Future Enterprise in the Era of AI

Why the AI-era enterprise collapses into two departments, Building and Selling. How cheaper building and near-free paperwork shorten the hierarchy, multiply the number of companies, and turn every employee into an AI-driven generalist with one deep specialization.

Issue
18/ 19
Reading
16min
Figures
5
Equations
6

Eryawan Presma Yulianrifat, Bryan Jeshua

  • economics
  • startup
  • notes

The shape of a company is a record of what used to be expensive. For a century, two things were: coordinating people, and processing paper. So firms grew tall, to route decisions through people, and wide, to staff the departments that process paper. AI makes both cheap at the same time. This post works through what the enterprise looks like once it does, using the economics of why firms exist rather than predictions about which jobs disappear.

Committed 2026-11-08: the AI-era enterprise has two departments, Building and Selling; its hierarchy shrinks to two or three layers; and the number of companies grows faster than the number of workers, because building and paperwork both get cheap.


Principles

Six lines that summarize the argument.

  1. A firm is whatever is cheaper to do inside than to buy outside. Change the two costs and the firm changes shape.
  2. AI cuts both costs, but not equally. Paperwork falls furthest, and paperwork is most of what it costs to deal with the market.
  3. Paperwork is a tax on existing. When the tax falls, companies multiply.
  4. Two departments survive. The one that makes the thing, and the one that gets it bought.
  5. Layers exist to route problems upward. When answers are cheap to find, fewer problems need routing.
  6. Generalist people, specialist agents. Division of labor deepens, but between a person and their agents rather than between people.

Why Firms Exist

In 1937 Ronald Coase asked a question economics had skipped: if markets allocate resources so well, why does anyone build a firm, inside which resources are allocated by command instead of by price? His answer was that using the market has a cost. Before every transaction someone has to find a counterparty, negotiate, write a contract, check that it was honored, and enforce it when it was not. A firm replaces a run of those transactions with one employment contract and a manager’s instruction.

Citation: R. H. Coase, “The Nature of the Firm,” Economica 4(16), 386–405, 1937.

Command is not free either. Every activity a firm brings inside adds to what its managers must coordinate, and coordination gets harder as the firm grows: more people to inform, more decisions to reconcile, more room for an instruction to be misread on its way down. So a firm keeps absorbing activities until the cost of organizing one more inside equals the cost of buying it outside. Write Cin(q)C_{in}(q) for the total cost of coordinating qq activities internally and TT for the cost of one market transaction. The boundary q∗q^{*} sits where

∂Cin∂q∣q∗=T\frac{\partial C_{in}}{\partial q}\bigg|_{q^{*}} = T

Everything to the left of q∗q^{*} is done by employees. Everything to the right is bought from other firms.

Oliver Williamson later sharpened which activities land on which side. An activity stays inside when it needs assets built for one specific relationship, when it recurs often, and when its outcome is hard to specify in advance, because those are the conditions under which a contract is expensive to write and easy to exploit. An activity goes outside when it is generic, occasional and easy to describe.

Citation: O. E. Williamson, Markets and Hierarchies: Analysis and Antitrust Implications, Free Press, 1975.

This is the frame for the rest of the post. AI does not change why firms exist. It changes both sides of the equation above, and the firm’s shape follows from which side moves more.

Where the firm's boundary sits schematic, not to scale activities done inside the firm (q) marginal cost q₀* q₁* boundary moves in internal coordination, before AI after AI market transaction, before AI market transaction, after AI
Coase's boundary. A firm grows until coordinating one more activity inside costs as much as buying it outside. AI lowers both costs, but it lowers the market's cost further, because that cost is mostly paperwork, so the crossing point moves toward a smaller firm.

What AI Makes Cheap

Two costs fall at once, and they fall for different reasons.

Building Gets Cheaper

The cost of turning an idea into a working product is mostly the cost of skilled hours: writing the software, configuring the systems, designing the interface, wiring the operations. Coding agents, generated designs and automated operations cut the hours per unit of output. A product that once needed a team of eight for a year can be prototyped by two people in weeks, and a prototype that works is most of what an early company needs to find out whether anyone will pay for it.

This is the part of the story most people already believe. It is also the smaller part, because building was never the main reason companies were expensive to run.

Paperwork Gets Cheap

Every company pays a tax for existing, whether or not it sells anything. It has to be incorporated. It has to keep books, file taxes, run payroll, register for licenses, write and read contracts, answer auditors and regulators, and keep records that prove it did all of the above. None of this makes the product better. All of it is text: forms, ledgers, contracts, filings, correspondence. And text is exactly what language models are good at producing, checking and reading.

This tax is also what Coase’s market cost TT is made of. Finding a supplier, negotiating terms, drafting the contract, checking the invoice against it, chasing the late payment: each step is a document. When AI makes the documents cheap, it makes the market cheap, which is why Fig. 01 shows the market line falling further than the coordination curve. Coordination inside a firm still involves people who must trust each other. A purchase order does not.

The Cost of Entering a Market

The cost of starting a company and keeping it alive long enough to compete can be written as a fixed entry cost with four parts:

F=Fbuild+Fpaper+Fcapital+FdistF = F_{build} + F_{paper} + F_{capital} + F_{dist}

where FbuildF_{build} is the cost of producing the first version of the product, FpaperF_{paper} is the administrative cost of existing, FcapitalF_{capital} is the physical capital the business needs (machines, inventory, premises), and FdistF_{dist} is the cost of reaching the first customers and earning their trust. AI shrinks the first two directly. It barely touches the third, because a delivery fleet costs what it costs, and it touches the fourth least of all, because trust is earned by people over time.

How much does a lower FF matter? In Steven Salop’s model of competition around a circular market, with LL customers, a cost tt that measures how much customers care about getting a product close to what they want, and a fixed cost FF per firm, free entry settles at

n∗=t LFn^{*} = \sqrt{\frac{t \, L}{F}}

firms. The number of firms rises as entry gets cheaper, and it rises with the square root: halving FF multiplies n∗n^{*} by 2≈1.41\sqrt{2} \approx 1.41, and cutting FF to a tenth multiplies it by 10≈3.16\sqrt{10} \approx 3.16. These are outputs of the model, not forecasts. The point is the direction and the shape: cheaper entry means more firms, each serving a narrower slice of the market more closely.

Citation: S. C. Salop, “Monopolistic Competition with Outside Goods,” The Bell Journal of Economics 10(1), 141–156, 1979.

The cost of starting a company, by part schematic, not to scale Build Paperwork Capital Distribution Capital Distribution build, shrunk paperwork, shrunk unchanged unchanged, and now the largest part Before AI After AI F = F_build + F_paper + F_capital + F_dist
Entry cost by part. AI shrinks the build and the paperwork, leaves physical capital alone, and does least to distribution, which is what is left to compete on.

Fig. 02 carries the consequence that the rest of this post builds on. When building and paperwork are cheap, the part of a company that is still expensive is getting customers to trust it. That is what the two-department enterprise is organized around.


Two Departments: Building and Selling

Michael Porter’s value chain splits what a firm does into primary activities, the ones that make and deliver the product (inbound logistics, operations, outbound logistics, marketing and sales, service), and support activities that keep the primary ones running (firm infrastructure such as finance, legal and planning; human resources; technology development; procurement). The typical org chart is the value chain with a manager on each box.

Citation: M. E. Porter, Competitive Advantage: Creating and Sustaining Superior Performance, Free Press, 1985.

Look at the support activities through the lens of the previous section. Finance is ledgers and filings. Legal is contracts and compliance. Procurement is requests for quotation, purchase orders and invoice matching. Human resources is offer letters, payroll, leave records and policy. Most of what these departments produce is the paperwork that AI makes cheap. They do not disappear, because the obligations behind them do not disappear: taxes are still owed, contracts still bind. But they stop needing a department of people. They become a layer the company runs on, the way it runs on electricity and a bank account.

What is left are the primary activities, and they fold into two departments.

Building makes the thing and delivers it: product, engineering, design, operations, fulfilment. Porter files technology development under support; in a company whose product is increasingly software and automated operations, it is the core of Building.

Selling gets the thing bought and keeps it bought: marketing, sales, partnerships, customer service. Selling is where the irreducible human work concentrates. A customer who hands over money, data or a critical workflow is extending trust, and trust is extended to people and to brands built by people. AI makes outreach, proposals and support replies cheap, which means everyone’s outreach is cheap, which means the outreach that works is the one backed by a relationship.

This is the same conclusion as the earlier post on software eating services: when AI makes the work cheap, distribution and trust are the moat, not the model. The two-department enterprise is what a company looks like when it organizes around that fact.

Before: one department per function After: two departments on an AI layer CEO HR Finance Legal IT Procurement R&D Operations Marketing Sales Service support primary Founder / GM Building product engineering, design operations, delivery Selling marketing sales, partnerships customer service AI layer HR · finance · legal · procurement · IT · paperwork support functions become infrastructure
The org chart before and after. The five support functions are mostly paperwork, so they move into an AI layer the company runs on. The five primary functions fold into Building, which makes and delivers the product, and Selling, which wins and keeps the customer.

Shorter Hierarchies

Two departments say what the company does. They do not say how many layers of managers sit between its leader and the person doing the work. For that, the most useful model is Luis Garicano’s theory of hierarchies as a way of organizing knowledge.

Citation: L. Garicano, “Hierarchies and the Organization of Knowledge in Production,” Journal of Political Economy 108(5), 874–904, 2000.

In Garicano’s model, work is a stream of problems. A frontline worker learns to solve the common ones, and learning costs something, so nobody learns to solve everything. When a worker meets a problem they cannot solve, they pass it up to someone who knows more. That person spends time on every problem passed to them. If a worker solves a fraction zz of the problems they meet, and a manager spends time hh on each problem passed up, then with a unit of time to spend a manager can support

s=1h (1−z)s = \frac{1}{h \, (1 - z)}

workers. This is the span of control. It widens when workers solve more on their own (zz rises) and when each escalated problem takes less of the manager’s time (hh falls).

A model only earns its place if the evidence agrees. Bloom, Garicano, Sadun and Van Reenen tested the distinction on American and European manufacturing firms and found that it does: technologies that make information cheaper to acquire, such as ERP systems for plant managers and CAD/CAM for production workers, went with more autonomy and wider spans of control, while technologies that only made communication cheaper, such as data intranets, went with less autonomy, because they make it cheap to send every decision upward.

Citation: N. Bloom, L. Garicano, R. Sadun, J. Van Reenen, “The Distinct Effects of Information Technology and Communication Technology on Firm Organization,” Management Science 60(12), 2859–2885, 2014.

AI is mostly the first kind. An agent that answers a worker’s question on the spot raises zz, because the worker can now solve problems they never learned to solve. An agent that drafts the analysis a manager would have written lowers hh. Both push the span wider.

A wider span is a shorter hierarchy. With span ss at every level, the number of reporting layers ℓ\ell needed to connect one leader to NN people at the front line is the smallest ℓ\ell with sℓ≥Ns^{\ell} \geq N:

ℓ=⌈log⁡sN⌉\ell = \left\lceil \log_{s} N \right\rceil

For a company of N=1,000N = 1{,}000: at a span of 6, it takes four layers, because 63=2166^3 = 216 falls short and 64=1,2966^4 = 1{,}296 does not. At a span of 32, two layers are enough, because 322=1,02432^2 = 1{,}024. These numbers are a worked example of the formula, not a measurement of any company. What they show is that the relationship is logarithmic: the span does not need to grow much for whole layers to disappear.

The same 1,000 people, two spans of control worked example, not data 1 2 3 4 1 2 4 reporting layers span 6: 6⁴ = 1,296 ≥ 1,000 2 reporting layers span 32: 32² = 1,024 ≥ 1,000
Layers grow with the logarithm of headcount, so a wider span removes whole layers. AI widens the span by letting workers solve more problems on their own and by cutting the time a manager spends on each one passed up.

The layers that go are the ones whose work was routing: collecting status from below, summarizing it for above, passing approvals down, reconciling what one team said with what another team did. Those tasks are reading and writing, the same material as paperwork. The judgment a good manager adds, deciding what matters and who should own it, does not go away. It is spread over more people, by fewer managers.


Many More Companies

Put the three previous sections together. The Coase boundary moves in, so each firm does less inside itself. The entry cost falls, so more firms enter. The hierarchy flattens, so a small firm does not need to grow headcount to grow output. The prediction is that companies get smaller and more numerous: the number of companies grows faster than the number of workers.

The evidence so far is suggestive and nothing more. In the United States, applications to form a new business, counted by the Census Bureau, rose from about 3.50 million in 2019 to about 5.67 million in 2025, a rise of roughly 62%, and 2025 was the highest year in the series. Two cautions apply. An application is not an operating business, and most of these are companies with no employees. And the jump started in 2020, two years before generative AI was a consumer product, so it cannot be credited to AI. What it does show is that the floor under company formation was already moving, and the costs that AI cuts are the ones that held it down.

Citation: U.S. Census Bureau, “Business Applications: Total for All NAICS in the United States” (BABATOTALNSAUS), Business Formation Statistics, via FRED; annual totals summed from monthly, not seasonally adjusted data, retrieved Nov 2026.

More companies does not mean every company stays small. The likely shape is a split. At one end, a long tail of very small firms, many run by one to five people, each serving a narrow market closely, the outcome Salop’s model predicts when entry gets cheap. At the other end, a few very large platforms, because the AI layer that every small firm runs on has strong economies of scale: models, compute, payments, distribution marketplaces. The middle, firms large enough to need many departments but too small to be platforms, is the part under most pressure, because its main advantage was absorbing the overhead that the AI layer now absorbs for everyone.

How many firms exist at each size illustrative, not data number of firms firm size, employees (log scale) 1 10 100 1k 10k+ more tiny firms before AI a thinner middle a few larger platforms
The predicted shift, drawn to show its shape rather than any measured values: many more very small firms, fewer mid-sized ones, and a handful of larger platforms that the small firms run on.

There is a second consequence for how work is organized across companies. In a world of many small firms, much of what used to be a department becomes a contract with another small firm. A company does not need a design team if a three-person design firm, itself running on the AI layer, will take the work on a contract that AI made cheap to write and cheap to enforce. The firm boundary from Fig. 01 does not just move in; the space outside it fills with other firms doing the same thing.


The All-Round Employee

If the company has two departments and two layers, what does a single person in it do?

Edward Lazear’s theory of entrepreneurship is a useful starting point. He argued that founders are jacks of all trades: running a business means being adequate at many things rather than excellent at one, so the people who start companies tend to have balanced skills, and in his data people whose careers had spanned more different roles were more likely to become entrepreneurs.

Citation: E. P. Lazear, “Entrepreneurship,” Journal of Labor Economics 23(4), 649–680, 2005.

The AI-era employee looks like Lazear’s entrepreneur. In a two-department firm with a short hierarchy, each person owns an outcome from end to end rather than a step in a process: a product area, a customer segment, a market. Owning an outcome end to end means touching design, engineering, operations, pricing, the customer conversation and the paperwork, which is to say being a generalist.

There is an obvious objection. Gary Becker and Kevin Murphy showed that how far a society divides labor is limited not only by the size of the market, as Adam Smith argued, but by the cost of coordinating the specialists. Lower the coordination cost and specialization should deepen, not retreat.

Citation: G. S. Becker, K. M. Murphy, “The Division of Labor, Coordination Costs, and Knowledge,” The Quarterly Journal of Economics 107(4), 1137–1160, 1992.

Both are right, and the resolution is where the specialization happens. AI lowers coordination costs, and the division of labor does deepen, but it deepens between a person and their agents. The agents are the specialists: one drafts the contract, one writes the code, one reconciles the ledger, one answers the support ticket. The person is the generalist who decides what each should do and whether what came back is right.

That last clause is the binding constraint. A generalist can only delegate what they can check. Suppose a person delegates work in several domains kk, an agent produces output yky_k in each, makes a serious error at rate eke_k, and the person catches an error with probability pkp_k, which depends on how well they understand that domain. As a sketch, the value that survives review is

Y=∑kyk(1−ek (1−pk))Y = \sum_{k} y_k \bigl(1 - e_k \, (1 - p_k)\bigr)

Breadth raises the number of domains a person can direct. Depth raises pkp_k in the domains where an uncaught error is costly. This is why the shape is a generalist with one specialization rather than a pure generalist: broad enough to direct agents across the whole outcome they own, and deep enough in one domain to be the person whose judgment the company relies on there. The specialization is what makes them trustworthy; the breadth is what makes them productive.


What This Means

For founders. Start with two departments and keep them. Every support function you are tempted to staff should first be asked whether it is a department or a layer. Spend the money you save on Selling, because Fig. 02 says that is where the remaining cost of competing sits. And plan for the company to stay small longer than instinct suggests: in a flat, agent-heavy firm, output can grow a long way before headcount has to.

For employees. The safest position is not the deepest specialization, which agents will imitate, nor the broadest generalism, which cannot check its own agents. It is a broad base that lets you own an outcome end to end, with one domain deep enough that your review is worth something. Learn to direct agents the way a manager directs people: say clearly what you want, and know how to tell whether you got it.

For incumbents. A large firm’s overhead used to be the price of scale, paid because the alternative was worse. When the AI layer gives the same back office to a five-person competitor, overhead stops being an advantage and becomes a cost the small competitor does not carry. The defensible assets are the ones Fig. 02 left untouched: physical capital, and the distribution and trust that took years to build. Reorganizing around those, rather than around the support departments, is the work.


Open Questions


The org chart of the twentieth century was a map of what was expensive: a department for every kind of paperwork, and a layer of managers for every round of passing a decision along. AI makes both cheap at once. What remains is the work that was always the point of a company, building something worth having and persuading people to trust you with it, done by fewer layers of more capable people, in many more companies than there are now.