Wild Scale

The digital economy already taught us something peculiar about scale: a company can serve another million customers without hiring another million people. Economists call one version of this scale without mass: information technology allows firms to replicate routines at low marginal cost, so revenue and reach can grow much faster than employment. AI introduces a more aggressive possibility: cognitive scale without organizational mass.

This is part 2 of After Software.

We’ve all seen the memes: The one-person unicorn – the founder who, equipped with an army of agents, bootstraps their business to 1 billion USD. Forbes recently documented variants of this possibility: tiny teams producing big-company output, “micro-unicorns,” AI-native founders building across functions that previously required staff.

Those are examples of what I call wild scale.

Organizational scale has historically provided two powerful advantages: more hands; and more kinds of minds. AI weakens both relationships, and asks, what was that organization doing for us that makes it worthwhile keeping? Once those variables start to separate, the size of a problem stops determining the size of the organization required to attack it.

Why did we need all these people?

Organizations are not arbitrary collections of jobs. They are engines for coordinating limited human capability.

Finance knows one part of reality. Sales knows another. Operations holds the keys to yet another part of the kingdom. So does Legal. Management layers move information between functions and people. Meetings synchronize them. Processes tell people what happens next. Hierarchy decides who may act. Software gives everyone bounded tools through which their particular piece of reality can be manipulated.

Robert Grant’s knowledge-based theory of the firm describes organizations precisely in terms of integrating specialized knowledge held by different people. That architecture makes sense when knowledge is distributed and expensive to move. It looks less inevitable when an intelligence can simultaneously hold the customer, inventory, financial constraints, legal rules and strategic context in working state.

The problem no longer has to travel through the organization in quite the same way. The organization can begin assembling around the problem.

The geometry of work starts to change.

Work no longer needs to be organized primarily around who knows what, which function owns the information, or which sequence of handoffs gets it to the next person. More work can collapse around the outcome itself: relevant knowledge is assembled when needed, the path through the problem is generated dynamically, and humans are pulled in where judgment, authority or consequence still require them.

Coase gets another turn

In 1937 Ronald Coase asked why firms exist when markets can theoretically coordinate production. His answer, simplified: using markets is not free. Searching, negotiating, contracting, monitoring, enforcing – all of that consumes money, time and attention. That’s why sometimes, putting capability inside a hierarchy is cheaper. Firms expand until coordinating another transaction internally becomes less attractive than obtaining it outside.

Almost ninety years later, economists are now asking explicitly whether AI agents could produce a Coasean Singularity. Peyman Shahidi and colleagues use the term for a world in which agents dramatically reduce costs around search, matching, preference elicitation, contracting, identity and enforcement. This isn’t entirely speculative anymore. Mastercard is building infrastructure through which agents can discover merchants and transact on a user’s behalf. The German Bundesbank is already analyzing “agentic payments” in which software not only executes a payment but chooses timing, instrument and rail within a delegated mandate.

The market itself begins acquiring cognition. The future of the firm becomes ambiguous.

AI can make enormous companies easier to coordinate. It can also make capabilities outside the company dramatically easier to find and orchestrate.

Internal costs fall.

External costs fall.

Which falls faster?

Nobody knows.

Some giant institutions may grow vastly larger because their scarce advantages – capital, infrastructure, distribution, proprietary state, regulatory authority – can be coordinated with less administrative and organizational friction. At the same time, tiny organizations or individuals may acquire capabilities that previously required enormous institutions. Between them sits a lot of expensive connective tissue whose economics start looking less inevitable.

The minimum viable institution

We already have early clues for what’s about to happen. In a field experiment involving professionals at Procter & Gamble, individuals using AI matched the performance of two-person teams working without it. AI also reduced the divide between technical and commercial specialists: professionals produced more balanced solutions outside their native expertise.

One experiment does not abolish the corporation, but it illustrates the mechanism. A person gains access not merely to more output, but to adjacent capability. A larger problem may no longer require proportionally more researchers, analysts, writers, developers and coordinators.

This begs the question: What was the organization doing for us that still makes it worth keeping?

The current one-person-company conversation provides one answer: perhaps surprisingly little.

A September experiment among patent lawyers gives us a warning. AI improved immediate performance – especially for junior lawyers – but durable improvements in expert judgment after three months were concentrated among senior practitioners. Junior lawyers gained capability without showing the same average increase in expertise.

The implication is uncomfortable.

AI will not just compress organizations. The compression might inadvertently also eliminate where junior staffers traditionally learned to become experts.

Wild scale can consume expertise. So far, it lacks a consistent mechanism for reproducing it.

What does an AI-native organization actually look like?

“Headcount reduction” is too boring a frame. A plausible Wild Scale organization probably doesn’t look anything like today’s company just with fewer boxes. It’s going to be radically different boxes, new shapes and configurations.

• A small human core of principals responsible for objectives, judgment, trade-offs, relationships and consequences.
• Around them sit portfolios of agents researching, producing, monitoring, negotiating, coding, analyzing and executing across what we currently call functions.
• Beneath both sits an authoritative context layer: proprietary data, institutional memory, customer history, rules, permissions, strategy, operating state and whatever tacit knowledge can actually be made legible.
• And wrapping the entire system is a control plane defining what each human and machine may know, spend, change, commit and escalate.

This architecture is already starting to emerge across industries. ServiceNow calls its version an AI Control Tower, built to discover, govern, secure, observe and measure agents and other AI systems across an enterprise; its September release adds runtime monitoring and a kill switch. Workday has developed an “agent system of record” for non-human identities. And Fortune describes a proposed AI-native org chart that no longer primarily maps who reports to whom, but maps system access, spending authority, consultation rights and human-agent responsibility.

While the classical org chart was largely a map of people, the Wild Scale org chart might resemble something like a map of authority.

The manager as a principal of intelligence

This also points at how the human role in an organization will evolve. “Managing an army of agents”, I heard someone describe their new work reality recently. But, the hard part isn’t having fifty bots.

Somebody still has to tell the AI what to do to begin with, somebody still has to decide:

• What context may they access?
• Which representation of reality are they reasoning against?
• Which tools can they invoke?
• What authority may they exercise?
• Where are the approval boundaries?
• What constitutes rewardable behavior?
• When should the machine escalate?
• And most importantly: who remains accountable when the machine doesn’t?

The growing agent-management literature is already describing orchestration as the next bottleneck: given a goal, how do you coordinate humans, agents, tools, information and time into one reliable outcome?

There is one giant caveat, however: orchestration is not simply a function of better models.

Wild scale depends on integration, context access, institutional memory, a usable ontology of the problem, decision architecture and its permissions, proprietary and tacit knowledge dispersed across the organization, monitoring, and also: trust.

A brilliant agent disconnected from organizational reality is just an extremely articulate outsider.

The human managerial residue therefore may become less “supervising work” and more architecting systems of delegated cognition: 

• Direction
• Allocation
• Authority
• Exceptions
• Conflict
• Taste
• Judgment
• Accountability

The wild scale equivalent of today’s manager may look less like a supervisor and more like a principal: someone entrusted with an outcome, commanding a shifting portfolio of human and machine capability to produce it.

Friction worth keeping around

So far we have discussed what comes next, when the cost of cognitive friction goes to zero. We tend to describe this friction as waste. 

And yes, some of it is: status meetings that exist because nobody knows what anybody else is doing deserve little sentimental protection.

But some friction might actually be contributing to productivity, instead of clogging up work.

Drafting something badly before learning to draft it well. Or wrestling with ambiguous evidence, until you are confident in the decision you are going to take. Defending an argument. Failing at something so you can learn from it, or disagreeing with someone and gaining a deeper insight from the conversation.

These are not merely inefficient routes to output that need to get optimized away. They are mechanisms through which judgment, expertise and sometimes identity are crafted.

Recent writing on “productive friction” is starting to make exactly this apprenticeship argument. If the junior work disappears, how do junior people become senior? If machines generate competent first drafts, where do humans learn what makes the drafts competent?

Our education system may eventually have to become more intentional about producing judgment precisely because the workplace stops producing it accidentally.

There is a diversity problem too: AI-generated ideas can outperform human ideas on average quality while remaining less collectively diverse: more capability does not automatically provide more independent ideas – “adjacent capability” and “cognitive breadth” are not the same thing. If one person can simulate twelve disciplines through the same underlying models, where does genuine disagreement come from?

What will wild scale actually feel like?

There is another trade-off hidden inside all of this: We assume eliminating drudgery makes work better. Maybe. But consider what remains.

If AI removes research gathering, slides, status reporting, spreadsheet manipulation, scheduling, coordination and much routine production, human work starts concentrating around:

• deciding
• rejecting
• editing
• approving
• escalating
• taking responsibility

AI could compress a week’s worth of judgment into an afternoon.

Sounds like liberation if you love deciding. It may be cognitively brutal if you don’t.

Bigger problems.

Wild scale is therefore not fundamentally an argument for tinier companies. Recent attempts to get there too quickly are instructive. Meta’s Project OT reportedly sought to replace large amounts of organizational mass with AI agents, then pulled back after systems underperformed and operational incidents increased.

Organizations do things beyond processing cognition: they hold assets, create culture, develop talent, accumulate trust, carry liability, define authority, remember decisions and maintain relationships.

Those functions do not magically stop being foundational, just because inference gets cheaper.

Cutting the institution does not create wild scale. You have to simultaneously increase capability per unit of organization.

The firm may therefore become less of a container for labor and more of a durable architecture for context, capital, authority, accountability and memory – with humans and machines assembling dynamically around those things.

Which brings us back to scale.

For most of human history, big ambition required big organization because knowledge and capability had to be accumulated in human bodies and coordinated through institutional machinery. To tackle a sufficiently large problem, you first had to assemble a sufficiently large institution.

The one-person unicorn, micro firms and new human-agent work arrangements show that this relationship may begin to loosen.

B the prize is not doing the same work with fewer people.

It is expanding the size of the ambition we can afford to wield. And once the scale of our organizations no longer has to match the scale of the problems we want to solve, we enter the era of truly wild scale.

 

 

Sources:

• Ronald H. Coase, The Nature of the Firm (1937)
• Friedrich A. Hayek, The Use of Knowledge in Society (1945)
• Herbert A. Simon, Administrative Behavior (1947)
• Jacob Marschak & Roy Radner, Economic Theory of Teams (1972)
• Kenneth J. Arrow, The Limits of Organization (1974)
• Oliver E. Williamson, Markets and Hierarchies (1975)
• Oliver E. Williamson, The Economic Institutions of Capitalism (1985)
• Thomas W. Malone, Joanne Yates & Robert I. Benjamin, Electronic Markets and Electronic Hierarchies (1987)
• Robert M. Grant, Toward a Knowledge-Based Theory of the Firm (1996)
• Erik Brynjolfsson, Wang Jin, Georgios Petropoulos & Xiupeng Wang, Information Technology, Firm Size, and Industrial Concentration (2023; revised 2025)
• Fabrizio Dell’Acqua et al., The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise (2025/2026)
• Gillian K. Hadfield & Andrew Koh, An Economy of AI Agents (2025/2026)
• Peyman Shahidi, Gili Rusak, Benjamin S. Manning, Andrey Fradkin & John J. Horton, The Coasean Singularity? Demand, Supply, and Market Design with AI Agents (2025/2026)
• Aaron Chatterji, Daniel Rock & Eduard Talamàs, Transformative AI and Firms (2026)
• Erik Brynjolfsson & Zoë Hitzig, AI’s Use of Knowledge in Society (2026)
• Rogerio S. Victer, Economies of Cognition: A Cognitive Theory of the Firm (2026)
• David Autor et al., Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting (2026)
• Christian Terwiesch, Lennart Meincke, Karan Girotra, Ethan Mollick, Gideon Nave & Karl Ulrich, Artificial Intelligence and Its Impact on Creativity and Diversity: An Empirical Study of Large Language Model-Generated Product Ideas (2026)
• Ajay Agrawal, Erik Brynjolfsson & Anton Korinek, eds., The Economics of Transformative AI (2026)
• Mastercard, Agent Connect / AI-Powered Shopping and Agentic Commerce (2026)
• Deutsche Bundesbank, Agentic Payments: When Payments Learn to Think (2026)
• Kristin Stoller / Fortune, Here’s What the AI-Native Org Chart Could Look Like (2026)
• TerDawn DeBoe / Forbes, AI Handed Your Small Team a Big Company’s Leadership Problem (2026)
• TerDawn DeBoe / Forbes, AI Gives Small Teams Big-Company Output. The Bottleneck Is You (2026)
• Aytekin Tank / Forbes, How Founders Can Build an AI-Native Startup Today (2026)
• Lisa Curtis / Forbes, How AI Is Giving Ten-Person Teams a Billion-Dollar Edge (2026)
• Joe McKendrick / Forbes, Small One-Billion-Dollar Businesses Are Almost Here, In a Way (2026)
• Kristin Gleitsman, Productive Friction in the Age of AI: Apprenticeship Is at Risk (2026)
• James Win, The Next AI Bottleneck Is Orchestration (2026)
• Olivier Gomez, The Next AI Advantage Will Be Management, Not More Agents (2026)
• ServiceNow, AI Control Tower and September 2026 release materials
• Fortune, AI Agents Are Going Rogue. CIOs Are Racing to Put Guardrails Around Them (2026)
• Reuters, Mark Zuckerberg Had a Bold Plan to Replace Meta Staff With AI—and It Imploded (2026)