The Revenge of the Enterprise
When AI changes the economics of coordination
I spent my career navigating technological change, from the media landscape changes ushered in by the internet to the massive wireless transformation driven by mobile growth. The patterns look the same: something that began as an added capability became an assumption about how the world worked.
That experience shapes how I look at AI. The useful question is what happens after a capability becomes widely available, cheap enough to use continuously, and embedded in ordinary behavior.
Most of the conversation today starts with the individual. How much more can one person produce? How small can a company become? Could a handful of people build something that once required thousands?
Those are real possibilities. I think they also obscure a larger one.
What happens when the thousands can finally learn and act together?
The tax on being large
A large organization has more people, more customers, more relationships, and more contact with the world. It also spends an extraordinary amount of effort getting those advantages to work together.
Someone learns something useful. They have to recognize its relevance, explain it, find the right person, reconstruct the context, and persuade someone to act. The recipient may be three departments away. By the time the information arrives, the situation has changed.
We built management structures, meetings, reports, and software to manage this problem. Each helps. Each also becomes another place where knowledge can slow down or disappear.
I think of this as the coordination tax. Part of the startup advantage comes from paying less of it. A small team can hold a shared picture of the business, make a decision, and move. A larger competitor can know more in total while making worse use of what it knows.
No company wires every pair together. Managers, meetings, and reports exist to cut the count. The count still shows why each new person costs more to coordinate than the last one did. Fred Brooks made the same point about software teams in 1975.
Most of that communicating is the tax in action: finding people, explaining context, and following up. Asana’s 2023 survey of 9,615 knowledge workers found the same thing, with 58% of the day going to “work about work” rather than the work itself.
Now ask what happens if AI removes a substantial share of that tax.
Every person is a point of contact with reality
The important unit in this argument is a human being.
A person occupies a particular place in the world. They hear a customer’s hesitation. They notice that an installation works differently in one building. They recognize a workaround that never appears in a process document. They carry experience, relationships, physical senses, and judgment that the organization cannot recover simply by searching its files.
They can also act. They can ask another question, try a different approach, repair something, or change what happens next.
That makes each person both a sensor and an actor. The observation and the ability to test it belong together.
AI could connect these situated experiences at a level we have struggled to achieve. It can help turn an observation into a question, find relevant experience elsewhere, suggest a test, and return the result to people who can use it. The human stays in contact with the world. The network helps that contact become collective learning.
In a field experiment with 776 P&G professionals, AI also erased the functional split. Without it, R&D staff proposed technical solutions and commercial staff proposed commercial ones. With it, both groups produced balanced proposals. People working with AI spent 12.7% to 16.4% less time. Teams with AI were about three times as likely to produce a top-10% solution. The authors note the limits: it was a one-day exercise with two-person teams.
There is a condition here: people need a reason to contribute honestly. A system that feels like surveillance will get edited accounts of reality. Trust affects the quality of what the network can learn.
Many employees already hide how they work, and a network fed those edited accounts would learn from the edits. In a separate experiment with 200 employees, those told they were being monitored were more likely to cheat, not less.
Parallel discovery changes the arithmetic
Imagine a business with hundreds of locations. This is a hypothetical, but the mechanism matters.
One employee notices customers repeatedly getting stuck at the same step. Elsewhere, another employee has found a workaround. A third location serves customers for whom the workaround fails.
Today those experiences might remain separate. In a connected learning network, they become evidence about the same problem. The system can surface the common pattern, preserve the important differences, and help teams test several explanations in parallel.
A useful result can then reach every relevant location. Each team begins its next experiment with more knowledge than it had before.
The gain has two parts. More people can explore different parts of the problem at once. What they learn can improve the next round of exploration across the network. Under the right conditions, discovery and reuse reinforce each other.
Adding a person then adds more than their individual output. It can add a new context, a different skill, another set of relationships, and another place where the network can test its understanding.
The new person benefits in the other direction. They gain access to accumulated experience without having to repeat every mistake that produced it. The learning cliff gets smaller. They still need practice and judgment, but they can begin contributing sooner.
The authors trace the effect to the AI passing along what the best agents already did, much of it tacit knowledge that had never made it into a manual. New agents picked up those habits without working alongside the people who had them.
This is the compounding effect I care about.
The enterprise already has the network
Established companies possess much of what this system needs: employees, customers, distribution, operating history, and a physical footprint. They have people encountering real problems every day, and places where proposed solutions can be tried.
Much of that potential is trapped in separate teams, systems, and habits.
If the cost of connecting it falls sharply, the competitive meaning of scale changes. The enterprise can use its existing reach to discover more, test more, and spread useful learning faster. Its size becomes a source of speed.
Firms with 250 or more employees use AI at about 1.7 times the rate of the smallest firms. The share climbs with every size class from 20 employees up. In this survey, adoption is furthest along where the networks are largest.
My shorthand for this is deliberately blunt: size is all that matters.
By size, I mean effective connected scale. How many distinct, useful perspectives can the network bring together? How much of its learning reaches someone who can act? How quickly does the result improve what everyone else does?
Ten thousand people repeating the same observation do not provide ten thousand independent discoveries. A smaller network with complementary experience, clear authority, and fast feedback can outperform a much larger one. Payroll is a poor measure of the thing I am describing.
When observations overlap, the count of independent views stops growing at one divided by the overlap, however many people are added. At an overlap of 0.01 the ceiling is 100. At 0.1 it is 10. To get more out of size, lower the overlap by drawing on different locations, roles, customers, and experience. Hiring more people who see the same thing barely moves the count.
The boundaries may also extend beyond a company. Customers, suppliers, partners, and independent specialists can participate where trust and permissions allow. A startup could build such a network. An incumbent could fail to connect the one it already has.
Small teams get the same improving AI tools. The question is whose constraint those tools relieve most. If a startup has been winning partly because its competitor cannot coordinate what it already knows, that advantage can shrink. The incumbent still has its customers, distribution, and operating reach.
Existing enterprises have a substantial starting advantage: their contact with the world is already there.
What has to become true
I am looking at the next 12, 24, and 36 months as successive tests of this thesis. I do not know the date at which coordination becomes cheap enough to change the balance in a particular industry.
The evidence we have is narrower than the outcome I am describing. METR measures progress on mostly well-specified software tasks; those results do not establish autonomous business management. Epoch AI has documented falling costs for a given level of benchmark performance; the total cost of reliable work includes much more. Both trends matter. Neither proves this thesis.
Both curves describe benchmarks, not businesses. The gap between the last two numbers matters most here. Agents can attempt long tasks but finish only short ones reliably, so checking the work is still the costly part.
The relevant operating measures are concrete: the time from observation to a useful experiment, the cost of a verified result, how reliably that result transfers between settings, and how quickly a new person can contribute.
| Measure | Best public evidence | Strength |
|---|---|---|
| Time from observation to a useful experiment | None directly. In the P&G experiment, people with AI reached solutions 12.7% to 16.4% faster. | Weak |
| Cost of a verified result | Benchmark performance costs about 13× less each year. Reliable agent work is still about an hour long, and verification cost is not tracked. | Partial |
| How reliably a result transfers between settings | A study of 122 best-practice transfers in eight companies found the barriers were knowledge, not motivation: the recipient’s capacity to absorb, unclear cause and effect, and strained relationships. No AI-era measure yet. | Pre-AI only |
| How quickly a new person can contribute | Support agents with AI matched six-month agents at two months. | Strong |
Some constraints will survive much better models. Conflicting incentives will remain. People will disagree about goals. Someone will still need authority to commit resources and responsibility for the consequences. Physical experiments will still take time, equipment, and money.
AI can also spread a bad inference much faster. If every location adopts the same mistaken explanation, connection has made the organization more fragile. A learning network needs independent checks, room for dissent, and the ability to preserve local differences.
Each writer improved while the group as a whole converged, which is the fragility risk on a small scale. A network that draws every location toward the same suggestion loses the local differences it needs to catch a wrong one.
The coordination tax does not have to reach zero. It has to fall far enough that the benefit of additional connected experience grows faster than the burden of coordinating it.
This is an illustration with made-up numbers, not a forecast. It shows the shape of the threshold. The value of added people grows in a straight line and the cost of coordinating them grows with the square, so every organization has a size past which growth stops paying. That size moves in proportion to the cost of a connection, so cutting the cost tenfold moves the turning point tenfold without the cost ever reaching zero.
That is the threshold I am watching. If adoption produces more messages and summaries while decisions and validated learning stay slow, the thesis has not shown up in practice.
Agent counts are climbing much faster than leadership alignment. The MIT figure is directional: it rests on interviews and conference surveys, and its six-month window may understate success. Taken together, they show more activity without evidence yet that the threshold has been crossed.
Where this leaves the individual
The personal implication extends beyond getting better at using AI alone.
I want to understand how to build, join, or own part of a network that learns. What can I contribute that expands its contact with reality? What experience does it give me access to? Does what I learn improve the work of others, and does their learning reach me?
Individual skill matters inside that system. Judgment, curiosity, relationships, and the willingness to test an assumption become ways to improve the network. A person with unusual experience may contribute something the rest of it cannot generate for itself.
There will be enormous opportunities for small teams. But the familiar assumption that large organizations must be slow deserves another look. Some of that slowness comes from the cost of making many people function together. AI is beginning to change the tools available for doing that.
I think we are underestimating what happens when an enterprise can learn across the full extent of its experience.
The next great advantage may belong to the organization that can bring the most distinct human experience to bear on a problem, test what it thinks it knows, and make the result useful everywhere.
That is the revenge of the enterprise.