinction, but economically it changed what could be governed, reused, shared, integrated, and coordinated across teams. Design became less about manipulating isolated files and more about managing an organizational design system.

Adobe therefore did not simply face a competitor with a better interface. It faced a competitor organized around a different unit of value.

An incumbent asking how to add AI to today’s object may miss an entrant asking whether AI makes today’s object obsolete as the natural unit around which work should be organized.

4. Horizontal strategies may move from standards adoption to AI-enabled translation

Digital platforms accumulated power by forcing participants into common interfaces - by establishing a standard and making everyone conform to it. The usual assumption is therefore that coordination requires either consensus around a shared standard or power strong enough to impose one.

AI weakens that assumption because it can reduce the cost of translation between incompatible systems. If machines can continuously interpret different schemas, documents, workflows, and domain languages, coordination can happen before everyone agrees on a standard. This creates a new entry strategy: instead of attacking an incumbent’s product directly, span the incompatible systems surrounding it and become the place where decisions are made.

Construction software illustrates the opportunity. Autodesk’s power partly rests on proprietary formats and workflows. An AI-native entrant does not necessarily need to replace Autodesk’s design tools. It can translate across Autodesk, Bentley, engineering systems, contractor documents, and project data to create a unified operational view. If users increasingly coordinate through that layer, the translator may eventually gain enough power to dictate interfaces back to the systems it once merely connected. Translation starts as interoperability and can end as governance.

5. Vertical integration returns when advantage lies in learning across the boundaries

For the past several decades, strategy rewarded modularization. Firms outsourced activities, standardized interfaces, moved infrastructure to the cloud, and concentrated on the layer where they believed they had an advantage. The implicit logic was that activities could be separated as long as interfaces between them were sufficiently stable.

AI changes the economics when performance depends on learning across those boundaries. If improving the customer outcome requires simultaneously adjusting product design, production, distribution, and demand sensing, then the valuable asset is not any individual layer. It is the proprietary feedback system that learns how the layers should change together. The coupling becomes an asset.

Shein is a useful illustration. Its advantage is not simply cheap manufacturing, superior fashion design, or better demand prediction. It connects social signals, design, small-batch production, market testing, supplier allocation, and replenishment into one rapid learning system. Each stage generates information that changes decisions elsewhere. A conventional retailer that outsources production and separately optimizes merchandising may possess many of the same components but not the same learning architecture.

The emerging form of integration is therefore not primarily about owning factories. It is about owning the feedback loops between activities.

6. In an AI economy, the by-product may become more valuable than the product

Most firms treat drafts, rejected options, failed experiments, intermediate reasoning, and abandoned designs as waste because the economic objective has historically been the finished output. But those discarded iterations contain information about why decisions were made, which alternatives failed, what trade-offs were considered, and how judgment evolved.

AI can turn that exhaust into reusable capital. Once intermediate work becomes machine-readable, the firm can search it, recombine it, identify conceptual gaps, train future workflows, and personalize outputs for different contexts. The finished product is only one bundle assembled from a much larger underlying knowledge base.

Publishing provides a strong example. A traditional publisher monetizes the finished book while most research notes, alternate structures, case material, rejected arguments, and conceptual connections disappear economically once publication occurs. An author who converts that material into a structured knowledge graph can produce multiple new narratives, detect gaps in the underlying theory, answer context-specific questions, or create entirely different products.

The strategic inversion is important: firms may discover that the asset was never merely what they shipped. It was the accumulated trail of learning generated while figuring out what to ship.

7. The machine that builds the machine becomes more important than the machine

When AI makes competent output widely available, selling a superior static output becomes harder to defend. Advantage moves toward the system that observes outcomes, learns from them, and modifies future production faster than rivals.

This changes the meaning of data advantage. The valuable data is not merely information about customers. It is measurement embedded throughout the production system: which design works with which manufacturing process, which recommendation performs under which conditions, which battery configuration works with which charging environment, or which intervention changes the desired outcome. The firm with the best measurement architecture can continuously improve every coupling that produces the result.

Tesla illustrates the logic better than a conventional car company. Its strategic system spans vehicle software, battery performance, charging infrastructure, driving behavior, service information, and continual software updates. The relevant advantage is therefore not reducible to producing a better car at a moment in time. It is the system through which information from usage feeds back into multiple parts of the offering.

As outputs commoditize, measurement becomes strategy because whoever defines what gets measured determines what the system learns to optimize.

8. In structural uncertainty, capabilities should be treated as options rather than execution assets

Classic capability building starts with strategy: decide where the firm will compete, then acquire the people, processes, and technologies required to execute that position. AI destabilizes that sequence because the external opportunity set and the internal capability set are changing simultaneously.

A new AI capability may suddenly make an adjacent market accessible. Employees experimenting with new tools may discover workflows executives never planned. A competitor from another industry may acquire capabilities that previously required years of specialized hiring. Capabilities therefore become less like fixed resources supporting a chosen strategy and more like options that reveal new strategies as they evolve.

Kodak provides a useful historical analogy. It is simplistic to say Kodak failed merely because it ignored digital photography; it invested substantially in digital technology. The deeper strategic problem was that digital capabilities were largely evaluated through the economics and identity of the existing photography business. The new capability was treated as something that should support the established game rather than as an option on fundamentally different ways of creating and capturing value.

AI makes this problem continuous. Strategy can no longer be an occasional choice followed by capability execution. Firms must repeatedly use emerging capabilities to test what new games have become possible.

Taking the Reshuffle thesis to the next level

Over the past year, I have been developing these ideas while applying the Reshuffle thesis across a wide range of industries. The work has involved asking where each industry is headed, how its structure is being reorganized, and what those shifts mean for strategy.

During that time, I have watched many of the thesis’s underlying hypotheses begin to play out in real time - often faster, and in more varied ways, than I initially anticipated.

Some of you will have explored the interactive companion to my book at reshufflebook.com. I am now expanding it from a companion to the book into a living platform for tracking the reshuffle as it unfolds.

As part of this expansion, I will be launching a series of industry observatories, indices, and maps that track how the thesis is playing out across different sectors. Alongside them, I will launch a cohort-based program in which I will launch a course based on my thesis, share my ongoing analysis as it develops, and explore what strategy is becoming in the age of AI.

The topics we will look at will include:

• Identify where value is moving,

• Which industry boundaries are weakening,

• What new bottlenecks are emerging, and

• How the sources of competitive advantage are changing.

I will make these announcements over the coming weeks. If you would like to get early access to the new research or participate in the program, please indicate your interest using the form below.

Sign up here

A network had twenty-four hours in a day, a handful of prime-time slots, and far more shows than it could have on air. The programmer’s job was therefore to decide what deserved one of those scarce slots, schedule it against competitors, and cancel quickly when the audience failed to appear. It was a case of exercising judgment under severe capacity constraints.

Streaming effectively removed the TV programming grid. There was no Thursday at 9 p.m. to win, no fixed shelf space to allocate, and far less reason to make every show appeal to a mass audience. Thousands of titles could coexist and viewers could consume them whenever they wanted.

The new game - The Streaming Reshuffle

Your average case study on Netflix will talk about the abundance of content and access, thanks to streaming.

However, removing programming-slot scarcity changed a lot more than distribution and access. It changed the structure of the industry behind the distribution pipeline.

Under broadcast scarcity, every show had to justify occupying one of a tiny number of slots. That pushed the industry toward broad appeal, predictable formats, recognizable stars, familiar genres, and concepts that could be explained quickly to advertisers and network executives. With shows competing to access a scarce scheduling slot, programming was much more conservative.

Once streaming removed the fixed grid, the viability threshold changed. A series no longer had to attract 15 million people on the same night. It could be valuable by attracting a smaller, highly specific audience over a longer period. That opened the door to narrower genres and serialized storytelling. It also allowed experimentation with darker themes, niche documentaries, and projects that would have looked too small or too strange for prime-time television.

It also changed who could be backed and transformed the careers of actors like Bobby Deol and Winona Ryder. In the old system, stars mattered partly because scarce slots made failure very expensive. If you only had a handful of bets, you wanted proven actors, proven showrunners, proven formats. Under abundance, platforms could make more bets and tolerate a longer tail of outcomes.

Broadcast television was organized around national markets because spectrum, advertising, scheduling, and distribution were national. Streaming made this market international. Squid Game, for instance, did not need to be designed as global television in order to become globally important.

Broadcast schedules rewarded self-contained episodes because viewers might miss one week, arrive halfway through a season, or encounter reruns out of order. On-demand viewing made long-form serialization much easier. Writers could assume continuity. Episodes no longer had to fit rigid 22- or 44-minute advertising structures. Seasons could be six episodes or ten rather than twenty-two. The unit of creative design could shift from the episode toward the season - or even the complete narrative arc.

It also changed how portfolios were managed. A broadcaster had to decide ex ante which few programs deserved scarce airtime. A streaming platform could commission a much larger portfolio and then learn from actual behavior: who starts a show, who finishes it, what audiences overlap, which markets respond, what gets rewatched, what drives retention. Programming therefore moves from something closer to editorial capital allocation under scarcity toward portfolio experimentation under abundance.

This, in turn, changed the power structure around talent. Under the old model, the network controlled the scarce distribution slot, so creators needed access to the network. Under streaming, platforms initially created much greater demand for producers, actors, writers, and showrunners because they needed huge volumes of differentiated content. Scarcity temporarily migrated toward high-quality creative talent and distinctive IP. That is one reason the streaming wars produced extraordinary bidding for creators and content libraries.

This is why technological transitions are so often misread. We ask how the new technology will improve the existing game: how streaming improves content access for consumers, how the internet improves retail, how AI improves knowledge work. We focus only on the job to be done.

The more important question is what happens after the old constraint disappears.

Because strategy is ultimately organized around scarcity. And when technology changes what is scarce, it changes where power, advantage and value can accumulate.

Hollywood initially misread streaming as a new distribution channel for the same old product. What it missed was that streaming removed the scarcity of schedule, which changed the game itself.

Many companies are making the same mistake with AI today. They are treating it as a productivity layer for the existing firm, while missing that the real game shifts elsewhere.

Scarcity and Strategy

One of the central ideas in my book Reshuffle is the relationship between scarcity and strategy.

Every generation of strategy is built around managing what is scarce.

When productive capacity was scarce, advantage came from owning productive assets. When distribution was scarce, advantage came from controlling channels. When attention and market coordination became scarce, platforms won by aggregating participants and setting the rules of exchange.

Reshuffle argues that AI is now removing scarcities around the production of usable knowledge work.

That is why so much of today’s strategy debate feels slightly off. We keep asking whether AI will automate jobs or lower costs - but these questions assume that companies will continue to play the old game.

What matters more is what becomes scarce once usable knowledge work can be accessed cheaply.

(Note that I constantly use the phrase ‘usable knowledge work’, not intelligence. The latter, as I call out in the introduction to Reshuffle, is a distraction unless it delivers the former.)

The history of strategy can be read as a history of migrating bottlenecks.

And the companies that get disrupted are often not the ones that fail to adopt the new technology. They are the ones that continue optimizing a game structured around managing a previously scarce asset that has suddenly become abundant.

Misdiagnosing scarcity

Even once we decide that we need to move to the new scarcity, it is easy to misdiagnose what’s really changing.

It’s easy to claim today that AI will make knowledge work abundant. But knowledge work is too broad a category, just as content was too broad a category for understanding streaming.

Much of the knowledge economy has been organized around the high cost of exploration. Serious analysis required scarce expert time, so organizations had to ration which questions deserved investigation before they knew which questions would prove valuable. Managers considered a few hypotheses, few alternatives, and few courses of action.

AI begins to remove that scarcity. It makes it possible to generate, simulate and recombine vastly more candidate explanations and courses of action before committing significant resources to any one of them.

This changes the game in three vastly different ways.

Direction 1: Judgment becomes more valuable because selection becomes the bottleneck

One view suggests that if AI can generate 100 plausible strategies instead of five, the strategist’s value moves away from producing options and toward deciding which ones deserve belief and commitment.

On this view, AI commoditizes production but strengthens the value of judgment. The scarce resource becomes the ability to distinguish signal from seemingly plausible noise.

The problem with this argument is that it traps us in our old arguments of human exceptionalism. If AI can generate alternatives, it can also rank them, criticize them, run adversarial checks, compare them against evidence, and learn which recommendations worked.

What looks like a permanent human bottleneck may merely be the next temporary inefficiency.

And organizations banking on this would simply be banking on an inefficiency arbitrage much like the prompt engineers of 2024 did.

Direction 2: Judgment becomes less important because experimentation replaces prediction

When exploration was expensive, organizations had to make strong ex ante judgments. An executive had to decide which three product concepts deserved testing because testing thirty was impossible.

As exploration becomes cheap, the organization can postpone judgment. It can generate many possibilities, test them cheaply, observe what happens, and progressively allocate resources toward those that survive.

The game therefore moves from upfront prediction to emergence through experimentation.

This is close to what happened when television moved from scarce programming slots to streaming. Network executives once had to decide in advance which few shows deserved scarce airtime. Once distribution became abundant, more bets could be placed and audience behavior could do more of the selecting.

Applied to knowledge work, the implication is unsettling. The future strategist may not be the person with the best judgment about which answer is right. The advantage may belong to the organization that can structure the cheapest and fastest learning process for discovering what is right - even if they entered your industry with much lower expertise.

If the cost of experimentation collapses sufficiently, the value of judgment collapses with it.

Of course, much of business cannot simply be A/B tested. Acquisitions, reputation, organizational restructurings and geopolitical bets often involve irreversible commitments. But wherever experimentation can reduce uncertainty, it allows the creation of a fundamentally different form of competitor.

Direction 3: The real scarcity becomes framing because abundance expands the search space faster than our ability to navigate it

There is a third possibility.

Cheap exploration does not necessarily make decision-making easier. It may make the search space explode.

When five hypotheses can be examined, deciding among them is manageable. When five thousand can be generated, the problem becomes defining which possibilities belong in the search space at all.

The scarce capability therefore moves further upstream, from producing answers to constructing the framework within which answers are generated.

What problem are we actually solving? Which variables belong in the model? Which causal relationships matter? Which evidence would change our mind? Which outcomes are we optimizing for? Which constraints are real and which are inherited from the old system?

AI can generate thousands of answers to the wrong question at negligible cost. Abundance therefore increases the value of a good lens.

The scarce asset is the learning architecture

These three arguments look contradictory because each places scarcity at a different point.

The first says scarcity moves downstream into judgment.

The second says judgment itself can be displaced by experimentation.

The third says scarcity moves upstream into framing.

The resolution is that none of those activities is independently the new source of advantage.

The scarce capability is the architecture that connects them.

An AI-native organization needs to frame the right search space, generate many possibilities within it, cheaply eliminate weak ones, test the promising ones against reality, determine where irreversible commitment is justified and then feed the results back into the next round of search.

The advantage is created through a superior learning architecture spanning possibility to evidence to commitment.

Strategy in a world of expensive exploration is largely the discipline of making good choices with limited analysis.

If you’ve enjoyed reading this post, dig furtehr into the ideas in Reshuffle:

Check it out

You’re right to be skeptical about every new ‘AI expert’ shouting out from the rooftops on LinkedIn.

But what about peer-reviewed AI research coming out from credible labs and top universities?

As it turns out, research can be right in the narrow sense and still misleading in the larger sense.

This, of course, applies to everyone writing about AI - including anything I put out there.

This post is about making sense of AI research - looking for gaps in a theory and bridging those with complementary bodies of research.

And I’m launching something super-interesting with this post - keep scrolling through the argument to learn more.

The three types of AI research out there

While researching for Reshuffle, I looked through the research on AI and came across three types of issues in the research:

1. Type 1: Rigorous research but from a different era - caveated as such by the researchers themselves but over-extended by enthusiastic proponents to make claims about how it applies to AI today.

   1. Possibly, the most glaring example is what I called out as Jevons Misunderstanding in my previous post, where Jevons Paradox is quoted without questioning where it applies and where it does not.

   2. The ALM hypothesis is another such example which was drafted in a pre-AI era when many of the activities that machines can do today were simply impossible for machines to extend to.

2. Type 2: Intellectually rigorous research but yielding narrow findings that are generalized far more than they should be. In many ways, these are the most difficult to identify correctly.

3. Type 3: Agenda-based research - The researcher or their sponsor needs to push an agenda and uses data to make that point. These were easy to weed out right at the beginning. A lot of this research was coming out of the BigTech and frontier labs in partnership with universities.

For a start, let’s ignore studies that are driven by an agenda (Type 3). As for Type 1 research, the researchers categorically state assumptions and limitations and yet proponents extend the findings and use them to bludgeon alternate views into submission, without looking at the assumptions backing those findings.

Type 2 is the most interesting category.

Even when AI productivity studies are intellectually rigorous, they still repeatedly make four assumptions about the impact of AI on work that are no longer true.

Four assumptions that intellectually rigorous studies on AI get wrong

There are four things that Type 2 studies often get wrong. With these fundamental misframings, the studies - no matter how rigorous - reveal results that cannot easily be generalized.

1. They assume a fixed capability frontier:

A productivity study estimates an effect at time T. AI capability is non-stationary; it changes every few weeks. So a point-in-time productivity study is partly obsolete by the time it’s published. Prior tools used to be static long enough to be measured. That’s no longer true.

2. They assume that value doesn’t migrate and hence productivity gains lead to value capture at the same point as yesterday:

AI’s productivity gains are often captured not by the firm or worker adopting it but upstream by the tool provider, or by the players ‘above the algo’. So the surplus migrates out of the frame that the productivity gains study is pointed at. This is why worker-level productivity can rise with no wage capture, and why the “AI boom will show up in the statistics eventually, like the Solow paradox resolved for computers” expectation is misplaced. The gain isn’t late to arrive. It’s just accruing at a different altitude, and firm-level metrics cannot count it.

3. They treat adoption as a one-time deployment:

In previous tech shifts, you could give the tool to workers and measure the lift. But AI adoption is continuous capability absorption, with tools, workflows, and practices co-evolving, plus an ongoing ‘human-in-the-loop’ bill to be paid for verification, evaluation, and re-tuning. Deployment is continuous so before-after studies assuming one-time deployment are not even structured correctly. There is no static ‘after’ destination.

4. They assume the unit of work stays the same:

A study measures a task or a role. But foundation models cut across functional boundaries and unbundle the role into fragments that get rebundled in other roles, teams, and even organizations (e.g. a task you performed could now be performed by your customer). The study measures a certain role or a certain org, not accounting for the fact that the work has moved to an entirely different location.

Every study has its merits and limitations. In writing Chapter 4 of Reshuffle, I had to work through all these studies mapping what each got right and what they didn’t.

I am launching that today as a structured interactive map of research in this space.

Here’s a full list of links to the resources:

1. The core framework - Why AI is different from previous automation technologies and how that should be factored in while analyzing different theories on AI.

2. An interactive map of the eleven schools of thought: The index draws on eleven established research traditions, not a single point of view - this breadth is the first test of rigor.

3. What is different about AI: Tests where each theory breaks with AI and explains how to fix it.

4. The contribution matrix: Shows exactly what each school contributes and where it falls short - no theory is over-claimed and every gap is explicitly identified.

5. The false binaries: Explains why we see superficial framings - automation vs augmentation, hard vs soft skills, generalist vs specialist - and why those framings fail.

Mapping and reconciling the different schools of thought

Instead of picking a single school, I took eleven of them - across labour economics, modularity theory, value migration, the ghost-work literature, organizational design, and the political economy of work - and treated each as a partial view of one phenomenon rather than a rival account of the whole.

The wager is that each of these traditions is right about something and blind to something else. There’s a method to the blindness - it’s not random.

View the interactive map

However, to compare these schools while accounting for the 4 problems mentioned above, these studies need to be placed on a broader causal chain.

To make the different schools comparable, I map them onto a single chain: capability, substitution, interface, bundle, capture.

First, AI changes the capability frontier. It changes what machines can perceive, generate, reason about, coordinate, or execute. But because this frontier is non-stationary, the capability layer cannot be treated as a fixed input.

Second, capability change creates possibilities for substitution or augmentation. Some activities can be automated, some can be assisted, and some become cheaper, faster, or more widely available. This is the layer most productivity studies observe most directly but also the layer they stop at.

Third, substitution and augmentation change the interface structure of work. Once AI can perform parts of a tacit workflow, new boundaries appear in the workflow.

Fourth, this changes the bundle. Activities that were previously fused inside a human role can now be separated and recombined across roles, workflows, and teams - even orgs. The job, role, workflow, or organizational unit no longer has to be held together in the same way. Work can be unbundled from one role and rebundled into another. It can move from employee to customer, from firm to platform, from team to agent, from professional service to software layer, or from human execution to human verification.

Fifth, the new bundle changes capture. Value does not necessarily accrue where productivity was measured. It may migrate to the owner of the model, the owner of the interface, the owner of the workflow, the owner of distribution, the owner of trust, or the actor that controls the new point of coordination.

See the central framework

Where Type 2 studies go wrong

Type 2 productivity studies usually miss all this.

They often see automation vs augmentation. But they often stop before interface, bundle, and capture. As a result, they may correctly measure a local effect while missing the system-level reshuffle that determines whether the effect generalizes.

This is why a study can be rigorous and still be strategically incomplete.

It may correctly show that AI improves a worker’s productivity on a measured task. But it may not tell us whether the capability frontier will hold, whether the workflow will be redesigned, whether the task will remain inside that role, whether the surplus will accrue to the worker or firm, or whether the entire bundle will migrate to another point in the ecosystem.

The Reshuffle thesis, therefore, does not reject productivity studies. It works across them, using them as evidence at one point in a larger chain, not as a full theory of AI’s impact on jobs. They tell us what AI can do to a task under observed conditions. They do not, on their own, tell us how work will be unbundled, where value will migrate, or how jobs, firms, and ecosystems will be rebundled around the new capability frontier.

What each study contributes

The real question

The real question is not only: what does AI do to this task?

It is:

What does AI make newly capable?
What does that substitute or augment?
What new interfaces does it create?
What work bundles does it unbundle and rebundle?
Where does value finally get captured?

Until we answer the full chain, we are not really studying the impact of AI on jobs. We are studying a local productivity effect inside a system that may already be reshuffling around it.

The Reshuffle Index - in beta

This research forms the bedrock of the Reshuffle Index, currently in beta with clients, and launching later this year to the market. The index tracks the extent to which different industries, jobs, and organization types are getting reshuffled because of AI.

The index is currently live in beta with clients, supporting their internal strategy and will be launching as a market-facing index later this year in partnership with other firms.

If you’d like to learn more about the index, just reply to this email and get in touch.

Once more - here’s the full list of links to the resources:

1. The core framework - Why AI is different from previous automation technologies and how that should be factored in while analyzing different theories on AI.

2. An interactive map of the eleven schools of thought: The index draws on eleven established research traditions, not a single point of view - this breadth is the first test of rigor.

3. What is different about AI: Tests where each theory breaks with AI and explains how to fix it.

4. The contribution matrix: Shows exactly what each school contributes and where it falls short - no theory is over-claimed and every gap is explicitly identified.

5. The false binaries: Explains why we see superficial framings - automation vs augmentation, hard vs soft skills, generalist vs specialist - and why those framings fail.

There are many convenient explanations people reach for when they want to make sense of AI. One of them is Jevons Paradox.

As the AI doomers began warning that AI would take away all jobs, the other side reached for Jevons:

• Efficiency expands demand;

• Cheaper work creates more work;

• Therefore, workers will be fine.

The problem with this binary debate is that it leaves no room for the actual economics. You are forced into one of two camps: AI destroys jobs, or AI creates more work.

But the real question is neither whether the market grows nor whether work disappears. The real question is who benefits from the demand expansion.

That is the Jevons Misunderstanding!

The Jevons Misunderstanding

The Jevons Misunderstanding - applying Jevons Paradox as a counter to all AI doomerism - mistakes augmentation for complementarity.

Augmentation - whether AI helps the worker perform the task - is irrelevant.

What matters is whether AI increases the worker’s claim on the new surplus generated.

The Jevons Misunderstanding assumes the old production system continues to scale when demand expands as a result of the new technology.

That was more plausible in earlier industrial cases where efficiency made the existing resource economy larger.

AI is different because it can re-architect the production system itself. It:

1. Separates the worker from the customer,

2. Turns the augmented worker’s continued expertise into raw material for training,

3. Codifies craft by progressively unbundling previously tacit knowledge into components that can be absorbed by the tool.

Expansion no longer has to pass through the old labor bundle. It can bypass it - meaning the worker stays employed but at progressively lower wages in worse conditions, not seeing the benefits of the market expansion.

That is what the Jevons debate misses. The issue is not whether AI expands demand. It often will. The issue is whether workers sit above the algorithm, using AI as leverage, or below the algorithm, feeding the system that captures the value.

An alternate format

Instead of provising the persuasive argument alone, as I’d do in most newsletters, I’m going to take another path here.

Trace the argument, but you’re also invited to play around with some data around the argument.

For instance, this divergence in demand and wages across 6 jobs that shows Jevons Misunderstanding play out at scale:

But instead of sharing it on the newsletter, I’m going to do one better.

You can play with the data and the argument at this deep-dive on Jevons Misunderstanding that launches today.

Open Deep-Dive

The argument architecture revisited

The Jevons Misunderstanding page is an example of a different publishing architecture.

The written word as a default mode of communication is rapidly getting commoditized. That does not mean all writing becomes worthless. It means the purpose of writing gets unbundled.

The old newsletter version of this idea would be a linear essay. It would explain the analogy, cite a few sources, and end with a conclusion.

The new version uses prose to frame the argument, but adds the cases - in a way that you can play around with them - to illustrate the argument.

The page shows the popular claim, the historical analogy, the counterfactual, the role-level details, the capture stack, the mechanisms, the falsifiers, and the positions through which workers might actually benefit from the expansion.

The reader is invited to inspect the architecture of the claim.

This is my attempt to disrupt this newsletter and move to an AI-native format.

Over the past few months, I've been building several indices to prove out the Reshuffle thesis. While there's a treasure-trove of data now loaded onto the underlying indices, this analysis, titled Jevons Misunderstanding, launching today, is the first application of the index that I'm putting out there.

This page is built on the index. And as the index continues to be trained on a monthly basis, all pages built on it refresh automatically with new data, providing a constantly updated dashboard on these ideas. As and when the data flips to contradict the older read of the page, a periodically scheduled agent brings that up for review - even as the pages that grow around the index keep increasing.

Here’s the page again.

I’d love to hear your thoughts and reactions.

And there’s a lot more where this came from!

The old newsletter was the argument.

The new newsletter is the interface to the argument system.

I started noodling with a question towards the end of last year - what should work look like now that producing work is no longer a constraint?

More broadly, what should AI-native work look like? And how can all of us - who’ve built careers around producing hard-to-produce outputs reimagine our work in an age when producing outputs isn’t all that hard anymore?

But before we get into all that, I’d like to start by launching something I’ve been working on over the past few weeks.

Launching: The living companion to Reshuffle

Many of you have enjoyed reading my book Reshuffle.

I’m delighted to announce the launch of the companion guide to Reshuffle.

Visit the companion guide

In the book, I had mentioned that I would be launching a companion guide to Reshuffle in early 2026.

When I started working on it, I originally imagined it as another book. But as I went further down that path, I realized that a book was not necessarily the right format for what I wanted to build.

Books remain an extraordinary medium for communicating a big idea in a persuasive, structured way. Reshuffle itself was written for that purpose. But in the age of AI, many ideas need to be encountered differently. Some need to be explored, queried, recombined, navigated, and applied. They need interaction, not just exposition.

The book - I believe - needs to be unbundled and rebundled.

That realization led me to rethink the companion guide. Instead of turning it into a static book, I have launched it as an interactive artifact — a living companion to Reshuffle.

For now, it has three core components, and I expect it to keep expanding over time.

First, and most important, you can work through the ideas in the book using the Reshuffle Map that maps out the core concepts and how they connect to each other.

You can explore specific threads/paths on the map. For instance, the coordination thread:

Or you can work through one of the four pathways through the book - which talks about AI’s impact at all 4 levels - jobs, orgs, ecosystems, and the larger economy.

And finally, if you’re just looking to scroll around, you can work through the central scrolly-telling narrative:

Have fun poking around the site, and let me know if you’d like to see anything specific in the comments below.

The AI-native book/newsletter:

Unbundling and rebundling a book/newsletter

This brings us back to a central question:

How should we reimagine our work in the age of AI?

For most of the knowledge economy, we have evaluated knowledge work through its outputs. We judged the report, the spreadsheet, the program, the design, the presentation, the book, the newsletter. These artifacts became the visible proof of the work.

But that created a subtle distortion. We began to overvalue the artifact itself and undervalue the problem the artifact was meant to solve. A book, for example, is valuable because it helps a reader understand, believe, remember, apply, or share an idea. A newsletter is valuable because it creates rhythm, attention, and connection around a set of ideas.

AI exposes this distinction because it makes artifacts easier to produce, with far less friction. When output becomes abundant, a polished artifact isn’t as valuable anymore.

What’s more valuable is understanding what job needs to be done and what form the idea should take to do that job well.

That shift changes the nature of knowledge work. Value moves from producing a single finished output to designing a system through which ideas can be structured, recombined, and expressed in different forms for different purposes. The same underlying ideas may need to become a book chapter, a map, a visual, a decision tool, or an interactive companion.

So the real work is no longer artifact production. What matters more is idea architecture. It is understanding the job an idea needs to perform, unbundling it into reusable components, and building mechanisms that allow those components to be rebundled towards the right output for the right context.

In a world where outputs can be created on demand, the scarce skill is not merely making the output. It is knowing what the output is for and owning the unbundling and rebundling logic to get it there.

Let’s take the Reshuffle companion. It is, by no means, a fully evolved solution, but it does two things that you’ll instantly see.

It unbundles the book into its core component ideas.

It then allows different forms of rebundling to bring those ideas back to life in new ways.

The map is one example of the rebundling.

The pathways are another.

Once you look thorugh those pages, you’ll see that both of them are different bundles of the same underlying components.

The book, for that matter, is itself a specific bundle of those ideas, tied together with a specific narrative style to aim a certain transformation for the reader.

Now, once the reader has read the book, it may no longer serve as the optimal artefact for what the reader next needs i.e. reference the ideas, apply them, build on them etc.

This is where creating a reusable idea architecutre - by unbundling and rebundling the book - helps.

Reimagining the unit of work in the age of AI

What does it actually mean to produce AI-native work?

Most people answer that question by reaching for AI as a tool. They use it to summarize a book, turn a keynote into a carousel, the carousel into a thread, the thread into a podcast, the podcast back into a blog post.

They get faster. They get into more channels. They feel productive.

They are doing exactly what Adobe did when the cloud arrived: same file, new pipe.

This is repurposing. It feels like transformation but it changes nothing structural. It is still built on the logic of the output, much like Adobe was built on the logic of the file.

Figma did not put the design file on the cloud like Adobe. Figma killed the logic of the file. From the article:

It replaced the file with the element - a button, icon, or type style - as the basic unit of work.

Because of the element-based architecture, Figma users could create shared libraries of reusable design components, like buttons, icons, type styles, and color palettes, that teams could use across multiple files and projects. Instead of duplicating these elements in each file, designers simply reference a single source of truth.

The ‘file’ was now a specific rebundling of these elements, not a siloed disconnected object.

Changes and permissions could be tracked and managed at the level of a design element. Each element was addressable in a database: change a component once and that change propagated everywhere it appeared.

This creates consistency, simplifies updates (change once, update everywhere), and enables cross-functional teams to work with aligned visual standards. Shared libraries shift design from isolated file ownership to coordinated, system-level collaboration.

By shifting the unit of work from file to element, Figma enabled real‑time collaboration, created a shared design environment that expanded who could participate, and made Adobe’s model feel increasingly constrained by its own architecture.

The Figma move and the Adobe move sound similar from a distance. They are opposites. One moves the same frozen artifact through new channels. The other unbundles the artifact and lets new artifacts emerge from new rebundlings.

Adobe’s move is a distribution story. Figma’s move is an architecture story.

Why the written form was always a workaround

Every output we produce - the book, the report, the deck, the course - is a bundle. Inside any one of them: frameworks, claims, examples, evidence, counterarguments, definitions, stories, decisions. What fuses the bundle is the narrative - a single chosen ordering of those ideas.

Ideas in your head are associative. You hold one concept and a dozen related ones hang off it at once - examples, exceptions, counterweights, applications. No writing technology has ever been able to transmit that directly. So the author compresses an associative web of related ideas into a single chosen line, discarding every other ordering that could have worked, because there was nowhere to put the leftovers. Then the reader decompresses, rebuilding the associations in their head from the one path they were given.

Once you can see the linear artifact - whether a book or a spreadsheet or a product design doc - as a property of the medium rather than of the thought, you have a path to making the Adobe-to-Figma move.

Unbundling, then rebundling

There are two distinct moves here.

Unbundling is the recognition that the output is not the unit. You stop thinking of the book (or any artefacts/output) as a thing to be sliced into formats and start thinking of the ideas inside it as items that have their own life. You stop asking “how do I get more out of the book” and start asking “what are the actual ideas I have, and what could each of them be in their own right.”

This is not something as pedestrian as content repurposing, which is the Adobe move - still taking the output as the unit of work and repurposing it for a new channel.

Rebundling is what becomes possible once you’ve unbundled: combinations the original sequence could never have expressed. New compositions for new contexts, new audiences, new questions. The expression of each idea is now a function of the context it is being delivered into.

Repurposing leaves the bundle intact and changes the wrapper. Unbundling and rebundling changes the bundle itself — and once you can do that, you can do it again, and again, in response to questions the original output never anticipated.

This is not artifact repurposing. It is interrelationship-between-the-ideas repurposing. The thing you reuse is no longer the output; it is the connective tissue between the ideas, and yesterday’s output was only one way of packaging that.

Super-charging the ideas economy

This has massive implications that extend far beyond what you can do with a book, a newsletter, or any artefact you create.

Let’s take the example of the book again.

Since back in 2012, when I first began taking my research and writing seriously, I have documented almost every idea I came across. This was largely for my own research discipline - to put everything in one place and find connections between ideas that, at first glance, seemed unrelated.

For most of that time, the process was entirely manual. I used Workflowy as my main knowledge base. There were more sophisticated mapping tools available, but I found most of them too complex for daily use. Workflowy worked because it looked and felt close enough to Microsoft Word, while still giving me the flexibility to move entire chunks of thought into new hierarchies.

Here’s what a segment of my Workflowy looks like:

At an average of 25-30 ideas added everyday across an average of 300 days a year for the past ~15 years, this has been built out to a corpus of more than 100,000 ideas that have been documented over this time. I say more than 100,000 becuase on days when I really get lost into a topic, I easily add 100-150 components in there.

Yet, only a tiny sliver of all this corpus made it through in books and newsletters.

More importantly, my ability to identify combinations and causalities across these ideas was constrained not just by my cognitive bandwidth but more importantly by the growing manual effort in recategorizing ideas and restructuring hierarchies as I added more ideas here.

Here’s how I’ve managed it over the past 15 years:

Every day, I would add what I was reading, along with the relevant links, into Workflowy. Every so often, I would go back and clean it up. And every December, I would take the last 2 weeks to review the year’s notes, reorganize them, and merge them with the ideas I had collected in previous years. The fact that I was usually at a beach resort with nothing to do but marinate in all the ideas through the year made it an annual ritual I’d look forward to.

That system served me extremely well. It helped me write three books. More importantly, it helped me keep seeing connections between everything I was reading, studying, and thinking about.

But I now realize that a knowledge base like this becomes valuable only if it can be activated. If it simply sits there as a large archive, its usefulness is limited.

This is where GenAI becomes especially interesting.

Because it can interpret language, identify concepts, and make connections between them, it can help turn a largely tacit body of work into something more modular, searchable, and generative.

Over the years, every post, essay, newsletter, and book chapter I have written has essentially been a narrative effort to connect ideas that had already made it to my Workflowy. Much of that work has happened tacitly - as it does for most knowledge workers. But with GenAI, there is now an opportunity to make those underlying concepts, connections, and patterns more explicit.

In many ways, this is an extension of what I have been thinking about since I began writing Reshuffle: the componentization of knowledge, the modularization of ideas, and the falling cost of translation between previously disconnected silos.

I have long been interested in how these shifts apply not only to industries and organizations, but also to the world of ideas, concepts, and theories.

Over the last several weeks/months, I have been experimenting with a range of tools, primarily Claude Code, but also a few tagging and taxonomy management systems, to organize and restructure my knowledge base. The larger ambition is not simply to create a better archive, but to build a living system that can identify connections, recombine ideas, and help me work across the full depth of what I have collected over more than a decade.

And in that process, I’d like to keep rebundling these concepts into fundamentally new bundles for readers and consumers of these ideas.

Narrative formats like the book and newsletter will continue to play a role where they add the most value. But there’s a lot more that can be achieved with the idea architecture in place.

What’s next!

If you’ve found this approach compelling, I’d love to hear how you’re applying it in your work (or intend to apply it).

For now, the next ‘rebundling’ that I’m working on is a jobs index - a product that takes data from the market and applies the Reshuffle thesis to it to determine the extent to which specific jobs are being reshuffled.

The first index will launch next week - stay tuned for more!

And once again, don’t forget to give this a spin…

Visit the companion guide

The US is betting on intelligence. China is betting somewhere else!

The US frequently frames its competition with China as an AI race, similar to the space race it ran with the USSR. The idea of a race was triggered around a year ago with the launch of DeepSeek. Ever since, much of the US media has been fascinated with the idea of the US winning the AI race against China.

Ironically, it isn’t much of a race if the US is the only one running it.

The American ‘AI race’ is framed around who will build the smartest models, as if superior intelligence alone decides the future.

China’s strategy reveals a different understanding of the game altogether.

It is not trying to win AI at all. It is betting that intelligence will become abundant, and that power will flow instead to whoever can reliably turn intelligence into economic value.

Great! Let’s dig in…

The US is building AI, China is preparing for the Reshuffle

The most consequential strategic bet China is making in the AI era is on coordination - and a very specific type of coordination -

The ability to reliably turn intelligence into energy-backed execution at scale.

China’s thesis is best understood as the following:

Economic value is created through the complementary effects of four factors:

• Energy creates the capacity to perform work.

• Intelligence creates the capacity to decide what work should be done, when, and how.

• Execution creates the capacity to perform task-level work.

• Coordination creates the capacity to align many task-level executions across time, space, and actors towards larger system-level outcomes.

None of these substitutes for the others.

Power accrues to whoever controls the binding constraint among them.

A tale of two bets

The US is betting that intelligence is that constraint. Their dominant assumption is that superior AI models - better prediction, reasoning, and abstraction - will sit atop a global system where energy, execution, and coordination can be purchased through markets.

Intelligence is treated as scarce and rent-bearing; everything else is assumed to be modular, interchangeable, and cheap.

This is why American AI strategy gravitates toward closed-weight models and proprietary control at the intelligence layer.

China is making a different, deeper bet.

It assumes that intelligence isn’t scarce, and treats coordination as the true locus of advantage.

Open-weight models are not a concession to hardware constraints; they are a deliberate move to commoditize intelligence so that value shifts to the layers that China controls today - energy infrastructure, execution capacity, and the ability to coordinate.

China has been steadily building dominance across these layers that turn energy into coordination - batteries, motors, power electronics, embedded compute, and algorithmic control.

China’s natural advantage in coordination

Together, these components mentioned above form a tightly coupled production system whose costs decline through cumulative output, learning-by-doing, and cross-layer complementarities. When each layer improves, it raises the returns to improvement in the others.

Because these components are reused across many products, learning compounds across domains rather than being confined to a single industry.

Firms like BYD, DJI, and Huawei could be misunderstood as manufacturers, but they are really system integrators that learn continuously through production, refine components through scale, and feed those improvements back into adjacent products. Coordination turns execution into a flywheel.

For any product, electrification becomes attractive only once the total cost of ownership, across capital cost, operating cost, maintenance, reliability, and performance, is lower than that of incumbent technologies. Once the threshold is crossed, adoption accelerates because the electric alternative is cheaper, better, or both.

When electric systems offer lower lifetime cost and superior performance, firms adopt them to remain competitive. Those adoptions increase scale, which pushes costs down further, pulling even more products across the threshold.

Once an application crosses the threshold, reversal is unlikely, because further learning and scale continue to improve the economics.

The strategic commoditization of AI

China’s open-weight AI push is best understood as a deliberate strategic commoditization play designed to force value migration away from the intelligence layer and into the coordination-heavy physical systems where China holds an advantage.

Strategic commoditization involves

removing scarcity from a layer you do not want as the profit center,

so that rents shift to adjacent layers you do control.

Open weights do exactly that: they reduce friction in adoption and accelerate experimentation. This speeds adoption of AI into products and operations. The point of open weights is to prevent models from becoming a choke point for anyone at all.

China’s bet makes sense. Intelligence does not create economic value at the moment of prediction or reasoning. It creates economic value only when tied to the ability to allocate resources, coordinate activity, and execute at scale.

This requires three complements:

• energy (cheap, reliable electrons),

• execution (the ability to build and operate machines), and, most critically,

• coordination (the institutional and industrial capacity to synchronize supply chains, standards, manufacturing, deployment, and learning loops).

If you already dominate these complements, abundant intelligence functions as an accelerant that raises utilization, throughput, quality, and the speed of iteration across systems you govern.

If intelligence becomes a commodity, rents migrate downstream to electrified products, integrated hardware-software systems, manufacturing scale, supply chain orchestration, and reliability in physical throughput.

A robot with a commoditized brain will not yield durable profits to the brain supplier; profits will accrue to whoever controls the motors, batteries, power electronics, integration, and deployment networks that make the robot economically valuable.

Understand China’s Reshuffle

The ideas in this post are based on the thesis originally proposed in my book Reshuffle.

Get the book

Why the US remains stuck in the intelligence distraction

The US’s quest to win AI assumes intelligence remains scarce and that coordination can be handled frictionlessly by markets. It treats execution and energy as modular, interchangeable, and cheaply purchasable through markets.

This belief was economically rational for a specific technological regime that we’ve lived through since the mid-1990s. It, however, becomes a fallacy when that regime changes.

Since the rise of the commercial internet, digital technologies have pushed marginal costs toward zero across much of the economy. Compute, bandwidth, and distribution became cheap and abundant over the past 30 years. In that environment, intelligence, expressed through algorithms, product design, and user experience, remained scarce and rent-bearing. If you controlled the user experience and harnessed user attention, you controlled the most important control point.

Ben Thompson’s Aggregation Theory describes this world. In this logic, value accrues to intermediaries that aggregate attention and demand by controlling user interfaces and informing user choice. Aggregation Theory assumes intelligence would remain scarce, coordination across the rest of the value chain would remain cheap, and the physical world would not reassert itself as a bottleneck.

These assumptions break down today. In capital-intensive, tightly coupled intelligent physical systems - factories, grids, vehicles, robots, supply chains - markets alone cannot absorb coordination complexity. Integration knowledge - how various value-creating components across the value chain fit together - becomes strategic.

China gets it! The US is still stuck in the Aggregation Theory mindset.

China’s strategy appears puzzling only if one remains trapped in the old regime. Open models look irrational if intelligence is assumed to be the choke point. They look brilliant once you make a bet on the commoditization of intelligence. By making intelligence abundant, China accelerates its absorption into coordinated physical systems - factories, vehicles, robots, grids - where learning compounds through execution.

In that sense, China doesn’t aggregate attention or users anymore; it coordinates across execution capacity, supply chains, manufacturing learning, and energy infrastructure .

Will the real AI bubble please stand up?

Aggregation dominated when intelligence was scarce and execution was cheap.

AI reverses that condition. When intelligence becomes cheap, it stops being the bottleneck and the rent-bearing layer.

What remains scarce, and therefore power-conferring, is the ability to coordinate complex systems and execute reliably at scale. Intelligence without coordination simply generates plenty of possible actions, but weak capacity to consistently realize them.

The fallacy, then, is not overvaluing intelligence; it is assuming intelligence will remain the scarce layer once a technology emerges whose primary economic effect is to make intelligence abundant.

When intelligence gets cheaper, power migrates to those who control the systems that turn intelligence into coordination at scale.

The real AI bubble is not that intelligence is overvalued, but that its ability to sustain rents is overestimated.

AI’s dominant long-run economic effect is not to concentrate intelligence, but to commoditize it, expanding the supply of intelligence faster than demand for proprietary control, and shifting durable rents to the coordination and execution layers that determine where intelligence can act at scale.

As intelligence becomes abundant, durable advantage shifts away from model ownership toward the coordination, execution, and energy systems that determine where and how intelligence can act at scale.

This is China’s play! This is where rents will eventually accrue.

Make America Great Again?

In the late 19th century, Britain and parts of Europe were the unquestioned leaders in “intelligence”.

They produced the foundational inventions of the age: steelmaking processes, internal combustion engines, chemical synthesis, electrical generation and transmission. Their universities, labs, and engineering culture were the envy of the world. They believed that if you invent the core technologies, industrial dominance will follow.

The United States made a different bet. It did not lead in invention. Instead, it focused on coordination.

American firms invested in standardized parts, interchangeable components, and production processes that could scale across geography. They built large factories organized around new coordination mechanisms rather than traditional craft. They developed managerial hierarchies, accounting systems, and planning routines to coordinate thousands of workers and suppliers. Most importantly, they built railroads as a coordination infrastructure that synchronized production, distribution, and demand across an entire continent.

British firms optimized for local performance. American firms optimized for system-level performance.

Eventually, engineering intelligence got commoditized as techniques and skilled workers crossed borders from the UK to the US, but there was no way for the coordination systems - the factories, the logistics networks, the managerial routines, and the institutional know-how required to run production at scale - to cross borders in the other direction.

The US today resembles the UK in 1880.

It leads in model architecture and training techniques and is pumping money into intelligence and the infrastructure to support it.

The implicit belief is that execution, energy, and coordination can be bought, outsourced, or handled by markets.

China today resembles the US in 1900.

It is not trying to monopolize invention at the intelligence layer. Instead, it is focused on embedding AI into factories, vehicles, robots, and grids. It’s investing in coordinating supply chains and standards and in integrating hardware, software, and energy.

Open-weight models play the same role that standardized parts played a century ago. They lower friction, encourage adoption, and ensure that intelligence diffuses rapidly into downstream systems. The goal is not to win at the frontier of invention, but to own the coordination system where intelligence creates physical capability.

Britain did not fail because it lacked intelligence. It failed because it assumed intelligence would remain the binding constraint. The United States won because it recognized that once technologies matured, coordination would dominate value creation.

When a technology lowers the cost of intelligence, power migrates to those who can coordinate its application at scale.

A century ago, that shift favored the United States over Britain.

Today, the same logic may favor China over the US, unless the US relearns the lesson it once taught the world.

Why Open China Works!

China’s openness in AI reflects an expectation that intelligence will commoditize, and a strategy to ensure that commoditization benefits China rather than undermines it.

By making intelligence ubiquitous, China increases the rate at which its factories, vehicles, robots, grids, and logistics systems absorb AI, and compounds learning where it already has advantage.

In this context, making intelligence abundant strengthens China’s position. Open models lower adoption friction, accelerate experimentation, and embed AI deeper into physical systems whose execution and energy layers China already governs. Intelligence becomes an accelerant, but not a control point.

The strategic objective of openness is not to get into some kind of AI race (never mind the pointless space race analogies) but to prevent intelligence from becoming someone else’s control point while profits and power migrate to coordinated execution.

Predictions are easy to generate and costless to abandon if the person making the prediction bears no responsibility for being wrong. And chatbots can do that better than you and me - producing fluent answers on demand.

What’s harder - and increasingly rare - is the work of reflection: slowing down, taking stock, and making sense of change.

And to make it worth your while, here’s a curated list of Holiday Readings to help you reflect more and reflect better in a world of relentless execution.

Laziness as a competitive advantage

Laziness gets a bad reputation because we confuse it with disengagement.

But a certain kind of laziness - the refusal to stay busy for its own sake - is its own form of competitive advantage.

In a world flooded with cheap execution, instant answers, and confident predictions, effort is cheap and motion is constant. What’s scarce is the discipline to stop. To not respond immediately. To not generate another take just because you can.

The ability to delay action until the question is clear and to conserve attention for what actually matters is key to leverage.

When everyone else is sprinting, the person willing to sit still can see the terrain.

So here’s to a few days of calm and quiet. Time with family - but more importantly, reflective time with yourself.

Our first three picks below focus on this idea.

Let’s dig in…

For one, stop chasing output

This is a good first read if you’re trying to unlearn the reflex to do more just because you can.

The piece makes three clear arguments:

First, when execution becomes cheap, whether it’s code, content, or prototypes, output stops being a signal of value. Václav Havel’s restraint in a suddenly free media environment shows why speaking less can matter more.

Second, advantage shifts from productivity to discernment: knowing what not to build, echoing examples like Muji and Miyazaki, who won by imposing limits rather than chasing scale.

Third, taste and coherence, not speed, become the real differentiators.

It’s a grounding read for year-end reflection, especially if you want to enter the new year thinking more carefully about where attention, judgment, and restraint actually pay off.

Stop digging faster…

Our second read is well-suited to our current moment in the AI gold rush.

In a gold rush, shovels help you dig faster - do what you already do, but more cheaply.

But when speed is cheap, what becomes scarce, and therefore valuable, is knowing where to dig with clarity.

Treasure maps help you see what you should be doing instead, and redesign your workflows, organization, and business model accordingly.

Good maps eventually spark your curiosity.

The thing about maps is you don’t need to see everything. You need to see what matters and what can be acted on.

In the early 20th century, London’s underground transit map was a tangled mess. It was geographically accurate as a map should be. But commuters had no clear mental model of how to move through the system. The map was technically correct, but practically useless.

That was until Harry Beck came along. You see below what happened next.

Beck, an electrical draftsman, treated the subway like a circuit diagram. Instead of mapping stations to their true geographic coordinates, he focused on usability. Straight lines. Even spacing. Logical connections.

The map wasn’t entirely true, but it helped you see what mattered.

Beck’s design has since influenced metro maps across the world.

That’s how a Treasure Map approach helps in a world of cheap execution. It elevates what matters and excludes what doesn’t.

And finally, here’s to a good year ahead of asking good questions…

Finally, this essay helps you step back and reset how you think. It makes three key points.

First, as AI makes answers faster and cheaper, the real bottleneck shifts to attention and judgment.

Second, real progress comes from reframing problems, not from producing more polished answers within old frames.

Third, in a world where conditions keep changing, the ability to stay curious and keep asking better questions becomes a durable advantage.

Advocating the spirit of The Holidays as a competitive advantage

This quiet period at the end of the year creates a brief but powerful advantage.

When answers are cheap and attention is scattered, the ability to step back, notice patterns, and reframe questions becomes a form of competitive edge.

Reflection isn’t passive. It’s how orientation is rebuilt before the next move.

Use these few days wisely. And see how you could incorporate such moments through the rest of the year ahead. In a world of cheap execution, the ability to step back and make sense of change will matter more than the ability to run faster and respond to change.

And with that… Happy Holidays!

If you like these ideas, get/gift a copy of Reshuffle

These ideas are based on my book Reshuffle.

Get/Gift a copy for the Holidays

If you’ve made it this far and want to dig into more recommendations, here’s the larger end-of-year reading list:

True, but utterly useless

This piece - now read more than 100,000 times- is the antidote to year-end prediction culture and LinkedIn-slogan thinking.

The piece starts with the Maginot Line, an engineering masterpiece built for a form of warfare that had already changed, and uses it to make three sharp points about how people misread AI.

First: the biggest mistake is framing AI as automation vs. augmentation. That keeps you optimizing tasks while the system of work is being redesigned. Then there’s the containerization example that I use in Reshuffle: the dockworker didn’t lose to a better dockworker; the whole port’s logic changed.

Second: productivity doesn’t automatically translate into advantage. When tools spread, execution becomes commodity input, and value often flows to whoever controls coordination, distribution, and the scarcest complements, not to the person doing the faster work.

Third: jobs and workflows aren’t stable objects. They’re temporary solutions to coordination problems. When AI shifts decision rights and execution sequences, roles get unbundled and rebundled, like basketball positions after analytics, or typists after word processors, so competing to use AI better can be the wrong race entirely.

If you stop chasing slogans and start asking how the system is changing, you can choose your next move with far more agency than the meme suggests.

Humans as luxury goods!

The AI skeptic says: “AI won’t replace the human touch.”

It sounds wise. Reassuring. It lets us relax into the belief that markets may change, tools may evolve, but people will always value people.

It is comforting. It’s also mostly nonsense.

Not because it’s false in a literal sense - humans will, obviously, remain human - but because it mistakes sentiment for economics and intrinsic value for market value. It answers the wrong question with great confidence.

Markets don’t reward what’s meaningful. They reward what’s scarce and measurable under the system that allocates value.

AI doesn’t need to replace the human touch to devalue it. It only needs to make some forms of previously human-performed work abundant, standardized, and interchangeable.

At that point, deeply human work can remain intrinsically valuable while becoming economically cheap.

When answers are abundant, the leverage shifts to people who can

(1) frame better questions,

(2) filter meaning from noise, and

(3) make calls under uncertainty while owning consequences.

The fugu guide to jobs in a world of AI

The AI skeptic says focus on what the machine can’t do.

The safe advice is to double down on “irreplaceable” tasks and wait out the transition.

It sounds prudent. Many nod along.

The trouble is that this way of thinking optimizes for the wrong enemy.

AI rarely destroys work by replacing tasks outright. It does something subtler—and more disruptive.

It changes the constraint in the system.

When execution becomes cheap and reliable, the value of doing the work collapses, even if the work remains human. Roles don’t disappear because machines outperform us; they lose relevance because the new system no longer pays for what used to matter in the old system.

This article offers a cleaner lens.

Instead of asking what AI can’t do, it asks what AI breaks: which coordination gaps widen, which risks concentrate, and which judgments suddenly become system-critical.

Like the licensed fugu chef, the winners are those positioned exactly where the new system breaks.

Making sense of hype!

We wouldn’t be wrapping up 2025 without a gentle nod to the topic of hype.

Much as we may hate it, hype is critical to capitalism today - it solves a very specific coordination problem.

Modern systems need many independent actors to move together before the payoff is visible. Capital, talent, regulators, complementors, customers - all have to act in concert under deep uncertainty. Traditional institutions used to do this work through standards, subsidies, and long-term planning. Increasingly, they can’t.

Hype fills that gap.

It creates a shared belief about a future state, early enough that people are willing to commit resources before the system actually works. It changes incentives not by force, but by reframing payoffs. Hype creates a focal point in a coordination game where no one actor can move first safely.

Yet, hype can make you consistently miss the point

AI proponents and critics land on two ends - AI will either unlock incredible prosperity or eat up all jobs.

People miss the point about AI because they argue about outcomes instead of architecture.

Both sides of the debate - doom and optimism - are trapped in the same lazy frame. They ask whether AI destroys jobs or boosts productivity, whether the pie shrinks or grows.

But AI doesn’t just change tasks; it reshapes the rules of the game. When systems are rebuilt - workflows, organizations, platforms, ecosystems - the pie can grow even as bargaining power concentrates. Growth and inequality rise in concert, by design. The mechanisms that expand output also determine who holds the knife when the pie is sliced.

The debate misses this because task-based thinking is cognitively easy and morally satisfying. It lets pessimists warn about job loss and optimists promise shared prosperity without confronting the uncomfortable middle: someone must control coordination, and whoever does captures disproportionate value.

In 2025, you can’t say hype without saying ‘agentic’

Agentic AI is framed as the next, more powerful wave of automation - an extension of RPA without the rule-based rigidity and with broader task coverage.

Success is measured in hours saved, headcount reduced, and processes automated.

The problem, though, is that workflows are treated as fixed, with agents dropped in to execute steps faster. Governance is bolted on after the fact, as compliance, audit, or monitoring.

This mindset assumes the system itself is stable and only needs efficiency upgrades. It rewards short-term gains and familiar metrics.

What really matters, though:

Agentic AI is not an efficiency technology but a coordination technology.

Its power lies in collapsing, eliminating, or radically reshaping workflows rather than automating them, step by step.

Value comes from redesigning how decisions are made, how agents interact, and how exceptions are governed across the system.

Instead of thinking about speeding up your workflows, ask the following two questions:

1. Which constraints does agentic AI remove, and which new constraints does it create?

   Stop asking what tasks get automated; start asking how the system’s bottlenecks move.

2. Where should governance, decision rights, and accountability be located once agents act in parallel?

   Stop optimizing execution speed; start designing the coordination architecture that determines who captures value.

The true opportunity lies in architecting AI-first

Most firms mistake tool adoption for architectural transformation.

AI-first means rebuilding your system around AI’s architectural properties - new atomic units, constraints, and coordination logic - rather than layering AI onto existing workflows.

It’s not about faster execution, but about changing how value is created, governed, and captured.

The difference shows up along three axes. Miss any one of them, and you’re bolting AI onto a legacy axle.

1. Atomic Unit Shift

Every real transformation starts by redefining the smallest unit of value the system works with.

Venice moved from chests of coins to ledger entries.

Figma moved from files to elements.

If your unit of work hasn’t changed, your larger system wont change.

2. System Shift

Once the atom changes, workflows, org charts, budgets, and decision rights must be rebuilt.

Walmart restructured retail around the new data created by barcode adoption.

This is where most incumbents stop, because systems change breaks power structures, reporting lines, and revenue logic.

3. Competitive Shift

Architectural shifts always change what it means to win.

Competition moves away from surface performance and speed toward control over coordination points.

Moats cease to matter when the axis of competition shifts, and incumbents usually keep defending the old one.

Most of us see the bento box and appreciate its near-ornamental visual aesthetic. But what’s most interesting about the bento box is not quite visible.

The bento is a collection of constraints that defines the structure of an entire industry behind it.

The box fixes the shape of the meal. Those constraints determine the workflow of the kitchen. And that workflow fixes the architecture of the entire supply chain.

Start with the box first. Its compartments dictate portion sizes and enforce consistency. Every meal must fit the grid.

This forces producers to engineer machinery that can portion ingredients with remarkable precision. Rice dispensers release exactly the right amount and cutters produce identically sized slices. The box’s dimensions determine the logic and parameters of the machines.

These machines, in turn, set production expectations. Farmers standardize produce and seafood suppliers target consistent cuts. Even the chemistry of sauces is stabilized so viscosity doesn’t disrupt portioning equipment.

Those production constraints also show up in logistics. Because bentos are perishable, the system must replenish stock many times per day. Cold-chain trucks are run on fixed schedule loops calibrated to store-level demand. Factories operate in short, intense bursts to prepare batches that match the expected sales by hour, not by day. On the distribution end, convenience stores design their shelves around predictable bento turnover, refreshing inventory multiple times per shift.

The architecture of the entire industry is determined by the constraints imposed by the bento box and its attached workflow. The stability of the bento workflow creates predictability. Predictability allows tighter coordination. Tighter coordination allows higher throughput.

The bento box illustrates how product and workflow constraints eventually determine industry architecture - the underlying structure that defines how work is divided, and how value and control are distributed across an ecosystem.

The ‘bento box’ guide to industry architecture

The surprising lesson of the bento box has important implications for the future of work today.

It shows that industries don’t simply choose their architecture. They are shaped around constraints. If those constraints change, the entire architecture of the industry changes alongside.

Professional services industries, today, are structured around similar constraints. Their workflows look nothing like a bento line, yet they display the same deep interdependence between what happens at the task level and what emerges as industry structure.

The work of a lawyer or auditor is document-centric, sequential, and heavily reliant on human interpretation.

Because humans can only process so much information at a time (the dominant constraint), the industry’s architecture is built around sampling techniques, multi-layered review chains, and periodic cycles to manage those constraints.

Specifically, the industry is structured around four key constraints.

1. Human speed and attention: Work must be done sequentially, step-by-step, with human review, because humans cannot monitor continuous flows of data.

2. Document-based evidence: Information is trapped in documents. Understanding that information requires humans to assess these documents.

3. Sampling and periodicity: These industries rely on episodic reviews in the form of annual audits, yearly underwriting, and periodic inspections.

4. Fragmented, heterogeneous data: Data is scattered across incompatible formats or unstructured documents. Humans do the stitching and translation to make sense of it.

These constraints also dictate industry architecture. There’s a structured cause-and-effect chain that locks in the industry’s architecture:

1. Regulation defines structure: Because of the constraints inherent to manual reviews and decision-making, regulators advocate fixed workflows and traceability requirements that ensure predictability.

2. Structure defines business model: Pricing models are anchored to billable hours. This is backed by the organizational pyramid structure, where lower-priced juniors help expand margin.

3. Business model defines skill/work patterns: Partner economics are structured around leverage i.e. how many juniors each senior could oversee. This produces junior-heavy, manual workflows.

This is a locked system. Rules lock methods, methods lock workflows, workflows lock economics, and economics lock incentives against transformation.

AI removes many of the constraints around which professional services workflows were built, and the moment the workflow changes, the architecture begins to reshuffle.

Dismantling constraints

As I explain in my book Reshuffle, AI doesn’t simply automate or augment existing tasks. That dual framing is a fallacy that distracts us from the bigger shifts. AI removes the constraints that shape today’s workflows.

AI dismantles all four constraints above.

1. AI removes human speed as a bottleneck

Agents can analyze full datasets instantly, analyze events as they happen, and even decide and act or route decisions and actions to the appropriate stakeholders for further oversight.

2. AI moves analysis from documents to underlying data

Models extract entities from static documents, map relationships between those entities, and develop a detailed understanding of the domain.

3. AI eliminates the need for sampling and periodicity

Continuous evaluation makes annual cycles obsolete. Risk and compliance are no longer episodic or reactive, but can be managed continuously and proactively.

4. AI stitches fragmented data into an integrated view

This enables what I call ‘coordination without consensus’ in Reshuffle. Actors don’t need to align on standards before they start speaking the same language or start seeing a shared view of the system.

Removing these constraints has obvious effects on the nature of the workflow, but it also eventually reorients the entire industry architecture.

Non-linearity enters film-making

To understand how a shift in constraints changes the entire industry architecture, let’s study an example where this has already played out, instead of merely speculating what might happen to professional services.

Traditionally, film editing was built around the physical limitations imposed by celluloid. ‘Film’ was literally ‘cut’. Rearranging sequences meant taping and re-taping. Mistakes were expensive, and redoing edits consumed days.

Because editing was slow, irreversible, and costly, the entire industry formed around that constraint. (I use a similar example of the impact of the word processor on the typists in Reshuffle.)

Then, non-linear editing arrived. Suddenly, cuts were reversible. Footage could be rearranged endlessly.

Films could be made faster on lower budgets. The old constraint of irreversible cuts was gone. With that, the logic of the entire system changed.

Directors no longer held unilateral power. Producers could reshuffle scenes deep into post-production. Junior editors leapfrogged veterans because software fluency mattered more than muscle memory with film. Shooting ratios exploded. Story structures changed as well because infinite rearrangement created new aesthetic possibilities.

Professional services are at a similar moment with AI today.

Law firms, audit firms, insurers, and compliance organizations operated under their own version of physical splicing: human limitations. Humans could review only so fast. Decisions had to be made sequentially. Entire industries grew around these constraints, just as film grew around celluloid.

The impact of AI is poised ot play out similar to the impact of nonlinear editing on filmmaking in the 1990s.

Reshuffling the film industry

So what exactly happened to the film industry?

First, there were the immediate, mechanical consequences. Shooting ratios exploded. Because everything can be rearranged later, directors shoot vastly more footage. Meanwhile, a cut can be reworked dozens of times in a single session.

But these mechanical shifts don’t change the industry architecture; they simply remove the constraint that shaped the craft.

Industry architecture changes once shifts in the workflow shift, where decisions are made and who sits at the new positions of power that emerge.

In movie-making, directors lose their monopoly on the cut because producers can now see and rework edits quickly.

With “fix it in post” becoming viable, production practices change as well. Actors perform with the knowledge that many imperfections can be smoothed later. Cinematographers and shooting crews change how they work as well.

The creative center of gravity moves into post-production.

With that, the role of the editor changes as mastery of software matters more than mastery of celluloid. Apprenticeships and time-based seniority lose power.

Eventually, rapid cuts, non-linear pacing, montage-heavy sequences, and complex structures become common because editors can try dozens of alternatives without cost.

Eventually, as we see with film-making, the entire industrial system changed.

Film editing is today a software discipline, built around version control. Nonlinear workflows also enable distributed editing. Post-production becomes globalized, with specialized hubs emerging - Los Angeles for direction, Vancouver for VFX, Mumbai or Manila for rotoscoping. New power centers emerge as well. VFX houses, post-production studios, and software companies like Avid and Adobe become more influential than equipment manufacturers or film labs.

Removing the physical constraint of film splicing first transformed how editors worked, then reshaped the hierarchy and economics of filmmaking, and finally produced an entirely new industry architecture built around software-driven, globally distributed workflows.

The ideas in this post are based on my book Reshuffle.

Get the book

Reimagining the nature of work

Alongside the reshuffle in industry architecture and power structures, the very identity of the movie has changed as well.

The shift from linear to nonlinear editing changed the nature of movie-making and storytelling.

In the early 1990s, film editing was still a mechanical art. Rebuilding a sequence was painful. That friction shaped the storytelling of that era with clear continuity, long takes, linear plots, and a relatively conservative approach to structure.

Sure, you could do something more experimental, but every experiment carried a cost.

Nonlinear editing changed the underlying limitations. You could try ten versions of a scene and revert to version three if things went wrong. That elasticity opened the door to new narrative habits.

The mid-90s wave of non-linear and self-reflexive storytelling - think of films like Fight Club or, a little later, Memento - reflected directors jumping to exploit the shift in constraints. Writers and directors now operated in a world where editors could rearrange time in post-production and test strange sequencing structures without destroying ‘film’.

Nonlinear editing made it cheap to experiment with non-linear storytelling.

Fight Club landed with unusual force because nonlinear editing made its fractured narrative structure feel natural rather than jarring. The film could ricochet between timelines, identities, and hallucinations with a fluidity that analog editing could never have supported. It allowed the story’s ‘content’ - the themes of alienation, split selves, and consumer-culture dislocation - to be embodied in the very form of the film.

Fight Club’s cultural impact was, in large part, attributable to the way the narrative felt like the world it described: disordered, unstable, and stitched together from competing realities.

By the late 90s and early 2000s, editing had been further transformed by digital tools that made it feasible to juggle dozens of layers, VFX plates, and micro-cuts. The Matrix exploited this in showing bullet time, slow motion, hard cuts between planes of perception, which would have been far more cumbersome in a purely analog workflow.

A few years later, the Bourne films pushed an opposite style, where fast cutting, and micro-fragments were stitched together as editors could manage hundreds of cuts in a short sequence. The ‘shaky cam’ effect gave a sense of realism in action cinema, shaping audience expectations of how physical conflict should feel on screen.

At the same time, fully digital production pipelines enabled the resurgence of fantasy and sci-fi genres. You can see it starting with The Lord of the Rings trilogy and rolling into the 2010s with the Marvel Cinematic Universe. These were films whose storytelling relied on visual density and visual continuity across dozens of interlocking movies. An ambitious cinematic universe like the MCU depends on an editing culture comfortable working inside massive, interdependent timelines. The core creative act shifts from choosing between a handful of takes to managing an evolving mesh of assets spread over years.

Editing technology also enabled a different narrative style of telling a story that ironically felt unedited. Films like Children of Men, Birdman, and 1917 used long takes and simulated one-shot structures to create immersion and tension. Those feats depended absolutely on digital editing, with hidden transitions and stitching, but the result felt like the opposite of fast cutting.

The editor’s identity changed yet again. The task was no longer to chop scenes into pieces but to hide transitions so well that the audience forgot cuts existed at all.

Today, Netflix’s binge-watching is made possible by the same technology - the ability to manage an evolving mesh of assets across different production timelines. These storytelling styles simply wouldn’t have existed in an era of linear editing. Cliffhangers, cold opens, and cross-episode callbacks are now central tools to sustain the binge and increase ARPU. The unit of narrative shifts from the film to the season, managing attention over hours.

Reimagining the architecture

That’s the real point of what’s happening around us today. AI looks like automation today. But…

When a new technology removes foundational constraints, it doesn’t just shave hours off the day. It makes different kinds of work thinkable, then practical, then expected.

The editor still edits but the underlying activity they perform and the universe around it have completely changed.

Once the constraint moves, the identity of the work moves. And once the identity of the work moves, the architecture of the industry follows.

It’s a convenient explanation.

It guarantees consensus theater.

Yet, it doesn’t quite explain why fast-moving incumbents fail.

The ones who do everything right by the textbook - the ones that both explore and exploit - the ones you talk about in HBS case studies.

Adobe checks all those boxes - one of the rare incumbents that made the leap to the cloud without imploding. In the early 2010s, it pulled off a full business model reboot, turning its one-time software licenses into recurring subscriptions. It rebuilt its products for continuous delivery and taught a generation of analysts to worship ARR (annual recurring revenue). By the mid-2010s, Adobe was thriving. The Creative Cloud became a case study in bold reinvention - the story of a company that understood the difference between digitizing its products and digitizing its economics.

The incumbent wasn’t slow.

If anything, the incumbent had moved fast and gotten it right.

And yet, as I write in Figma - the untold story, Adobe failed to counter Figma’s growing dominance.

Why exactly do the fast incumbents fail?

Below, I unpack 15 counterintuitive lessons on this topic, based on a podcast discussion I had with Aidan McCullen on The Innovation Show.

The ideas I cover apply every bit as well to the dilemmas that firms face today with AI adoption and covers:

1. Why productivity isn’t the real prize

2. Why expertise can become a liability

3. Why ‘learning AI’, while important, is not at all sufficient.

4. How power shifts not just between companies but inside careers, as algorithms strip away agency from some roles while new value accrues to others.

Read this as both a map of corporate disruption and a set of ideas to challenge how you think about disrupting your own career path.

Thank you to the first 100 paid subscribers supporting the newsletter over the past 6 weeks.

If you’ve found this newsletter useful, do consider supporting its operations with a subscription:

Subscribe now

The slow incumbent fallacy

The slow incumbent fallacy is the belief that incumbents fail because they move too slowly.

Well, here’s the thing:

The myth of slowness is comforting because it preserves the illusion of control.

It lets everyone nod along in boardrooms and MBA classrooms: just be faster, more agile, more experimental. It helps you sell Lean Startup and Agile bootcamps into organizations as some form of silver bullet.

But speed within the old frame is still the old frame.

The fallacy persists because it panders to existing hierarchies and treats failure as a matter of efficiency rather than imagination.

But architectural revolutions, like the cloud, platforms, and now AI, are not efficiency games. They are system redesign games.

There are three broad reasons incumbents love the slow incumbent fallacy, particularly when it’s sold to them in the form of workshops with a lot of post-its stuck to ‘innovation room’ walls.

1. Architectural change is harder to perceive than operational change

Operational change shows up on dashboards and KPIs. Architectural change is not quite as visible because it makes older metrics completely irrelevant.

Managers can see faster workflows, but not the invisible shift in the system’s underlying logic.

That invisibility makes architecture both the slowest-moving variable and the most dangerous to ignore.

More on this topic:

2. Execution bias amplifies path depandance

Speed is good if you’re headed in the right direction.

But if your expertise, incentives, and infrastructure are all tuned to preserve yesterday’s logic, every process, skill, and metric reinforces the old logic and makes you progressively blinder to the new one.

3. Movement is misunderstood as progress

Motion is visible, quantifiable, and consensus-friendly.

It produces the illusion of progress without confronting the existential dread of redesign.

Architecture, by contrast, requires leaders to admit they don’t fully understand the game board.

Measuring motion soothes anxiety; re-architecting threatens identity.

That’s why organizations fetishize execution while avoiding the deeper work of redefining which execution matters.

Fallacy case study - Adobe vs Figma

Let’s unpack this further using the Adobe vs Figma case.

For starters, here’s the video with Aidan exploring the case:

Now digging into the ideas in here:

1. How architectural shifts determine incumbent fortunes

2. How value migration determines incumbent fortunes

3. The shift from execution to governance

4. How moving fast in the wrong is worse than not moving at all

5. How to reimagine your career to avoid the incumbent trap

And finally, a. diagnostic to apply this to your business or your career.

Read more

For a time, canals delivered exactly what they promised: lower transportation costs and smoother flows of commerce.

Railroads, when they appeared, seemed at first like a faster version of the same idea. Yet their impact was of an entirely different order.

Like canals, railroads reduced the cost of moving goods. Far more importantly, though, they changed the entire logic of commerce.

Trains ran on fixed timetables, and as railway lines stretched across hundreds of miles, local timekeeping started to become a problem. Before railroads, every town set its clocks by the sun, which meant noon in one town could be fifteen minutes different from noon in the next. That was tolerable for canals, where barges moved slowly and deliveries were measured in days or weeks. But it made railroads unworkable.

To coordinate trains safely across long distances, the industry had to impose standardized time zones.

That act of forcing distant towns and cities to operate on the same clock had far-reaching economic consequences.

1. It meant grain in Chicago could be priced against demand in New York on the same schedule.

2. It allowed financial markets, shipping companies, and manufacturers to plan production and deliveries with a new degree of precision.

Canals lowered the cost of moving a barrel of flour from one city to another. But Railroads created a world where supply and demand could be matched across vast regions in real time, because the movements of goods and information could now be coordinated on a shared clock.

Canal engineers thought like cost optimizers: How do you cut the time or expense of moving goods along an existing path?

Railroad builders were forced to think like system designers: How do you align schedules, enforce standardized time-zones, and orchestrate the movements of thousands of passengers and shipments across a continent?

Those who clung to the mindset of canals missed the real value of railroads. They saw a faster way to move coal and cotton, but couldn’t have imagined the invention of national markets.

This is the classic trap when a new technology resembles the old. We instinctively draw on the mental models of the last breakthrough. The trouble is that those models smuggle in assumptions that limit the possibilities that the new tech offers.

Canal logic never yielded time zones. Only when people realized the railroad demanded that clocks in Chicago and New York strike the same minute, so that freight arrived predictably, did the system-transforming potential become clear.

This essay works through various ideas, including:

1. The problem with today’s agentic AI experts

2. We’ve seen the same problem before with lean manufacturing, cloud adoption, ERP, and big data

3. Two core failures in agentic AI implementations today

4. Agentic AI’s railroad moment - and what it will take to get us there

Let’s dig in!

The problem with Agentic AI experts

On the surface, agentic systems appear to be an extension of automation tools, Many of today’s agentic AI experts came of age in the world of Robotic Process Automation (RPA), where success was measured in headcount reduced or hours saved. Their language and mental models reflect that background.

In their view, agentic AI is a more powerful tool for automating tasks.

The problem is that this interpretation is the modern equivalent of seeing railroads as faster canals.

RPA was built to optimize discrete tasks within existing structures. It delivered cost savings by replacing clerical labor at narrow points in the process.

Agentic AI, by contrast, has the potential to reimagine entire workflows, and with that, reimagine the organizational systems that need to be built around them.

In Reshuffle, I use the system of work framework to explain these interdependencies.

When workflows change, reporting lines, incentives, compliance structures, and even the logic of competition change in response.

A bank that treats agentic AI as another round of task automation might save money on reconciliations, but a competitor that uses it to redesign end-to-end customer onboarding could collapse cycle times from weeks to minutes and reorient the economics of the industry.

The former sees automation and efficiency; the latter achieves coordination, and with it, an advantage that compounds over time.

As I explain in Reshuffle, agentic AI is not primarily a technology of efficiency but of coordination. When experts carry the RPA mindset into the agentic era, they risk misframing the opportunity.

Reshuffle continues as #1 bestseller in all categories

Nearly three months after its launch, Reshuffle remains a #1 bestseller in all its categories.

If you haven’t yet got your hands on a copy, now is a good time.

Get the book

The Toyota way to agentic AI

Before we unpack the limitations of current agentic AI implementations, it is worth noting that this is not the first time an entire generation of consultants has failed to grasp the true value of an emerging technology or practice.

Lean manufacturing is one of the clearest examples of how the same set of practices can produce entirely different outcomes depending on the frame applied.

In the United States and Europe, lean was often reduced to a toolkit for cutting costs. They saw the kanban boards as scheduling devices, just-in-time as inventory reduction, kaizen as incremental productivity programs.

Consultants packaged lean as a new efficiency initiative, and executives measured its success in working capital improvements and headcount reductions.

This was great for short-term results, with less stock on factory floors, leaner balance sheets, but the systemic gains never materialized.

Instead, production became more fragile when suppliers faltered, and employees learned to view lean as a euphemism for austerity.

Toyota, by contrast, never conceived lean as an efficiency drive.

The Toyota Production System was built as a coordination architecture, one that aligned the company with its suppliers, empowered line workers to halt production when defects emerged, and created dense feedback loops so that learning could compound across the organization.

Just-in-time was less about stripping out inventory and more about reorganizing workflows across the entire supply network. Quality circles - where frontline workers regularly met to identify, analyze, and solve production problems - were a lot more than the productivity tricks that Western consultants were making of them. They were mechanisms for embedding problem-solving capacity at the edges of the system. This led to a steady rise in quality and resilience, enabling Toyota to scale globally while Western rivals wrestled with recurring defects and costly recalls.

The techniques - kanban cards and andon cords - were visible in both Japanese and Western plants. But Western managers treated lean as another canal, an efficiency program that trimmed waste. Toyota treated it as a railroad, a system that required new standards, coordination mechanisms, and governance to deliver compound benefits.

That distinction in mindset explains why the same vocabulary produced fragile savings in Detroit but an enduring advantage in Toyota City.

The limits of the traditional automation view

RPA’s origins explain why its worldview is so narrow. It was designed to automate repetitive, rules-based tasks performed by clerical workers. The technology assumed a fixed structure in the inputs and processes involved.

This history conditioned automation practitioners to identify tasks in isolation and to script deterministic logic around them, and eventually, to justify projects with short-horizon cost savings.

That orientation was well-suited to the problems RPA set out to solve, but it carries hidden constraints.

It works well with structured inputs but fails in dynamic environments. It treats exceptions as errors rather than as signals of where coordination breaks down.

Accordingly, it builds automation as one-off projects instead of as evolving systems.

These habits hold back the potential of coordination that agentic AI offers. As a result, you could employ all the right technologies and still end up stuck within yesterday’s workflows.

That’s the problem with most agentic AI implementations today.

In fact, this problem shows up in two big ways that hold back most agentic AI implementations:

1. Agentic AI fails when you try to automate workflows instead of eliminating or collapsing them.

2. Agentic AI fails when you focus entirely on workflow execution at the cost of workflow governance.

Below, we explore these two ideas in further detail.

1. Which workflows should stop existing?

Let’s start with the first point.

The real potential of agentic AI is not to automate the steps of a workflow but to eliminate the workflow itself.

In talking about Reshuffle in a post that went viral on LinkedIn last week, Howard Yu makes a similar point about the book’s central message:

Workflows exist because organizations have historically needed to break down complex objectives into linear sequences of tasks performed by different roles. RPA thrived in this environment because it could pick off individual steps and automate them, but the structure of the workflow remained in place.

RPA taught practitioners to see a workflow as a string of discrete steps that could be mapped, scripted, and optimized one by one. That logic carries over when they design agents. Each agent is framed as a substitute for a task, rather than as a participant in a network of interactions.

This task-by-task orientation of RPA experts completely misses the reason agentic AI matters.

When agents can perceive context, make decisions, and negotiate with one another, the need to route work through a rigid sequence of steps no longer holds. Instead of a claims process that moves from intake to validation to adjudication in a linear chain, a network of agents can operate in parallel, gathering data, identifying anomalies, consulting policies, and resolving exceptions dynamically. The goal is no longer to make each step cheaper, but to redesign the system so that many of those steps disappear or collapse into a coordinated interaction.

This is why treating agentic AI with a task-substitution lens misses the point.

Counting hours saved at the task level is the wrong metric because the system advantage lies in eliminating the constraints that made workflows necessary in the first place. A well-designed agentic system can compress weeks of back-and-forth into minutes, not because some tasks were automated more efficiently, but because the structure of the work was reimagined.

Where RPA automated the clerks, agentic AI makes it possible to rethink why the clerks were needed at all.

It is important to acknowledge that the practitioners who came up through RPA do bring genuine strengths. They understand enterprise processes in detail. They have experience navigating change management in conservative organizations, where risk and audit concerns dominate decision-making.

But these strengths come with blind spots. Deep process knowledge is not the same as systemic vision. Comfort with task automation can make it difficult to reimagine workflows around agentic capabilities. The risk is that the very people best positioned to guide adoption of this new technology are also the ones most likely to limit its potential. Unless they evolve their perspective, the organizations they advise will see only incremental efficiency when coordination could be transformative.

2. Moving from execution to governance

Rather than automating tasks within a fixed process, agentic AI enables multiple agents to perceive context, make decisions, and interact with one another to achieve outcomes.

This brings us to the second point.

Its value does not lie in how cheaply a task can be executed but in how effectively a system of tasks can be synchronized, governed, and adapted as conditions change.

In the world of RPA, the emphasis was on execution: how quickly a task could be completed, how many hours of manual work could be displaced, how reliably a form could be processed.

When agents are not simply automating tasks but making decisions and interacting with other agents and humans, the central question is no longer how efficiently they execute but under what rules, policies, and standards they coordinate.

Governance becomes the key determiner of a successful agentic AI implementation. It defines:

1. what goals agents pursue,

2. how much autonomy they are allowed,

3. when they escalate exceptions, and

4. how accountability is assigned when outcomes diverge from expectations.

With RPA, governance was an afterthought. It showed up as compliance checks at the end of the process, audit trails to satisfy regulators, oversight to make sure bots were doing what humans used to do.

This distinction is crucial.

Poorly governed agents can amplify errors, undermine trust, or generate coordination failures that break things down in adjacent workflows even if the immediate one is managed.

Well-governed agents, by contrast, can catch problems earlier, synchronize across functions, and adapt dynamically as conditions change.

Instead of counting task-level savings, leaders need to ask how agentic systems are governed and how the architecture of workflows changes when decision-making is redistributed.

The organizations that understand this shift will capture the benefits of coordination; those that don’t will find they now have faster clerks but no systemic advantage to justify their expensive investments.

From canals to railroads - from execution to governance

Canals were essentially about execution. Once a waterway was dug, the task was straightforward: load goods at one end, move them slowly but steadily along, and unload them at the other. Each barge was largely independent. The system did not require different towns or operators to coordinate their schedules in any precise way. As a result, there was minimal need for governance. Execution capacity, in terms of more barges, wider locks, sturdier boats, was the main variable of performance.

Railroads, by contrast, made governance a central performance driver.

Trains traveled fast, shared tracks, and passed through dozens of towns in a single day. If every locality kept its own solar noon, then scheduling was impossible and collisions were inevitable.

The success of the railroad was based on how rules, standards, and coordination mechanisms were designed so that thousands of journeys could interlock without chaos.

Canals optimized execution within a fixed process. Railroads redefined governance so that the larger system could scale.

Agentic AI needs the same reframing today: moving beyond faster execution of isolated tasks to building the governance frameworks that allow many agents to act together coherently across a system.

The distinguishing feature of agentic AI is therefore not performance on an isolated task but the ability to reconfigure relationships across tasks, workflows, and actors. A supply chain improves not by merely speeding up data entry, but by acting as a network where agents coordinate forecasts, resolve disruptions, and optimize flows in real time.

With the RPA hat on, experts frequently frame agentic AI as a cheaper clerk rather than as an orchestrator of systems. Yet just as railroads restructured markets by demanding new forms of coordination, agentic AI will restructure organizations by requiring new ways of structuring governance, standards, and interactions.

Agentic AI: A case of wrong experts with the wrong frame

Every major technological wave has gone through this cycle.

Efficiency experts dominate early, extract savings, and declare victory.

Meanwhile, a different group reimagines the technology as a system architecture and captures the real prize.

Many companies invested heavily in ERP systems in the 1990s with the narrow goal of cutting back-office costs. Walmart, instead, used the technology to rewire relationships with suppliers, integrating information flows so tightly that it could orchestrate inventory across an entire retail ecosystem. Walmart saw ERP as a coordination platform, not as an accounting tool. That choice gave it a control point in the value chain that competitors could not match.

With the rise of cloud computing, CIOs in large firms justified migrations as a way to cut capital expenditures by outsourcing servers. Startups like Netflix and Stripe saw cloud as the foundation for an entirely new system architecture: elastic, API-driven, and capable of supporting products that scaled globally.

With the rise of big data, enterprises built larger dashboards, but Amazon and TikTok built personalization engines and logistics flywheels.

In each case, the wrong experts applied the old frame, extracted incremental value, and left the systemic transformation for others to capture.

Agentic AI now sits at a similar juncture.

The choice is not between adopting the technology or ignoring it.

The choice is between treating it as a canal or as a railroad.

Efficiency experts will continue to deliver savings. But the leaders who frame agentic AI as a coordination layer, who build the standards, governance, and architectures that allow agents to orchestrate complex systems, will capture the gains from transformation.

Until the mindset shifts, we will keep digging canals in an age that demands railroads.

Read more

This post is based on ideas from my new book Reshuffle.

Get the book

The story of the Library of Alexandria is often told as an example of ancient openness, a hub where the world's knowledge was collected and shared with scholars.

But the mechanics of its growth reveal something entirely different.

Every ship that entered Alexandria’s port was required to hand over any manuscripts on board. The library’s scribes copied them, and those copies stayed in Alexandria.

What appeared as openness was, in practice, a system designed to capture knowledge that leaked through trade and travel routes. The library’s power was based on its ability to intercept and absorb information that originated elsewhere.

Take yoga’s journey from India to the West. In the West, yoga is packaged as a gift of openness - an ancient practice generously offered to anyone seeking balance and well-being. Studios frame it as a kind of cultural commons, open to all who wish to participate.

Yet what gave yoga its global economic power was not its openness but the leakage of specific practices out of their original religious and cultural context.

Breathing techniques, postures, even the Sanskrit names seeped into new domains, unbundled from their context, and rebundled into fitness routines, lifestyle brands, and billion-dollar studio chains.

The flows of knowledge were never fully controlled by their originators; they leaked, and others built systems to capture and commercialize them.

Open source is not too different. It is often held up as the pinnacle of digital openness, where code is freely shared, modified, and redistributed.

Yet the firms that have extracted the greatest economic value are not the contributors but the cloud providers that take open-source tools, integrate them into large-scale infrastructure, and monetize them as proprietary services.

The openness of the community creates an environment where leaks of code and ideas are inevitable. The firms with the systems to absorb these leaks end up with the advantage.

Openness is a misdirection.

What’s really at work is leakiness.

If openness provides the appearance, leakiness supplies the mechanism.

The reason openness has become such a powerful cultural ideal in business and technology is that it works as a social signal. Declaring yourself open attracts participation, lowers resistance from users and regulators, and reassures partners that they are entering a fair system.

But the real source of durable advantage lies in how systems capture what escapes through that openness.

What matters is not whether a system is nominally open or closed, but whether it has the absorptive capacity to catch and use what leaks.

Consider Facebook’s Login API. It was presented as a convenience for developers and users: one password, access everywhere. On the surface, this looked like openness. Yet the true advantage came from the way every login generated data about user behavior across the web. That information leaked into Facebook’s ad infrastructure, strengthening its targeting engine.

Apple’s privacy posture offers a different case. The company has built its reputation on being closed to outside surveillance. Still, it allows data flows that feed its own advertising system. The signal to users is closure, but the mechanism is selective leakiness.

Stripe offers yet another angle. It does not advertise itself as open or closed. Its position at the boundary of payments means that every transaction leaks economic context to Stripe on what is sold, when, and by whom. Stripe captures and integrates that information into a broader system of financial intelligence.

In each case, the signal of openness or closure matters less than the actual underlying game of leakiness. The real economic logic is not whether you open your system to others, but whether you stand at the junctions where activity produces spillovers, and whether you can absorb them.

Platforms and ecosystems - business models that dominate the internet today - are really not about being open and closed but about creating the conditions for value to leak out of one context and into another.

This shift reframes competitive advantage in the digital economy.

Firms that succeed are not those that are most transparent or most closed off, but those that can spot, engineer, and capture leaks from surrounding systems and redirect them into their own.

The ‘openness’ signal may win attention, but the ‘leakiness’ condition decides who captures the value.

Why leakiness matters

Classical economics has long grappled with the problem of spillovers. Knowledge generated in one firm often escapes to others; ideas spread through labor mobility, reverse engineering, or casual observation. These spillovers were traditionally seen as externalities - valuable, but difficult to capture, and often wasted.

What digital technologies have done is change the character of those spillovers.

They have made them

1. Observable,

2. Weakly excludable, and

3. Rapidly absorbable.

Together, those three features explain why leakiness has become the engine of advantage in digital systems.

Observability

In the early industrial era, market activity was largely invisible till the railroad ticketing system came around. Suddenly, the flows of people and goods could be tracked, priced, and optimized, not just guessed at. The simple act of issuing tickets created a data trail.

Today, Stripe plays a similar role across internet activity. Every payment becomes an observable event.

Stripe might seem like a payments company but it’s really a financial intelligence layer for the internet. It captures patterns of activity - seasonality, demand cycles, geographic shifts - that would otherwise dissipate.

The observability of these traces is what makes leakiness possible. Without a trail, nothing leaks.

Weak excludability

Even when firms attempt to wall off their data, much of it slips out through other channels.

In economics, excludability refers to whether you can prevent others from using or benefiting from a good without your permission. A good is highly excludable if you can enforce property rights around it i.e. lock it behind a fence, put a password on it, or charge a fee for access. A good is non-excludable if, once it exists, others can use it whether you like it or not, for instance, clean air or a public broadcast.

Weak excludability is the gray zone in between. It describes situations where, in theory, property rights exist, but in practice they are hard to enforce or incomplete. Information and data are classic cases. You might own the copyright to a piece of text, but once it is posted online, it is difficult to stop it from being copied, scraped, repurposed, or used to train an LLM.

Apple’s App Tracking Transparency campaign was meant to shut down the use of device identifiers by advertisers. Yet within months, marketers had shifted to other methods, like probabilistic attribution and fingerprinting, that reconstructed user behavior from the fragments that still leaked.

Browser extensions that promise coupons or shopping help to online shoppers often end up capturing vast amounts of browsing data, and leak user data into external systems.

This incompleteness of property rights is what makes leakiness possible. If firms could perfectly exclude others from using or seeing the by-products of interactions, spillovers would remain locked down. But because excludability is weak, value escapes. The organizations best positioned to capture and absorb these leaks - whether through algorithms, networks, or institutional systems - are the ones that gain advantage.

So, weak excludability in the context of property rights means that rights may be formally defined but cannot be fully enforced.

The gap between formal ownership and practical control is where leakiness occurs, and where competitive advantage in digital systems often resides.

Absorptive capacity

Absorptive capacity is the ability of a system to take what leaks in and turn it into a durable advantage.

During World War II, Allied intelligence would intercept radio chatter. On its own, the information was fragmented and noisy. But signals intelligence operations, from Bletchley Park to the U.S. Navy’s codebreaking units, had the organizational and computational capacity to absorb those leaks and transform them into actionable strategy.

Today, Tesla relies on its fleet of cars as an absorptive engine. Every driver correction, every braking event leaks back into the central system, strengthening the autopilot. Google’s ad system functions in much the same way: trillions of search queries become raw material for auctions that match advertisers and consumers with uncanny precision.

Observability and excludability set the conditions, but without absorptive capacity, the leaks would remain unusable.

When information leaks, it flows to whoever has the absorptive capacity to catch it, structure it, and redeploy it.

This is why platforms orchestrate ecosystems, why AI systems accelerate so quickly off the back of publicly available information, and why digital moats are less about ownership than about position.

The most defensible firms are those that sit at the junctions of activity, watching what leaks through, and building the systems to turn those leaks into leverage.

Sign up for briefings

I’ve started a new series titled ‘Briefings’ which go out every week only to paid subscribers. Here’s the first one from last week:

Subscribe now

How leakiness transforms the rules of competition

Leakiness changes competition not through the ownership of assets but through the management of flows.

Its effects play out in three important ways:

1. The internalization of knowledge spillovers,

2. The creation of new control points, and

3. The destabilization of complements.

Each of these can be traced back to the way firms position themselves to absorb what escapes elsewhere.

Internalization of knowledge spillovers

In most markets, spillovers are considered a public good: valuable ideas and practices diffuse outward, often benefitting rivals as much as originators.

Digital ecosystems reverse this dynamic. Firms with absorptive capacity capture the by-products of others’ activity and fold them into their own systems.

On Amazon, sellers may think of themselves as independent businesses, but every product listing, pricing experiment, and fulfillment choice leaks into Amazon’s data infrastructure. The lessons of one seller are abstracted and applied across the entire marketplace, and in some cases, appropriated directly into Amazon’s private-label offerings.

The emergence of control points

In the industrial economy, control came from owning bottlenecks: railroads controlled track, oil companies controlled pipelines, and manufacturers controlled plants.

In the digital economy, control is exercised less through ownership and more through placement at leaky boundaries.

Stripe does not own merchants, banks, or consumers, but by handling the flows of payments, it positions itself where valuable data leaks from one domain to another. That boundary becomes a point of leverage.

Similarly, Facebook did not need to own the web to control how people moved across it. The Login API gave it a seat at the junction where users flowed between sites without changing identity, turning an apparent convenience into a strategic chokepoint.

These control points are difficult to see from the outside because they rely on leakiness rather than control.

Destabilization of complements

Platforms often encourage complements to flourish - developers building apps, sellers stocking shelves, curators creating playlists - because they make the ecosystem more attractive.

But complements cannot prevent the leakage of their contributions into the platform itself.

Spotify’s rise illustrates the pattern. Independent curators built followings and added value through their taste. But every skip, like, and playlist addition leaked into Spotify’s algorithms, which absorbed and automated the work of curation. In many ways, Spotify has today captured playlists as a mechanism to redirect attention rather than as a way for users to curate.

Over time, the complements were commoditized. Their insights were captured, abstracted, and built into the system, leaving them with little bargaining power.

The same story plays out in most ecosystems: complements thrive only so long as their activity continues to leak into the platform’s advantage.

Together, these outcomes explain why leakiness compounds power. Firms that capture spillovers strengthen themselves with every new participant. Firms that sit at leaky boundaries convert other people’s activity into control. And firms that absorb the contributions of complements eventually destabilize the very ecosystem they cultivated.

Competitive advantage in this environment is not about scaling production or locking down assets. It is about ensuring that what leaks flows in your direction

This post is based on ideas from my new book Reshuffle, now available in Hardcover, Paperback, Audio, and Kindle formats.

This post is based on ideas from my new book Reshuffle, now available in Hardcover, Paperback, Audio, and Kindle formats.

Get the book

Leakiness and power

Leakiness can reshape entire markets, confound regulators, and shift the balance of power in ways that few participants anticipate.

For one, leakiness accelerates concentration. Once a firm has positioned itself at a leaky boundary and built the capacity to absorb what flows through, each new participant strengthens the system.

Regulation often misfires in this context. Policymakers are drawn to the visible signal of openness. They debate whether platforms should be more transparent, more interoperable, more open to competition. They talk about data residency and algorithm auditability.

But the real lever of power is not openness; it is leakiness.

Firms can appear open or closed depending on the audience they are addressing, while ensuring that the leaks flow inward. Apple’s privacy positioning is the clearest example: the signal is closure, but the structure is selective permeability. By focusing on the posture, regulators miss the logic that creates defensibility.

Leakiness distorts the relationship between contributors and orchestrators. Developers who contribute to open-source projects believe they are enriching a commons; in reality, much of their work leaks into corporate systems that monetize it at scale. Cultural practices such as yoga migrate across contexts in much the same way: they appear to be shared openly, but the long-term value accrues to those who build systems that capture what leaks and package it for new audiences.

These dynamics create a strategic paradox. Firms must leak enough to attract users, partners, and complements, but they must capture enough to sustain advantage. Too much closure and the system withers; too much openness and the system diffuses without defensibility. The art of digital strategy lies in managing this tension—engineering just enough permeability to generate activity, while ensuring that the flows of data and behavior leak in the right direction.

Leakiness, in the end, is an outcome of observability, incomplete property rights, and the capacity of systems to absorb what leaks. It explains why platforms orchestrate ecosystems and why moats today are built less by what you own and more by what leaks toward you.

That is the real source of digital power, and it is the reason the firms that appear most open often end up the most entrenched.

Every Briefing will explore one single idea.

Here’s the first one…

Read more

Get the book

Netflix vs. Blockbuster is one of those well-worn stories that suggest quick and obvious explanations - all of which successfully miss the point.

The first misconception is that Netflix’s streaming beat Blockbuster’s DVD rental. But Netflix was way ahead way before streaming entered the picture.

If you rewind the tape a little further, another generic explanation is Netflix’s better customer experience - no more late fees, no more humiliation at the checkout counter when you returned “Titanic” three days late.

Those explanations fit the narrative we want to believe, that a plucky upstart delighted customers and the lumbering incumbent was too slow to adapt.

The trouble is, Blockbuster did adapt. Once Netflix’s no-late-fee policy started luring away customers, Blockbuster eliminated late fees too. When Netflix’s DVD-by-mail model gained steam, Blockbuster copied that as well, complete with its own red envelopes. On the surface, it matched Netflix feature for feature. Yet Blockbuster still collapsed.

Finally, the last bastion of defence: Netflix had better data, which helped it personalize recommendations for its customers.

Sounds right, you might expect.

But it’s yet another case of true, but utterly useless.

The real reason - fulfilment architecture

This post is not about Netflix vs Blockbuster but understanding why Netflix beat Blockbuster really helps us understand a larger point about where competitive advantage really lies.

In a 2020 post, Amazon is a logistics beast, I explain what really helped Netflix stand out:

The one thing that Blockbuster could never compete with was the integration of demand-side queuing data (users would add movies that they wanted to watch next into a queue) with a national-scale logistics system. All this queueing data aggregated at a national scale informed Netflix on upcoming demand for DVDs across the country.

Blockbuster could only serve users based on DVD inventory available at a local store. This resulted in:

1) low availability of some titles ( local demand > local supply), and

2) low utilization of other titles (local supply > local demand).

Netflix, on the other hand, could move DVDs to different parts of the US based on where users were queueing those titles. This resulted in higher availability while also having fewer titles idle at any point.

Queueing data improved stocking and resulted in higher utilization and higher availability. It allowed Netflix to serve local demand using national inventory.

I explain this further in Why offline retailers fail at online marketplaces:

There was no way for Blockbuster to compete with Netflix within the framework of local inventory - local fulfilment. Netflix fundamentally changed that framework.

This Netflix parable is important - it helps us identify a key factor that drives the competitiveness of companies in the age of data and AI.

Debunking the ‘data is our moat’ fallacy

The “data is our moat” fallacy is the belief that simply accumulating large amounts of proprietary data automatically creates a durable competitive advantage.

The logic goes like this: data is scarce, data is valuable, and the company that collects the most will be hardest to dislodge.

This assumption is flawed for several reasons. Data is often substitutable, its marginal utility declines rapidly in most cases, and features built on data are copyable.

Defensible advantage in data-driven businesses does not come from having data in isolation, but from the data-informed architecture a company builds.

What matters is the architecture that emerges when data is embedded in the system itself. Netflix’s queue created an entirely new operating model. Competitors could mimic features and even acquire similar datasets, but they could not easily rip out and rebuild their architecture without unraveling their existing model.

The puzzle of Netflix versus Blockbuster, in other words, was not about who had the data, but about who used it to reconfigure the logic of the business. Blockbuster never escaped the gravity of its old architecture.

In short, the fallacy is mistaking data as an asset in isolation for data-informed system design. The former is often transient; the latter, when well-architected, can produce durable advantage.

The architecture is the moat

Competitive advantage in data-driven businesses comes less from the data itself and more from the architecture that data makes possible.

In any complex system, outcomes are shaped less by individual parts than by how the parts are arranged - the feedback loops, buffers, and flows that govern behavior over time.

Data provides visibility, but it is the architecture that determines how variability is managed and how guarantees are met.

A well-designed architecture creates reinforcing loops: better forecasts reduce variance, which improves reliability, which attracts more usage, which in turn generates better data. These loops compound, widening the gap between firms that embed data into their structures and those that treat it as an add-on.

Once an architecture is built, it channels behavior in ways that are difficult to reverse. Changing it requires not just new inputs but dismantling and reassembling the system itself.

In this sense, architecture becomes the true moat. It is the pattern of interconnections, informed by data, that locks in compounding performance improvements and makes imitation prohibitively costly.

The AI-first series

This post is fourth as part of an ongoing series on AI-first companies.

You can see the previous posts below:

Data unlocks architectural moves

Data can help unlock a range of architectural moves. Four prominent ones are particularly interesting in the case of order fulfilment as with Netflix above.

1. Risk pooling

Without data, firms guess demand locally, leading to frequent stockouts in some stores and excess in others. With data, firms can forecast across a wider population, consolidate inventory, and reduce variance.

Centralizing stock reduces the safety buffer required to meet uncertain demand. Netflix did this with DVDs. Instead of guessing what each neighborhood wanted, it pooled demand nationally, shipping from a central stock based on queue data.

Amazon uses a similar principle with regionally optimized fulfillment centers, dynamically reslotted as demand signals shift.

These data-informed fulfilment architectures can promise availability with fewer assets, something a competitor tied to a local inventory model cannot easily match.

2. Demand shaping

Data allows companies to manage demand, not just respond to it.

Netflix’s queue was an early example. By asking customers to line up what they wanted, it smoothed out variability and gave the firm foresight into future demand.

Amazon does this through delivery-date promises; the moment a customer sees arrives Thursday, they are committing to a slot that Amazon has calculated it can meet.

In India, quick-commerce players use batching windows - “delivery in 10 minutes” versus “within 30 minutes” - to steer demand toward times and routes that make the network most efficient.

Here, the promise itself becomes the product. Competitors may have similar goods, but without a system designed to manage promises at scale, they cannot compete on reliability.

3. Postponement

The later you commit, the more accurate your decision, because uncertainty reduces as time passes. Data makes postponement feasible at scale.

Amazon leverages robotics and sortation systems that hold orders in a “ready-to-ship” state until late in the cycle, when routing decisions can be made with fresher data.

Quick-commerce firms prepare pods of popular items in dark stores, holding off on final assignment until the last minute. The architecture absorbs uncertainty by delaying commitments until predictions are more accurate.

The outcome is higher reliability without excess cost. Replicating this requires both better forecasts and physical and organizational systems designed for late-binding decisions.

4. Economies of density

Perhaps the most powerful effect of data-informed architecture is the shift from scale to density.

Traditional firms think in terms of growing volume nationally, but in last-mile systems, what matters is the number of orders per square kilometer per hour. Data allows dynamic routing and clustering so that each new order in a neighborhood reduces delivery cost and time.

Dark stores placed close to demand, riders dispatched in batched clusters, and vans leaving fuller as local demand grows - all these effects create non-linear efficiency gains. This is what makes quick-commerce viable: the system compounds as density increases.

A competitor operating on a broader but thinner footprint can grow overall volume but will never match the cost curve in dense zones without restructuring around the same principle.

This post is based on ideas from my new book Reshuffle, now available in Hardcover, Paperback, Audio, and Kindle formats.

Get the book

Deep-dive and framework

Now that we’ve covered the core idea, the rest of the post (paywalled) gets deeper into the application:

1. Unpack the ideas above futher by diving deeper into e-commerce and quick-commerce fulfilment architectures.

2. Extend the AI-first thesis that we’ve been developing across the past 3 posts.

3. Close with the final framework: The five tests for an architectural moat

Read more

Get the book

Traditional publishers in the gaming industry have repeatedly attempted to replicate Roblox.

They launch platforms with familiar ingredients: simplified graphics, accessible scripting tools, creator marketplaces, and virtual currencies. On paper, these efforts should work.

The incumbents control world-class studios, operate at a scale far larger than Roblox, and have the financial resources to fund creator incentives.

Yet, they fail to replicate Roblox’s success.

The challenge lies not in replicating surface features but in replicating the underlying structure.

Roblox is not just another game platform; it is built around a fundamentally different atomic unit of value.

Inverting the logic of gaming

The traditional gaming industry has been structured around the title as the unit of production and monetization.

Studios raise budgets, allocate teams, and measure returns on a title-by-title basis. Incentives throughout the system reinforce the centrality of the title.

Roblox begins from a different unit. It treats the experience block - an environment, an asset, a behavior - as the foundational building block. Blocks can be created quickly, remixed easily, and reused across many contexts.

Once the block becomes the operative unit, the entire system above it reorganizes.

1. Discovery shifts from promoting titles at launch to recommending flows of activity across many blocks.

2. Monetization shifts from unit sales to streams of microtransactions within an ongoing economy.

3. Identity persists across worlds rather than resetting with each new release.

Incumbents cannot simply bolt this model on because their systems assume the title as the atomic unit. Their payout structures, content policies, and discovery algorithms are written for packaged games, not for composable experiences.

Even when they copy the visible features, they remain locked into the economics of the old atom.

Unbundling the atom to create a new system

The unbundling of the game title into experience blocks set off a cascade of effects.

First-order effects - New sources of innovation

First, this unlocks a new source of game experience creation.

Developers assemble experiences from shared libraries, deploy them instantly, and update them continuously.

Avatars and inventories can travel across worlds, making identity portable in a way that the title-based model never allowed.

Second-order effects - New economic logic

At the second level, a market begins to form.

Different types of creators emerge: some focused on world-building, others on scripts, others on avatar skins.

Roblox introduced its virtual currency Robux to enable exchange and designed recommendation systems to optimize for overall time-in-world rather than for the success of individual titles.

The result was an economy where coordination mattered more than production.

Third-order effects - New governance

At the third level, institutional infrastructure changes.

Roblox had to invest in trust and safety operations, enforce age bands, detect fraud, and design payout contracts that stabilized creator income.

Standards for versioning and dependency management became critical for ensuring interoperability across the building blocks.

Roblox’s economy looks less like a traditional game studio and more like a self-contained economy with its own labor force, currency, and governance.

Anyone, from a teenager in their bedroom to a micro-studio, can create and upload building blocks that plug into this economy, coordinated through a few common mechanisms:

1. Robux acts as the medium of exchange

2. Creators earn payouts based on time spent and in-game transactions

3. A marketplace allows buying, selling, and remixing of assets

4. Governance structures manage trust, safety, and age-appropriate content.

This progression explains why incumbents failed to adapt. Even if they copy the surface-level ‘features’, they are architecturally constrained unless they unbundle their atomic unit - the title - into more foundational building blocks.

Doing that would mean pulling apart their entire business and rebuilding a new one from scratch.

That is the power of an architecturally-native business model.

For more on the idea of an architecturally-native business model, read:

The incumbent architectural lock-in

The inability of incumbents to replicate Roblox lies not in their technical capability but in the lock-ins embedded in their operating model.

The first hurdle is accounting.

Incumbent accounting logic is built around titles. Their revenues, greenlight approvals, and performance metrics all assume the title as the unit of measurement.

The second hurdle is org design.

Their organizational design reinforces this, with teams structured as discrete studios producing individual games on multi-year cycles.

The third hurdle is internal product fiefdoms.

The IP regime privileges ownership and control, making it difficult to accommodate derivative rights and remix culture, and accept interoperability with less successful titles within the same studio.

The fourth hurdle is risk.

Their safety and compliance processes are modeled on batch testing before launch, not on continuous governance of a live marketplace.

The final hurdle is design.

Sunk investments into high-fidelity graphics and immersive engines may create resistance to moving to standardized, low-variance building blocks.

Each of these lock-ins is rational when the title is the atomic unit of value.

Together, they make sense of why incumbents thrived in the previous system.

But when the atom shifts, the very structures that once conferred advantage become obstacles.

To copy Roblox would require breaking these lock-ins simultaneously, which is not simply a matter of launching a new product or platform but of abandoning the entire economic logic on which the incumbent firm is built.

So the next time you hear advice about moving from a product to a platform or about building out user communities, think again about whether your atomic unit allows you that flexibility.

The Roblox way

Roblox’s durability comes from being architecturally native to its chosen atom.

As we noted in last week’s post, four properties define this:

1. The shift in the atomic unit from the title to the gaming environmental block.

2. The embrace of constraints as design principles. Low-fidelity graphics combined with simplified scripting act as standards that enable interoperability and scale.

3. The recomposition of systems around the new architecture

4. The reframing of competition. Roblox competes on coordination i.e. who can orchestrate the most creators, sustain the deepest engagement, and govern the most resilient economy.

We noted these properties with the shift from Adobe to Figma as well, and in explaining why Adobe - structured around the design file - is structurally incapable of replicating the success of Figma - structured around a design element as the atomic unit.

For more on the shift from Adobe to Figma, check out my interview on The Innovation Show below:

By shifting the atom, Figma embraced the constraints of the browser, rebundled collaboration workflows, and reframed competition around design workflow and governance in the enterprise.

The pattern is the same as in gaming: change the atom, adopt its constraints, rebuild the system, and redefine the axis of competition.

But there is also a difference in scope.

Figma rebundled workflows, but Roblox rebundled an entire economy around the new atomic unit.

The same structural forces play out, but in the case of Roblox, they extend beyond workflow into labor markets, currencies, and governance.

Unbundling knowledge work in the age of AI

All of this matters because AI today unbundles knoweledge work into fundamentally new atomic units. I explain this in Chapter 7 of Reshuffle:

Historically, expertise and specialized knowledge were tightly bundled with human labor - to access expertise, you had to hire, train, and manage workers. As a result, organizations paid a premium to access knowledge, and hit bottlenecks when they tried to scale it.

AI changes this. It unbundles expertise from the expert, turning knowledge into a capital asset rather than a labor input. Instead of hiring someone to perform a task, you can now rent the associated capability.

When knowledge is unbundled from human labor and becomes accessible as capital, it gains three essential traits.

It becomes rentable, as you can access it without long-term commitments.

It becomes recombinable, since different forms of expertise can be recombined without the overhead of coordinating across siloed teams.

And it becomes scalable: once a solution is built, it can be deployed repeatedly at near-zero marginal cost, unlike human labor, which scales linearly with cost. 

The availability of expertise as building blocks changes productivity, but more importantly, it changes power.

Knowledge workers who once sold their labor can now package and deploy it as a building block. On the other hand, solopreneurs and creators gain leverage by combining these building blocks into new businesses.

The nature of competition also changes as a result.

Capabilities bundled with underlying assets were confined to the boundaries of a specific industry. However, as building blocks, they are now available to be leveraged across various industries. Industry boundaries don’t matter and competition, instead, plays out in connected ecosystems where these building blocks are now available across industry boundaries and success is determined not by what you own, but by how well you assemble and coordinate the building blocks that others provide.

Get the book

With this unbundling and shift in the atomic unit - from the performance of knowledge work tied to skilled workers to components of knowledge work accessible on-demand - we have the opportunity to rebundle not just new workflows as Figma did, but entirely new economies as Roblox did.

What does it take to create a new economy? - Lessons from the rise of TV programming

In the mid-twentieth century, the movie studio was the unquestioned center of the entertainment economy. Everything revolved around the feature film.

Studios financed production on a per-title basis, distribution schedules were structured around theatrical runs, and revenues were tallied in box office receipts.

The incentives of the system, from star contracts to marketing budgets, all assumed the feature film as the atomic unit of value.

Then television arrived, and changed the atomic unit.

The feature film gave way to the episode.

Much more importantly, the episode enabled the creation of the time slot for TV programming - an entirely new atomic unit, enabling the rise of an entirely new economy.

A film could be two hours long or three; an episode was standardized at twenty-two or forty-four minutes, carved up by advertising breaks. This constraint quickly became the basis for a new economy.

Advertisers could price and buy predictable slots. Writers’ rooms could develop story arcs across serialized episodes. Syndication markets emerged to sell bundles of episodes into different geographies and time zones. The entire structure of incentives reorganized itself around the new atomic unit, unlocking new value flows and new ways in which players could participate.

In fact, it was only with the rise of Netflix that this atomic unit of the programmable time slot got dismantled. Today, TV channels struggle in a Netflix era not because Netflix has better UX or a superior content library but because their entire architecture is structured around an atomic unit - the programmable time slot - that has today become irrelevant.

Ironically, this is also why sports rights have increased in value becuase live sports is the last remaining bastion preserving the logic of the programmable time slot and the entire economy around it.

But that’s a topic for another day…

Before you read further…

Starting this week, this newsletter is now open to paid subscriptions.

I’ve not turned on paid subscriptions so far even though there have been many pledges for this newsletter. I didn’t want to hold back my writing from readers.

I learnt recently, however, that Substack only promotes paid newsletters. Ironically, keeping this newsletter completely free limits its growth as the platform chooses not to promote it.

I turned on paid subscriptions this week and the newsletter immediately shot up to #4 in the Technology category. That has dramatically increased subscribers as the platform is now pushing readers to the newsletter.

Shout out to AllThingsADAS for pledging support as a founding subscriber.

I will continue to ensure that there is really good material available for non-paying subscribers - as you can see with this post.

The core argument I am making will continue to be available for free subscribers. However, most posts will have the last 40% behind a paywall where we

1. Crystallize the framework,

2. Apply it to today’s business dilemmas, and

3. Conclude with a diagnostic or a set of questions to think through.

Accordingly, what follows through the rest of today’s post is the following:

1. Why it is difficult to identify the shift in atomic unit - and what to look for

2. How new economies get unlocked around a new atomic unit

3. How the convergence of railroads and telegraph unlocked the modern corporation around a new unit

4. Designing an AI-native economy

Let’s dig right in…

Read more

Get the book

Every startup deck claims to be AI-native.

Every incumbent insists it is becoming AI-first.

But when you press them to explain what those phrases actually mean, the answers tend to collapse into clichés: faster automation, smarter tools, agentic workflows.

Press harder and the answers are always ‘operational’ - a faster move within today’s game.

Most executives talk about AI as if it were electricity: a general-purpose input that can be plugged into any process. The metaphor is convenient but misleading.

Electricity mattered less as an input than as an architecture that redefined how production was organized.

Steam-driven plants had been organized around a single massive drive shaft; electric motors allowed production lines to be reconfigured entirely.

You can’t really be ‘AI-first’ or ‘AI-native’ unless you reimagine your business around the architectural properties of AI.

Through economic history, the organizations that defined new technological eras were rarely those that adopted a new technology, but those that allowed its architecture to reshape the foundations of how they worked, coordinated, and competed.

To understand what it truly means to be AI-native, it helps to step back from today’s slogans and look at the larger structural shifts that have played out whenever a new underlying technology showed up.

This post unpacks these ideas.

What follows is a framework drawn from economic history, identifying four recurring properties that mark when a system is genuinely native to a new architecture:

1. The redefinition of the atomic unit of value,

2. The integration of constraints as design features,

3. The rebundling of organizational systems, and

4. The reframing of competitive advantage.

From Venetian ledgers to movable type, from the telegraph to containers and barcodes, we see this pattern repeatedly show up. The winners - in every instance - were those who stopped bolting new tools onto old logics and instead rebuilt their systems in the image of the new architecture.

The ideas in this post are based on my new book ‘Reshuffle - Who wins when AI restacks the knowledge economy.’

Get the book

Architecture eats execution for breakfast

Being architecturally native means designing a system from the ground up around the capabilities and constraints of a new technological input.

Consider the shift to containerization, the story I open Reshuffle with.

For ports that treated the shipping container as merely a tool for the automation of cargo handling, the gains were modest. But for those that rebuilt their operations around the box, primarily, by transforming from just an automated port to an intermodal transportation hub, the container became an engine of transformation around which they reorganized their economy. The container, eventually, imposed a new logic on the entire supply chain, unbundled vertically integrated manufacturing, and reorganized global trade routes.

The advantage did not flow to the ports that handled containers most efficiently within the old frame, but to the ones that redesigned themselves around container standards and positioned themselves at the centre of this evolving ecosystem.

The introduction of the barcode - a story I explore in Chapter 3 of Reshuffle - had dramatically different effects on the fortunes of Kmart - which adopted it as a tool - and Walmart - which pioneered the barcode-native retail architecture.

Kmart saw it as a faster way to ring up purchases, saving a few seconds at checkout. Walmart understood it as an architectural shift. The barcode created a stream of item-level sales data that could be used to reorganize and reorient the supply chain, cross-dock warehouses, and enforce vendor-managed inventory. This resulted in a lot more than store efficiency.

It redefined competition in retail, where power moved from brand manufacturers to data-rich retailers who governed replenishment.

These stories make a case for a broader thesis.

To be architecturally native is to stop competing on execution within the old system and to rebuild the system around the architecture of the new one.

The arc of transformation

Every architectural shift restructures three levels of the system.

1. The task or the unit of work: The atomic unit of value shifts, enabling new workflows and rendering incumbents’ unit of work obsolete.

2. Organizational logic: Value creation moves from doing work faster to coordinating and governing work differently.

3. Ecosystem structure: Entire industries reorganize: who participates, how coordination happens, and where defensibility lies all change.

Figma - the overlooked insight

Last week’s article titled Figma - The untold story unpacks this by explaining why Adobe finds it difficult to compete with Figma because the former uses the cloud as a delivery mechanism while the latter reimagined its entire architecture around the properties of the cloud:

Adobe's design logic is built around the design file (.psd, .ai) as the atomic unit of work.

Figma’s design logic is built around an element in the design file - a button, icon, or type style - as the atomic unit of work.

Changes and permissions could be tracked and managed at the level of a design element. Each element was addressable in a database: change a component once and that change propagated everywhere it appeared. Permissions replaced ownership; engineers, product managers, and marketers could view or comment without being sent anything.

The shift from ‘file’ to ‘element’ as the atomic unit of work had another important effect. Because of the element-based architecture, Figma users could create shared libraries of reusable design components, like buttons, icons, type styles, and color palettes, that teams could use across multiple files and projects. Instead of duplicating these elements in each file, designers simply reference a single source of truth.

This creates consistency, simplifies updates (change once, update everywhere), and enables cross-functional teams to work with aligned visual standards. Shared libraries shift design from isolated file ownership to coordinated, system-level collaboration.

This architecture created strategic separation from Adobe. Adobe used the cloud to deliver the same file‑based logic more efficiently. Figma used the cloud to replace that logic entirely.

By shifting the unit of work from file to element, Figma enabled real‑time collaboration, created a shared design environment that expanded who could participate, and made Adobe’s model feel increasingly constrained by its own architecture.

This simple shift eventually transformed the structure of the entire industry:

In traditional file-based systems, value was created and captured inside closed loops: files lived on local drives, changes were tracked by humans, and tools were optimized for ownership and execution.

The dominant logic was self-contained workflows: a designer edited a file, exported assets, and handed them off, often using proprietary formats inside siloed tools.

But element-level architecture unbundles the design process into modular, reusable pieces. This naturally dissolves the boundary between inside the tooland outside the tool.

With components living in shared libraries and third-party tools can plug into atomic design elements through APIs, interoperability was invevitable.

This shift fractured the vertically integrated model Adobe had dominated. Just as the modular web displaced proprietary desktop software, Figma’s architecture enables a loosely coupled, composable ecosystem of tools. integrating at the level of individual design elements.

Value no longer accrues to those who own the file, but to those who coordinate the system, through reusable design tokens, shared standards, and governance mechanisms.

Figma’s story is a reminder that architectural shifts often look deceptively small.

Moving from files to elements might sound like a technical detail, but in practice, it redefined the unit of work, the basis of coordination, and the structure of competition across an entire industry.

The atomic unit that rearchitected Venice

Trade in medieval Europe was a bit of a gamble. A merchant might finance a voyage to Alexandria, but his entire fortune was tied to a single, indivisible shipment of metal currency: if the chest was to a storm in the Aegean or to pirates off the Barbary coast, the entire venture collapsed. Risk was binary and concentrated - one chest tied to one voyage, yielding one outcome.

Venice changed that by altering the atomic unit of trade.

In the fourteenth century, Venetian merchants developed double-entry bookkeeping and redefined the atomic unit of commerce from the chest of coins to the balanced transaction.

Each entry in a ledger served as the new atomic unit of trade, capable of being audited, transferred, and reconciled at a distance. Similarly, the bill of exchange allowed money to move as paper promises, negotiable across cities and payable in different currencies. A sack of coins could vanish at sea, but a paper claim could be settled in Bruges or Genoa, unbundled from the voyage that carried the goods.

When the atom shifts, scale and coordination follow new logics. This new atomic unit - an entry in a ledger - enabled the creation of balance sheets, external investment, and eventually the rise of the modern corporation.

This atomic shift transformed the nature of risk. Loss was no longer catastrophic and indivisible; it could be diversified across dozens of ledger entries, pooled across convoys, and insured through collective instruments.

Venice’s muda system institutionalized this principle further: merchants no longer sent ships individually but banded together in state-protected fleets with fixed schedules. The convoy itself became the new organizational system of commerce, turning unpredictable voyages into routinized channels of trade. Pirates might capture a ship, but they could not bankrupt an entire city.

Once risk was unbundled into these smaller, more fungible units, the Venetian economy expanded dramatically.

Families that once tied their fortunes to a single voyage could spread investments across many. Credit could be extended at scale because the underlying records were reliable and auditable. Insurance markets developed, backed by the predictability of convoys and the credibility of the books that recorded their transactions. And most importantly, trade networks could extend deep into the eastern Mediterranean and northern Europe without requiring every transaction to rest on personal trust between two individuals. The ledger and the bill stood in for personal reputation, enabling anonymous exchange at a distance.

The unintended consequence of this shift was that Venice became more than a city of merchants; it became the hub of a financial system. Once recorded in ledgers, wealth could be leveraged to finance new ventures, armadas, and even wars.

By redefining the atomic unit of commerce, Venice transformed risk from a personal gamble into a system-level variable that could be diversified and insured. This unlocked scale, making Venice, perched on its lagoon, one of the strongest economic powers in Europe.

Being architecturally-native - Four properties

To be architecturally native is to allow a technology’s structure to become the foundation of your system, not an accessory.

Not simply to adopt a technology, but to be defined by it.

The difference between adaptation and nativeness is the difference between layering a new component onto an old design versus allowing the underlying architecture of the system to be reconstituted.

Four recurring properties determine this:

1. A shift in the atomic unit of value

Systemic shifts start with redefining the smallest unit of value.

Complexity is built from atomic units - basic building blocks that determine how larger systems scale. Change the atom, and the system above it must reorganize, as we noted with the example of Figma.

When the unit of exchange or production changes, transaction costs, coordination mechanisms, and value capture shift with it.

Much like the shift from Adobe to Figma, Gutenberg’s invention of the movable type changed the atomic unit of how knowledge was organized and reproduced.

Before Gutenberg, the book was the atomic unit. A manuscript was copied from beginning to end by a scribe, and every book existed as a discrete, indivisible object. If you wanted another, you had to start again from scratch, one page at a time. The cost of duplication was high, and the variation between copies was inevitable.

Movable type unbundled that unit. Instead of treating the book as the smallest whole, Gutenberg treated each character, cast in lead, as the atomic unit of value. By rebundling these atomic units, printers could construct words, sentences, and pages, then disassemble them and reuse the type to build entirely new texts.

Knowledge was no longer bound up in singular artifacts; it became modular, replicable, and composable. That shift allowed for standard editions, catalogs, and the beginnings of an information market where texts could circulate at scale.

The analogy to Adobe’s file versus Figma’s element is obvious. Once the unit was unbundled, everything else changed: workflows, collaboration, industry structure, and even the economics of scale.

2. The embrace of constraints as design principles

Architecturally-native systems don’t treat the limits of a new technology as weaknesses to be engineered away; they adopt them as operating principles.

Systems are defined as much by their constraints as by their capacities. As I explain in Reshuffle, constraints provide stability, reduce degrees of freedom, and direct behavior into predictable patterns.

The muda - Venice’s fixed, state-protected convoy departures timed to seasonal winds - locked merchants into rigid schedules. These constraints reduced movement, but they also reduced piracy risk and synchronized cash cycles, allowing for pooled insurance, enabling Venice to dominate Mediterranean trade.

The telegraph, with its narrow bandwidth, forced traders and journalists into a language of codes and ticker symbols. Brevity became the norm, and from that constraint emerged new grammars of communicating finance and news.

Even containerization required ports and shipping companies to accept the rigidity of standardized box sizes, a compromise that might have seemed costly on the face of it but unlocked global interoperability.

Constraints, when internalized, create predictability and faster learning loops.

3. The recomposition of systems around the new architecture

Once the atomic unit changes and constraints are absorbed, entire workflows and organizations are rebuilt.

Double-entry bookkeeping made it possible for investors in Florence or Antwerp to entrust money to managers they would never meet, because the ledger itself became a trustworthy account of performance. That in turn allowed for external auditing, separation of ownership and control, and the rise of the joint-stock company.

These are second and third-order effects, not always visible at first. Small changes in components or rules can generate large-scale structural shifts.

The barcode at the supermarket checkout was more than a faster way to check out items. It created a continuous stream of demand data, recomposing retail from store-level merchandising to system-level replenishment. Walmart used that flow to cross-dock warehouses, enforce vendor-managed inventory, and reverse the traditional balance of power between retailers and suppliers.

In the hands of an architecturally-native player, a simple technology becomes the basis for rethinking how an entire organization coordinates and governs itself.

The rise of the railroads alonside the telegraph is possibly the best example of such a systemic shift.

Before the railroads, most businesses operated at a local scale. A factory might employ dozens of people, a trading house might operate a few routes, but management could be handled directly by owners or a handful of trusted clerks. Information moved at the speed of letters and couriers, and decision-making was largely in-person.

The railroad changed that, and the telegraph made it possible. Rail lines stretched hundreds of miles and trains ran simultaneously in both directions, forcing the need for precise coordination. A single delay in one town could have larger effects across others. Without real-time communication, chaos and collisions were inevitable.

The telegraph solved this coordination problem. It allowed railroad companies to centralize dispatching: trains could be scheduled, rerouted, or held back based on telegraphed updates from stations along the line. To make this work, railroads also needed standardized time, which is why the U.S. and Britain adopted time zones, an institutional response to the demands of rail and telegraph integration.

But perhaps the most enduring change played out organizationally. Railroads became the first truly large, geographically dispersed corporations. Owners could no longer manage by proximity, so they invented new structures: regional divisions, professional managers, and reporting hierarchies.

These were the foundations of what Alfred Chandler later called the visible hand or more bluntly, the rise of managerial capitalism.

4. Reframing of competition

Once a new architecture takes hold, the game itself changes.

In the age of railroads, competition moved from who had the fastest locomotive to who controlled the timetables and the through-rates.

Containerization rewrote shipping competition from handling capacity at individual ports to the efficiency of end-to-end intermodal routes.

Each time, incumbents who clung to the old axis of competition - whether engine horsepower or port size - found themselves blindsided by rivals who had redefined the game.

Shifts in architecture change the system’s boundary conditions and the criteria on which firms differentiate and capture rents. The Venetian merchants who embraced convoys timed to seasonal winds illustrate this best. The competitive axis shifted from individual risk-taking to collective risk-pooling and insurance capacity, making Venice the most powerful merchant shipping hub of its time.

These four ideas determine how workflows, organizations, and business models restructure around new technologies - not based on their performance as inputs as much as around a new logic of value creation across the system.

Reshuffle is now 1 month and 50 reviews old

Here’s one:

If you’ve enjoyed the ideas in this post, go ahead and get the book to access the larger framework!

Get the book

The slow incumbent fallacy

When new architectures emerge, incumbents rarely stumble because they lack speed, vision, capital, talent, or the will to compete.

They falter because their existing systems are locked into a particular logic, and abandoning that logic would mean unraveling the very structures that sustain them.

Adobe illustrates this dilemma. Its dominance was locked-in to the logic of the file as the atomic unit of design work.

To abandon that structure in favor of Figma’s element-based architecture would have required dismantling the entire system of connected tools, workflows, and customer habits that Adobe had cultivated for years.

Path dependence amplifies this lock-in. Factories in the early era of electrification often installed electric motors but left their layouts unchanged, still organized around a central shaft designed for steam. Their workflows, contracts with suppliers, and even union agreements were tuned to the rhythms of the old architecture.

The incumbent is rarely blind or inert. They see the new architecture. But to adopt it fully would mean rewriting their workflows, revenue streams, and identity.

We often over-emphasize the need for operational agility, suggesting that incumbents fail because they move slowly.

The real reason incumbents struggle is not operational agility, it is structural agility - the inability to unlock their existing locked-in architecture.

Adopting ‘agile’ will not solve the real problems your organization faces - that of structural agility.

A final diagnostic

If you really believe you’re pursuing an architecturally-native approach, ask yourself the following questions:

1. Atomic: What new atomic unit replaces the old one?

2. Constraint: Which hard constraints are you embracing as design?

3. Rebundling: Which workflows, budgets, and authorities are rewritten around these new constraints?

4. Reframing: On what new axis of competition will you win, and which incumbent moats become irrelevant?

Reshuffle on the podcast circuit

I’ve been doing a range of podcasts on the ideas in Reshuffle, including BCG Ideas and Thinkers, AI@Wharton, the Futurists pod, and several others.

If you’d like to discuss Reshuffle on your podcast or feel it should be brought to a podcast in your network, just hit reply and let me know.

In the meantime, here’s one pod to explore some of the ideas in the book further:

Both paperback and hardcover are now available in India as well.

Get the book

Figma has had quite a ride in recent years, from Adobe’s acquisition attempt and regulatory scrutiny to a massive IPO followed by a quick reality check.

Crashing the gates of a well-defended incumbent is one thing. Getting them to pay a premium for you, and eventually getting investors to shower you with an even higher premium and devalue the incumbent in the process - that’s no joke.

Figma’s success is often explained with familiar reasons: it enabled real-time collaboration, grew through strong network effects, delivered great user experience, and cracked enterprise adoption.

All of that is true. But if those really are the factors that explain its success, why couldn’t Adobe, with its resources and dominance, simply do the same?

Adobe had the talent, the cash, and the incentive. Collaboration isn’t a deep trade secret, and enterprise distribution is Adobe’s home turf.

The answers to this question, again, circle back to well-worn tropes

• Incumbents are slow, startups move fast.

• Ideas are cheap, execution is the differentiator.

These explanations are, in a phrase - true, but utterly useless - and examples of what I’ve previously called consensus theater.

The problem with consensus theater is that the topic ends right there. Everyone leaves the room feeling smart, yet not a single person has a clue on how to apply this newly acquired insight the right way.

So the real answer must lie somewhere deeper.

This is the untold story of Figma.

Figma - the well-worn story

If you’re not entirely familiar with Figma, let’s first start with the well-worn story used to explain Figma’s success.

Figma is a cloud-hosted design tool that allows multiple people to design, comment, and iterate on the same design project.

Figma saw that design isn’t only about designers, but about how entire teams collaborate around the design workflow. Designers, product managers, engineers, even execs. Unlike traditional design software, it treats design as a collaborative, multiplayer activity rather than a solo task.

That shift from individual productivity to organizational collaboration created Figma’s advantage by driving virality and network effects within and across teams.

The more designers used Figma, the more they pulled in non-designers. And the more non-designers used it, the more value it had for designers as feedback got faster and decisions were made sooner. This became Figma’s compounding engine, as its usage expanded across entire companies and across networks of collaborators.

This ‘collaboration drove network effects’ story is the well-worn story used to explain Figma’s success.

Adobe and Figma - divergence on the cloud

Both Adobe and Figma have built large sucecssful businesses on the cloud. But their approach couldn’t be more different.

Adobe moved from selling a box of software with one‑time licenses to selling subscriptions, but the mental model stayed the same: a powerful tool for a single user. The primary customer, for all purposes, remained the same - the designer.

Instead of merely delivering design software over the cloud, Figma reimagined design from a single-player execution-oriented activity to a multi-player coordination-oriented activity.

All these differences are non-trivial. And execution is important.

But it still doesn’t quite explain why it wasn’t just difficult - but impossible - for Adobe to copy Figma.

Adobe’s software was architected in a world of desktop software. Figma’s software was architected for the cloud.

In other words, Figma is cloud-native, Adobe is not.

That doesn’t mean Adobe didn’t build a business on the cloud. It did - and a very successful one at that. However, it ported the architecture of its desktop business and layered it onto the cloud.

In other words:

Adobe used the cloud as a new distribution channel and a new revenue model. No small feat - and not the sign of a slow-moving incumbent struggling to execute.

So all those tropes used to explain Adobe’s inability to ‘do a Figma’ don’t quite make sense.

Figma, instead, reimagined every aspect of its business around the capabilities of the cloud.

This is where things get interesting.

Let’s dig in.

Get the book

If you’d like to get daily updates on lessons from Reshuffle, follow me on LinkedIn where I post ideas from the book regularly.

How Figma built a cloud-native business

The most important difference between Adobe and Figma is the one that is also least understood.

The most important difference is in how the two companies view the very logic of design work.

This might seem like an insignificant difference but ends up changing the nature of business models each company can support, the stakehodlers it can serve, and even the structure of the larger industry around them as well as the basis on which other industry players collaborate and compete.

Adobe’s shift to the cloud

Take Adobe’s file-based architecture.

When Adobe moved its Creative Suite into the cloud, it didn’t rethink its core assumptions about how work happened. Photoshop and Illustrator were still built around the same atomic unit: the file.

Designers opened a PSD, worked in layers, saved, and sent it off, repeating the process every time someone else needed to make edits. Files lived on individual machines, and sharing meant duplication. Each round of revisions spawned new versions, which needed to be tracked and reconciled manually. This file‑centric architecture defined how Adobe’s products and business were structured.

Technically, that approach imposed hard limits. Adobe’s products assumed that a single designer would work on a file at a specific stage in the workflow. Simultaneous collaboration was not possible.

Figma’s shift to the cloud

The cloud introduced four architectural enablers that Adobe’s model didn’t fully exploit.

1. First, always-on connectivity ensured that design assets could reside on the network, rather than on a hard drive.

2. As a result, one version of the truth could exist on a server, accessible simultaneously by everyone. This single source of truth enabled collaboration.

3. Collaboration was further strengthened by moving most of the processing to the cloud, enabling users to collaborate in real-time even if they had low processing power on their devices.

4. Finally, thanks to an API‑driven architecture, individual design elements could be stored and referenced dynamically, rather than frozen in a single monolithic file.

Figma reimagined all of design work around these four capabilities of the cloud.

It replaced the file with the element - a button, icon, or type style - as the basic unit of work. The ‘design document’ existed only on Figma’s servers. Instead of saving and sending files, users accessed a shared space directly.

Because everything lived in the cloud, Figma reimagined design for collaboration. Multiple stakeholders on a project, such as designers, product managers, and marketers, could be working on the same file simultaneously, seeing changes as they occur, without worrying about whose version is the latest.

Changes and permissions could be tracked and managed at the level of a design element. Each element was addressable in a database: change a component once and that change propagated everywhere it appeared. Permissions replaced ownership; engineers, product managers, and marketers could view or comment without being sent anything.

The shift from ‘file’ to ‘element’ as the atomic unit of work had another important effect. Because of the element-based architecture, Figma users could create shared libraries of reusable design components, like buttons, icons, type styles, and color palettes, that teams could use across multiple files and projects. Instead of duplicating these elements in each file, designers simply reference a single source of truth.

This creates consistency, simplifies updates (change once, update everywhere), and enables cross-functional teams to work with aligned visual standards. Shared libraries shift design from isolated file ownership to coordinated, system-level collaboration.

This architecture created strategic separation from Adobe. Adobe used the cloud to deliver the same file‑based logic more efficiently. Figma used the cloud to replace that logic entirely.

By shifting the unit of work from file to element, Figma enabled real‑time collaboration, created a shared design environment that expanded who could participate, and made Adobe’s model feel increasingly constrained by its own architecture.

How Figma restructured an entire industry

The shift from file to element delivered three critical blows to Adobe.

1. WORK: It broke Adobe’s control over the unit of work

Files reflected individual work, handed off between collaborators in sequential stages. Figma upended this by unbundling the file into its constituent elements.

This unbundling turned design into a live system, rather than a sequence of files.

It enabled co-editing, granular permissions, and instant updates across all design artifacts. The value shifted from discrete deliverables to ongoing coordination.

Unbundling changes the unit, which in turn changes architecture. In music, the album was once the atomic unit because the architecture of vinyl and CDs forced songs into bundles. With MP3s, the atomic unit shifted to the song. That simple shift enabled entirely new architectures of distribution: Napster, iTunes, and eventually Spotify.

Image source: Reshuffle

2. ORGANIZATION: It shifted value and power from execution to governance

In the old world, the unit of value was the act of execution. Who could design faster, deliver cleaner, export assets better.

But with shared elements and real-time updates, execution is commoditized. What matters more is managing permissions, rights, auditability, and collaboration across a multi-player workflow. Governance becomes the new source of leverage.

Design teams can now design coherently at scale. This requires governance: control over how components are created, reused, evolved, and deprecated.

When value shifts from execution to governance, the organizational budgets that pay for your software also change.

In execution-led tools, software is paid for by the teams doing the work - designers, engineers, marketers - because the tool helps them complete specific tasks faster. But in governance-led systems, value comes from managing consistency, control, and coordination across the organization. The budget often shifts upward, because the tool becomes strategic infrastructure, rather than just a productivity aid.

3. INDUSTRY: It eroded Adobe’s closed-loop power and changed industry structure

Once elements were unbundled from the file and individually addressable, they could be packaged into shared libraries, which acted as central sources of truth reused across files and teams.

This restructured the industry from siloed, file-based workflows to interoperable design ecosystems.

When Figma moved design from static files to dynamic elements, it reshaped the structure of the ecosystem. In traditional file-based systems, value was created and captured inside closed loops: files lived on local drives, changes were tracked by humans, and tools were optimized for ownership and execution. The dominant logic was self-contained workflows: a designer edited a file, exported assets, and handed them off, often using proprietary formats inside siloed tools.

But element-level architecture unbundles the design process into modular, reusable pieces. This naturally dissolves the boundary between inside the tool and outside the tool. With components living in shared libraries and third-party tools can plug into atomic design elements through APIs, interoperability was invevitable.

This shift fractured the vertically integrated model Adobe had dominated. Just as the modular web displaced proprietary desktop software, Figma’s architecture enables a loosely coupled, composable ecosystem of tools. integrating at the level of individual design elements. Value no longer accrues to those who own the file, but to those who coordinate the system, through reusable design tokens, shared standards, and governance mechanisms.

So when people say Figma succeeded because of collaboration, they’re not wrong. But they’re also not getting at what really drove Figma’s success.

The real transformation lies in how element-level coordination reshaped the structure of work, the boundaries of the ecosystem, and the source of strategic control.

Applying Figma’s lessons today

Figma delivered three shocks to Adobe’s architecture:

• The unit of work changed

• The nature of the organizational system changed

• The structure of the competitive ecosystem changed

In my book Reshuffle, I explain how the transformative effects of technology play out across the entire system of work:

Consider, for instance, how an AI-native legaltech firm drives changes at all three levels:

Image source: Reshuffle

Much like Figma did with the cloud, companies leveraging AI are poised to trigger a similar structural shift.

Where Figma unbundled the file, AI unbundles today’s human-dominant knowledge workflows into tasks, which can be recombined into fundamentally new workflows.

This eventually changes how work is structured, how workflows and orgs are organized, and how companies compete.

Yet, most players today are doing what Adobe did - they are slapping AI onto the old logic without rearchitecting their business around what AI makes newly possible.

That is the untold story of Figma’s success - and that is the lesson that will remain lost on companies that merely interpret Figma as a beneficiary of collaboration software alone.

Get into the weeds with Reshuffle

If you’ve enjoyed reading this post, you should check out my new book Reshuffle.

Get the book

If you’d like to get daily updates on lessons from Reshuffle, follow me on LinkedIn where I post ideas from the book regularly.

Follow for daily updates

Get the book

People often ask me about my favorite books on AI.

Now that Reshuffle is out, the question has shifted to "Which books inspired it?"

The honest answer is almost ironic.

I can't name a single book on AI that served as inspiration. I wasn’t looking for more predictions of the "This changes everything" type.

I was looking to understand the deeper, enduring structural forces that shape how economic systems evolve and how firms are organized.

So the five books that most influenced my thinking aren’t conventional 'AI books' at all.

They talk about structure, systems, tensions, and power.

If read closely, these books give you the sharpest lens on how AI will change work and business.

Through this newsletter, I’m sharing the 5 books that I believe are most essential to understanding how jobs, organizations, and competitive ecosystems change in response to a new technology like AI being introduced.

Add this to your summer reading list.

Some of these are pacy and entertaining, some are slow. But all of them are essential reading.

And if you haven’t picked up Reshuffle already, now is a good time!

Get the book

Now on to the list…

1. The Content Trap - Bharat Anand

Perhaps the single best book on the economics of digital media.

I particularly like Anand's discussion on the importance of complements, and that's informed my own 'theory of change' on new tech.

Anand explains why content on its own rarely creates lasting power, and why the real leverage comes from the complements that surround it: the networks, experiences, and ecosystems that make the content valuable in the first place.

It’s not enough to ask what the shiny new tool can do; you have to ask what it makes newly essential, newly scarce, or newly possible.

The Content Trap rethinks how you approach strategy in the digital age. The book didn’t do as well as it should have - the name probably didn’t help.

2. Design Rules - Carliss Baldwin and Kim B. Clark

A dense, demanding book, but absolutely essential.

It outlines how modularity and architecture influence the evolution of complex systems. Written with a narrower lens at a different time, the principles travel far: once you understand modularity, you see how AI will restack industries.

Not easy summer reading though, it’s the kind of book you wrestle with.

Baldwin and Clark dissect modularity and system architecture, showing how the way a system is carved into parts determines how it evolves over time.

Once you absorb their framework, you notice how companies carve up value chains, how industries fracture and recombine, how control points shift when the rules of design change.

3. Clockspeed - Charles Fine

Written before the 'digital era', let alone the AI boom, this book is more important than ever today.

Clockspeed explains how different parts of a value chain evolve at different speeds, and how that creates hidden tensions. Companies get locked into slow-moving layers while competitors dance across faster ones. Strategies that make sense at one speed suddenly break when another part of the system evolves at a madly faster rate.

Many of my arguments regarding the friction between AI tool providers and solution providers are built directly on Fine’s framework.

Even though Fine was writing in the late ’90s, his framework feels eerily current. Once you see the world this way, every industry looks like a set of interlocking gears, some spinning fast, others grinding slow. With the rapid pace of AI development attacking slow-moving sectors like professional services, Clockspeed is more essential than ever at helping explain how these industries will collide and evolve.

4. The Goal - Eliyahu Goldratt

Often dismissed as ‘business school folklore,’ but one of the sharpest books ever written about constraints.

Goldratt’s core idea is simple yet radical: every system has a single point that limits its performance, and everything else is noise until that constraint is addressed. It’s not about pushing harder everywhere; it’s about knowing where to push, and when to stop.

Much like Design Rules and Clockspeed, the book may have been written with a narrow lens, but its principles travel far - in this case, way beyond operations and manufacturing.

AI is supposed to transform workflows and organizations, but before you can redesign them, you have to understand why they’re structured the way they are today i.e. what hidden bottlenecks they’ve been built around.

Goldratt’s classic provides that foundation.

5. The Box - Marc Levinson

This history of the shipping container might seem like the oddest choice of all, but it was the spark for Reshuffle.

I have long felt that the discussion on 'intelligence' fundamentally misses what matters in the impact of tech on systems.

That is why the introduction to my book is titled Why unintelligent AI matters.

Levinson shows how dumb tech like the container transformed the global economy by reconfiguring systems around it.

If you really want to understand how AI will transform firms, don’t read the latest AI manifestos.

Read these five books.

They offer enduring principles about complements, modularity, constraints, and system redesign - principles that explain not only why past technologies reshaped industries, but how AI will do it again.

Reshuffle - Get the bonus chapter

Have you had a chance to read Reshuffle yet?

I’d love to hear your thoughts and reactions. I’m also looking to launch a companion guide to Reshuffle later this year to help readers apply these concepts to their work and strategy. I would love to hear the questions that emerge as you read the book and what you’d like to see in the companion guide.

Also, if you’ve enjoyed reading Reshuffle, I’d be grateful if you left a review on Amazon.

As a thank you, please reach out to tripta@platformthinkinglabs.com with a link/screenshot of the review, and we’ll send you an exclusive bonus chapter on Agentic Competition.

On one side, there’s the anxious chorus of doomsayers, warning that automation will eliminate jobs and render people obsolete.

On the other side, the techno-optimists appear, brushing off the fear and proclaiming that innovation is a tide that lifts all boats. History, they’ll remind you, is full of moral panics about machines replacing people, and yet, look around, the economy’s bigger than ever.

There’s a problem with this polarized debate - both sides are having the wrong conversation. They’re locked in a tug-of-war over whether the pie grows or shrinks while missing what’s really happening: AI, like many other technological breakthroughs before it, grows the pie, but the growing pie isn’t sliced in a way that everyone gets their fair share.

A wider lens to understand the impact of AI

We tend to treat economic growth and inequality as separate issues, as if one belongs to macroeconomics and the other to ethics. But they’re deeply intertwined.

A growing system does not mean shared prosperity. In fact, growth and inequality often rise together, because the very mechanisms that expand the pie also determine how it’s divided.

Today’s polarized debate on AI is rooted in task-centric framing. It frames the problem in terms of job loss or productivity gains, but overlooks the more significant shift.

AI doesn’t just change what we do. It changes how systems are restructured and who has control over them.

To clarify this, let’s map the landscape, not as two opposing viewpoints, but as four distinct positions based on how the pie grows and how it’s sliced.

The Luddite fallacy of machines taking over jobs in a fixed system sits at the bottom-left. It’s a fear that’s easy to mock in hindsight, but it persists because even with system shifts, people do lose out. Entire communities built around a specific skill or workflow can vanish in a generation.

Scoot over to the bottom-right, and you have the techno-optimist fantasy that innovation will expand the pie for everyone. The myth that AI will only make the world a better place for all is hyped incessantly by startups selling a vision of future productivity. However, a closer look reveals rising evidence of wage stagnation, power concentration, and growing inequality, which tells a different story.

In the top-left quadrant, we find task-based misdirection. People sense that work is changing, but still treat the pie as fixed. They believe jobs will look different because AI will take over tasks, not because the system itself is changing. In this view, humans govern and supervise AI while it executes at scale, but the underlying architecture of value creation stays the same. It’s a comforting simplification, but also a trap.

This brings us to what’s really happening: the pie is growing, but not everyone will benefit from the gains. This is the quadrant of rebundling, where entire industries are being rebuilt around new logic.

All three other quadrants still run with the task-centric framing. Only in this quadrant is AI seen as the engine for a new system.

In this world, success is no longer determined by getting better at playing yesterday’s game using AI. It’s defined by whether you’re playing the right game.

Former giants collapse, not because they didn’t adopt the tools, but because they adopted them into old systems.

Kodak went digital, and Barnes and Noble had a website and an e-reader. In the end, it didn’t really matter.

Most debates about AI still live in the bottom half of the grid: Will jobs vanish? Will productivity rise?

The real question is up and to the right: Who defines the new rules? Who captures the upside? And who’s holding the knife when the new pie gets sliced?

The growing pie doesn’t get shared equally. It gets divided by those who control the means of coordination.

This is the fundamental idea of a reshuffle. New winners and losers emerge, not despite a growing pie, but precisely because of it. 

Reshuffle

The above is an excerpt from my newly launched book Reshuffle.

BUY THE BOOK

No, not another book on AI?

Reshuffle is as much a book about AI as the Dark Knight is a movie about Batman.

As I mentioned in the last post, the Joker drives the movie plot, escalates every central turning point, and forces all the other characters to react. The film’s most profound questions, of chaos versus order, moral compromise, and the fragility of social contracts, are posed by the Joker.

Unlike traditional superhero films, this movie doesn’t build toward the hero’s triumph, but rather toward the moral cost of survival in a world shaped by the Joker.

The real story of AI’s impact lies less in understanding when a certain technology hits a certain performance benchmark and more in understanding how economic systems change in response to new technology. Systems, not technology, determine the story.

AI’s impact is often framed in narrow terms - what work it replaces, what jobs it threatens, and how it boosts productivity.

However, the real impact of AI comes not from how it performs a task,
but from how it restructures the entire system around that task.

If we stay trapped in task-based thinking, we remain stuck in one of the 3 quadrants above.

Reshuffle moves the discussion into the fourth quadrant and provides a comprehensive framework to explain
(1) How technology changes the rules of competition
(2) How organizations transform in response to (1)
(3) How jobs transform in response to (2)

Conversely, it also explains
(1) How individuals and teams leverage new tech to innovate
(2) How workflows and orgs change as a result
(3) How some firms accordingly rewrite the rules of competition in their favor.

A combination of these inside-out and outside-in shifts transform the overall system.

An overview of the book

Instead of dwelling on AI’s technological advancements, this book focuses on the systems into which AI is introduced - jobs, customer journeys, workflows, organizations, and even entire competitive ecosystems - and how they must evolve to exploit AI’s capabilities fully.

Central to this approach is the framework of unbundling and rebundling: breaking down existing systems into their components and reassembling them in new ways to create value. 

To explore this transformation, the book is structured in three sections that follow the arc of AI’s impact, from

(1) How we think about it, to

(2) How it changes how we work, to

(3) How it restructures competitive advantage and power.

We begin by reframing the mental models through which we understand AI.

Reshuffle - Section 1

The chapters in Section 1 demonstrate that AI is not merely a tool of automation or substitution, but a force that transforms the economic logic of any system by changing how value is created and distributed, and how these changes reallocate power. 

Reshuffle - Section 2

In Section 2, we zoom in on the world of work and the opportunities and limitations introduced with AI. We look at how AI restructures organizations and the roles and workflows within them. We also examine the opportunities it creates, not in terms of new jobs, but in terms of new value. Many tasks that once commanded high value may decline in relevance, while new points of leverage emerge elsewhere in the system.

In a system that’s constantly shifting, it’s a mistake to keep looking for the next job title once old ones no longer apply. The real opportunity for workers is to track where value is headed and position themselves in its path.

Reshuffle - Section 3

Finally, we scale out to the level of ecosystems and economies. In this section, we examine how AI restructures competitive advantage, how it elevates new winners and eliminates previous ones, how it changes erstwhile partners to competitors and makes friends out of yesterday’s foes, and how it changes the very logic of an industry. 

This is not a book about technology. There are many ways to discuss technology, but writing a static book that tries to capture the specifics of a rapidly changing field isn’t the most helpful approach.

Instead, this book focuses on something far more enduring. It examines how technology shapes the systems around us and changes who wins and who loses every time a significant technological shift occurs.

By tracing this arc, from how we frame AI to how we reimagine work and rethink power, I aim to offer a practical lens for understanding how AI changes economic systems as a whole, irrespective of which specific technologies improve or fade.

Your move

Get your hands on the book. And hit me back with your thoughts and reactions.

If you enjoy reading the book, do drop in a review on Amazon that helps others make an informed decision. Or share about it further on social media to spread the news further.

As a thank you, please reach out to tripta@platformthinkinglabs.com with a link/screenshot of the review, and we’ll send you an exclusive bonus chapter from the book, on agentic competition.