The Economics of Intelligence: How and when does AGI lead to economic transformation?

This essay draws on the paper Organizing Intelligence which presents the full economic model and derives the results formally.

How do capability improvements transform the economy, and when? In this essay, we define thresholds of economic capabilities, e.g., minimal AGI, as a function of how they transform the economy. We then track how the economy–market structure, employment, and wages–changes during the transition across capability regimes.

The last few years have seen exponential growth in the capabilities of machine intelligence. Every few months, we see a frontier AI model clear another milestone that seemed years away: writing production software, outperforming domain experts, and solving Millennium Prize problems in math. But looking around at the broader economy, it seems remarkably unchanged. Most people still go to recognizable jobs, headcounts look very similar to a few years ago, and aggregate productivity numbers have barely budged.

Part of the reason for this disconnect is that the individual AI agent is the wrong economic object. The vast majority of capability benchmarks measure individual AI agents solving problems in isolation, but economic transformation is an equilibrium outcome where organizations compete, subject to the forces of supply and demand. Workers do not operate in a vacuum. Organizations assemble people with complementary strengths and divide up tasks to create teams far more capable than any solo worker. Because existing benchmarks only test whether AI can outperform individuals, a model can exceed the performance of every person one-on-one and still lose to a well-structured team. This is why organizations have sometimes been referred to as a form of superintelligence.

At the same time, standard economic models would miss the transformational potential of AI because they still tend to model machine intelligence as previous technological innovations—either as a better tool in a worker’s hands or a machine that is swapped in for people one task at a time. General-purpose machine intelligence does not fit that mold due to two main features. First, while the supply of human workers is fixed, AI agents and their associated capabilities can be scaled on demand at the cost of compute. Second, the constraints that limit the size of human teams such as span of control, communication frictions, and misaligned incentives need not apply to machine agents. When they can eventually work directly with one another rather than sitting inside human hierarchies, we will be dealing with a genuinely new kind of economic input.

So, when does machine intelligence actually lead to economic transformation—and what capabilities need to be achieved in order to see the impact of AI on market structure, wages, and jobs?

To answer this question, we derive economic scenarios under different levels of AI capabilities. By examining how intelligence functions inside organizations, the answer to when the economy transforms depends on when two key capability thresholds are crossed:

• Competence: Gains in individual skills like math, writing, or coding will smoothly shift wages, prices and the mix of humans and AI agents in a firm while leaving industry structure largely intact.
• Coordination: Gains in the ability of agents to work together within a firm alter how organizations scale. Once coordination reaches a critical threshold, it triggers large, abrupt jumps in market concentration, employment, and wages.

Mapping where AI sits along both dimensions reveals distinct regimes of capabilities based on how they impact what the economy looks like, who gets employed, and how much they get paid. We term these regimes minimal AGI, supercompetent AI, supercooperative AI, and Artificial Super Intelligence (ASI). Tracking the transition between regimes shows why the real visible transformation may come from improvements in coordination—a capability most benchmarks don’t yet capture.

Jagged agents, mosaic organizations

The vast majority of current benchmarks don’t attempt to measure coordination. They treat intelligence as an individual score rather than a jagged profile that can coordinate with complementary teammates.

To show why this is important, we gave 100 people a quiz spanning six domains: math, verbal reasoning, science, patterns, culture, and economics. Their performance in one domain turned out to be not very predictive of their performance in another. The population didn’t contain copies of the “average intelligence”; it contained specialists who were jagged along different directions (strong in some areas, weak in others).

For example, one person may be very good at many of the domains but abysmal at pattern matching; on the other hand, another participant could excel at pattern matching while showing middling performance on everything else. (Figure 1 shows data from six participants.)

Figure 1. Jagged human capabilities across six domains and organizational mosaic intelligence.

But workers almost never participate in consumer and enterprise markets in isolation. They are part of organizations that exist largely to do something really clever with humans’ inherent jaggedness: they combine complementary specialists so that the team’s capability profile covers ground no member covers alone. This is Ricardian comparative advantage operating inside the firm—the idea that people sort into jobs that they are relatively better at. Roles aren’t filled randomly, but structured to maximize each specialist’s comparative advantage.

The result is mosaic intelligence: the organization builds a collective capability that captures the strengths of every specialist while eliminating the gaps between all of them. By filling in the jaggedness across worker profiles, mosaic intelligence outperforms every individual in it. Figure 1 shows how this works in our data: the blue envelope represents the average ability of the six participants on the left, but the red (much larger) envelope corresponds to the intelligence of an organization that matches specialists to the skill that they are best at. When participants are paired to complement each other to solve a specific task, they outperform random pairs by 32%.

Now apply this logic to an entire team. Consider building and shipping software, which requires both strong coding and clear writing. Pairing specialists in both areas balances out individual weak spots. The resulting productivity gain means each coder and writer within the firm can earn more than they can on their own outside of it—not because they each become more capable in isolation, but because complementary specialists reinforce one another’s comparative advantage, allowing both to command higher wages. (The left panel of Figure 2 shows how this works with a simple numerical example.)

Organizations also unlock value for a third kind of worker: coordinators. Judged by any standalone single-agent benchmark, a pure coordinator looks like the weakest worker in the economy, average at coding and average at writing. Working by themselves, their economic output is negligible. Inside a team, however, a skilled coordinator eases handoff friction across colleagues, enables the team to grow larger without losing focus, and makes every coder and writer noticeably more productive. The coordinator allows the team to benefit from what each specialist individually excels at. In equilibrium, coordinators—or as they’re more commonly called, managers—earn some of the highest wages in the economy, not because they excel at a particular task, but because they maximize the output of the mosaic.

Now what about machine intelligence? Much has been said about the jaggedness of AI agents-–they can solve advanced math problems but stumble on what appear to be fairly easy problems. Human capability is uneven too, but while human intelligence is jagged in different directions, AI models are jagged in ways that are very similar to one another. We can add a complementary human specialist to a team to fill a gap, but spinning up another copy of a model only adds capacity without increasing breadth of capabilities.

A model whose capability profile outperforms the average of every individual human can therefore still lose in the competitive market to a team of jagged humans who reinforce one another, none of whom could best the AI on their own. Asking "is the model better than a human?" is therefore focusing on the wrong object of economic analysis. In a competitive market, the right question is whether an organization of AI agents can surpass an organization of humans, or a blend of humans and agents.

Figure 2. Mosaic intelligence across skills and organizational scale curves across firm sizes.

How AI differs from humans

Beyond how capability profiles fit together, AI agents differ from human workers in two key ways that are key for understanding economic transformation: duplicability and coordination.

Human expertise is in fixed supply. Each person has a finite number of hours in a day and cannot be scaled at will. To expand, industries are required to hire further up the talent distribution, which inevitably drives up wages. This is the essence of what Ricardo (the same one as before) taught us: the highest economic returns accrue to those whose skills cannot be reproduced.

AI removes that scarcity bottleneck at scale. Once a model reaches a given capability level, it is available in any quantity at the rental rate of compute. This becomes the backbone of how wages change when AI is introduced into the economy: whenever a model can match human capability, human wages become tied to the cost of compute.

The second difference is coordination. As the Nobel-winning economist Ronald Coase famously asked: if organizing work inside a firm cuts costs, why isn’t all production carried out by one big firm? A century of organizational economics has answered Coase by pointing to human limits. Two colleagues working side by side hold the full design of a project in mind and resolve mismatches immediately. Expand that team to 20 or 200 people and attention starts to stretch thin, implicit assumptions drop out in handoffs between workers, and private career incentives create misalignment amongst employees. Much of human knowledge is tacit—captured in Polanyi’s Paradox of "we can know more than we can tell"—and passing work from one person to another compresses rich context into lossy summaries. Jacob Marschak and Roy Radner built an entire theory of teams around the information that is lost in communication alone.

History shows us what happens when there is a shift in coordination constraints. In the 19th century, the emergence of the telegraph and the railroad didn’t make individual workers better—they raised the speed and fidelity of handoffs, and that alone reorganized production into structures that had never existed. It was improvements in coordination technology rather than advances on individual skills that led to the rise of the modern firm.

Importantly, none of the constraints that limit human coordination have to apply to AI agents. Today’s AI systems coordinate poorly in practice. But because of their high communication bandwidth, their potential for coordination is immense. Copies of an agent share identical weights and representations. They can pass along a complete working state—including entire codebases, intermediate derivations, and full context windows—and execute a common objective without conflicting career concerns.

The limits that have historically bounded firm size stem from human bandwidth and misaligned incentives. In order to predict an AI-driven economic transformation, it’s critical to understand if these limits apply to agents.

Competence sets the level but coordination sets the scale

Now imagine a market where firms assemble any workforce they like, hiring scarce humans at market wages and renting AI copies at the price of compute. Competition then selects organization structures that survive new entrants.

Any organization’s overall capability can be decomposed into the competence and coordination dimensions discussed earlier. Competence sets the level of output per worker at a fixed firm size. Coordination determines what happens to that output as the organization adds more humans or AI agents. Put another way, competence sets the potential productivity level; coordination determines how big an organization can get without becoming too messy to function.

Track output per worker as an organization grows and you get what we term organizational scaling laws—a parallel to model scaling laws. Whereas model scaling laws govern returns from parameters and data, organizational scaling laws govern what happens as you scale a team of agents. They have three phases:

Separating competence and coordination yields an important insight about economic transformation: competence reshapes the economy gradually, whereas coordination abruptly transforms it. Raising domain competence shifts prices, wages, and the mix of humans and AI agents smoothly while keeping firms relatively small. The economy looks relatively similar to the outside observer. But once coordination passes the supercooperative threshold, it triggers abrupt jumps in firm concentration, employment, and wages. The economy looks nothing like the one we see today.

The Four Regimes of AGI

Rather than treating AI progress as a single slider moving up, our framework tracks improvements in competence and coordination independently of one another. This generates four capability regimes based on what these capabilities imply for market structure, employment, and wages. We call these regimes minimal AGI, supercompetent AI, supercooperative AI, and artificial superintelligence (ASI).

To see how these regimes play out across the entire economy rather than inside a single industry, we track three kinds of sectors.

• Deep scalable markets: Large, standardized industries like software, logistics, or finance where demand can absorb massive volume.
• Niche markets: smaller, specialized, or custom settings where bounded demand limits the payoff to sheer scale.
• The relational sector: Goods, services, and experiences where human involvement is part of the core value - from healthcare and teaching to services and craft.

Below, Figure 3 illustrates what happens to market structure and employment across these three sectors as we transition between the AI capability regimes.

Figure 3. Evolution of firm organization across niche markets, deep scalable markets, and the relational sector across AI capability regimes.

Minimal AGI

First consider minimal AGI, where reproducible AI matches or exceeds the average capabilities of humans across tasks. At first glance, almost nothing structurally changes in the economy. In deep markets, balanced human mosaic intelligence teams still outcompete AI firms because the joint capabilities of complementary experts dominate a firm with average capabilities. Headcounts stay intact and industries remain fragmented across many small to medium firms.

All of the changes occur in wages. Workers whose skills are matched by AI will see their wages decline toward the cost of compute, while specialists and human coordinators can demand even higher wages than before. Because copies of minimal AGI agents can be rented at the cost of compute, small AI firms enter price-sensitive and smaller niche markets, driving down the price of goods.

The key takeaway has implications for today’s debate about AI and the economy: because wage effects arrive before organizational restructuring, macro statistics of employment rates provide poor evidence that AI is not having an effect on the economy.

Supercompetent AI

Now let’s consider the case where AI systems outperform every human expert at every capability except coordination, a regime we call supercompenent AI. What happens when AI agents become more competent without necessarily becoming better at coordinating?

Surprisingly, there is still little structural transformation. Even when no team of human specialists can surpass the model on individual tasks, coordination still caps how big an organization can get without falling apart. Introducing supercompetent AI means that firms become far more productive without getting larger; in fact, headcounts and scale likely shrink. Industries stay competitive across many lean firms, goods and services grow cheap and abundant, and the economy delivers something counterintuitive: abundance without concentration.

Inside those lean firms, however, roles flip in a striking way. Because AI can handle standalone specialist work at the cost of compute, pay for routine specialist tasks gradually declines. But large teams and projects still suffer without a skilled human coordinator. Someone who sets direction and keeps the whole effort aligned multiplies the output of a team of supercompetent AI agents. People do not vanish from production; their comparative advantage shifts toward judgment, orchestration, and creating the architecture for effective coordination.

Supercooperative AI

Now flip the dimension of improvements: hold an AI agent’s standalone competence at the level of minimal AGI but improve agent coordination past the supercooperative threshold.

Here, the economy experiences a sudden phase transition. Each additional agent in an organization now makes the team more productive rather than adding friction. In mathematical terms, returns to scale flip from being concave to convex. In deep, standardized markets, these types of highly synchronized AI organizations can grow large and then suddenly outcompete smaller firms, even if no individual agent is particularly impressive on an isolated benchmark.

The potential difference between two economies sitting on opposite sides of the supercooperative threshold can be massive: even if the competence of the AI models is nearly identical, one economy can look fragmented and familiar, while the other reorganizes rapidly into a few large AI-only firms. In the latter, the reorganization effect also causes wages to drop abruptly.

But this doesn't happen in the entire economy. In niche markets, buyer demand is naturally bounded, making it hard for an AI organization to recoup enough through scale to be competitive with smaller human teams based on skill. However, an influx of workers moving into the niche and relational sectors will drive down wages in those markets, which shows how a breakthrough in machine coordination can lead to massive changes in the entire economy.

Artificial superintelligence

When capabilities cross both the supercompetent and supercooperative threshold, we reach the regime of artificial superintelligence, or ASI. In deep markets, production remains concentrated in large AI-driven firms. Because ASI also surpasses human specialists on standalone tasks, smaller niche markets no longer shelter human teams: lean AI firms can now serve specialized demand without needing massive scale.

Looking only at the goods and tasks that machines can replicate, it would seem that human labor has nowhere left to go. But workers may not sit idle once we consider the entire economy. To see what happens to work and value as a result of automation, it’s useful to consider the economics of structural change from a historical perspective. When automation increased productivity in agriculture and later in manufacturing, abundance in those sectors made society richer overall. And automation did not decrease the number of jobs—people spent their newfound income in sectors that were not automated (services), and those sectors grew in both value and employment.

But AI is different from previous technological change, and the wave of automation could be much broader. As AI drives down prices, the majority of goods and services may become commoditized and society would grow much richer in real terms. So where do spending and employment go? One possibility is that this wealth flows toward goods and services that are valued but remain scarce—the relational sector, where human involvement is itself part of the value. This likely includes familiar fields in which most tasks can be automated but human delivery increases its value, such as healthcare and education. And if history has taught us anything, we will also see brand-new jobs that are hard if not impossible to imagine today (in 1950, who would have thought a large share of the population would work in something called “software engineering”?). So even if human pay in commoditized production is anchored by the cost of compute, much as Pascual Restrepo has analyzed for fully automated industries, the relational sector may expand substantially and employ much of the workforce at high real wages.

Seeing the future

The relatively modest economic changes we see today make complete sense through the lens of our framework: AI models are getting exponentially better in individual competence while lagging behind in coordination. Across dozens of professions, models already match or outperform experienced specialists on self-contained deliverables. Spin up a swarm of agents, however, and output still runs quickly into diminishing returns.

Frontier deployments encounter the same problem. Reflecting on running thousands of agents in parallel on open math problems, OpenAI’s Noam Brown notes that “it’s actually very difficult to get these agents to coordinate in a productive way.” Observing that swarms readily collapse into working on the problem independently rather than dividing labor effectively, he noted that he “wouldn’t even attribute 10% of the credit to multi-agent” capabilities over single-agent performance.

Today’s labor market reflects the current capabilities of AI systems: not mass unemployment or sudden AI monopolies, but leaner startups, flatter org charts, softer hiring in entry-level roles closest to benchmark tasks, and steady or rising demand for experienced people who can direct complex work across teams.

If the deepest structural shifts hinge on coordination, our measurement toolkit needs to catch up. Today’s leaderboards overwhelmingly test solo agents in a vacuum. As economic predictors, these scores overstate how well the models compete against complementary human teams and give little warning of how close AI systems sit to a cooperative tipping point.

Alongside individual leaderboards, we need benchmarks for organizations of agents: measuring how well multi-agent systems divide unfamiliar problems without duplicating work, share context cleanly across handoffs, and catch rather than compound one another’s mistakes over long horizons. Economic statistics need a parallel update, tracking both payroll and compute spending to see how people and machines are teaming up in practice.

In the meantime, wages and job roles—not headline GDP or industry concentration—remain the earliest signals of change, moving well before firm boundaries visibly shift. For too long, debates over AGI have treated intelligence as a solo capability race between a person and a machine. What matters is the organization that turns individuals’ jagged capabilities into something far greater than the sum of its parts.