How to think about the economic impact of AI

The scarce factor is humans’ cognitive ability

Economists struggle to predict large productivity gains resulting from AI.¹ The truth is that, with a given structure of tasks or jobs, improving how we do a portion of those tasks is unlikely to make a big difference in productivity. But in an economy where time and knowledge are the key scarce resources the advent of AI, with its ability to solve problems at scale, can change how the entire economy is organized. This has the potential to significantly increase the potential gains from AI, and also affect the entire labor market. Rather than focus on existing tasks being replaced, to obtain a precise numerical estimate, it's essential to determine how much total problem-solving "brain power" current levels of compute equate to in terms of the number of operations they can perform and the level of ability (in human terms) at which they operate.

The Knowledge Economy

In the modern economy, time and expertise are the main inputs needed to do most jobs. For consultants, lawyers, doctors, managers, repair technicians or production workers in the factory floor, doing their work involves facing problems that need solution. A doctor needs to diagnose a patient; a manager needs to decide on which input to purchase; a repair technician needs to diagnose the broken machine; the shop floor employee must implement a changeover of the product. In all these cases, knowledge is the indispensable input in production. So the ability to manage efficiently its acquisition, communication, and use is a key determinant of the efficiency of any organization, be it a team trying to keep the production of a manufacturing company going or the entire economy.   

Think of knowledge as the ability to solve the problems that naturally arise in any production process. The organizational problem arises because this knowledge is embedded in individuals who have limited time to work. Hence, although the knowledge they possess can be used repeatedly as an input in production, the person that possesses this knowledge faces a time constraint that limits in practice how often this knowledge can be used. How to use the knowledge better?

The solution is to organize production in “knowledge hierarchies.” To make the best use of the time of an expert lawyer, protect her from wasting her time on routine problems — let some associates handle the easy part of the job, and let her deal with the exceptions. The “expert” only comes in when the problem becomes sufficiently hard to be worth her time. In the factory floor, the supervisor directs the tasks of shop floor employees by substituting her expertise for theirs on the problems they do not know how to solve. The direction here is not authority — it is knowledge. The hierarchy is a knowledge hierarchy — the workers do as the engineer tells them not because she is their boss, but because she knows the solution (that is why she is the boss).

The theory of knowledge-based hierarchies  shows how organizations allow agents to relax this time constraint by working in teams, where less skilled workers deal with routine production tasks, thereby economizing on the time of more skilled agents, and allowing them to specialize on giving directions on the harder tasks. The organizational problem then becomes the problem of determining who knows what, who they communicate with, and how many workers of each type are required. That is, it becomes the problem of determining how to use knowledge efficiently.²

The growing importance of knowledge in the economy is the reason the return to cognitive skills (the wage of those going to college versus not going, of those having a graduate degree versus having only an undergraduate degree) has been increasing. And it is also the reason for the increasing importance of assortative matching on the jobs – better people going to firms where other top people are. Why do we see the best lawyers together in “the best” law firm? Because an expert lawyer wants to spend her time as effectively as possible, and  hence only work on problems that are worth her time. That requires having the best associates under her. ³

The impact of artificial intelligence on a knowledge economy

AI can radically alter this picture. Essentially, it allows the economy to eliminate the time constraint organizations face on the agents' utilization of their knowledge – at scale. That is, what is crucial with AI is that, once trained, the capacity of the model to solve problems is almost unlimited (up to small costs of electricity).  Consider for instance the recent work by Erik Brynjolfsson, Danielle Li & Lindsey R. Raymond (2023), who study the introduction of a chatbot to assist 5,179 customer support agents in answering queries. This tool raised their productivity by 14% on average, with a particularly large improvement (34%) for those with less expertise.

How will AI change the organization of work and the rewards of expertise in a knowledge economy? In a 2024 working paper, Enrique Ide and Eduard Talamàs from IESE Business School in Barcelona pose this question in the context of  a knowledge economy with individuals with different levels of knowledge — understood as different abilities to solve problems– organized in knowledge hierarchies.  I follow their analysis in what follows. Note that it only has two layers of workers, and we could have (particularly post AI) multiple layers — a case they do not study.⁴

The crucial distinction we want to draw is between AI that needs humans to solve problems (the world of the paper by Brynjolfsson and coauthors where AI helps sales employees solve problems) and the world where AI can solve problems without humans. 

When the humans are in control

Start by considering an AI that, like the chatbot in Brynjolfsson et al. (2023), is not autonomous: It can only be used alongside a human. For example, doctors are still needed to diagnose patients, but they may use AI to assist them in the process. In this world compute is abundant in the sense that every human can use AI if she wants to. 

Workers with different skills are organized in two layers. Less skilled workers handle simpler problems and direct their questions to more skilled workers. Workers ask questions to problem solvers. The figure below shows the earnings of workers relative to their skill levels. Skill is on the horizontal axis, earnings on the vertical axis. The problems that AI can solve are marked by the point labeled AI — an artificial intelligence system can solve a large number of problems, significantly more than workers just below or above it, as long as the problems are no harder than  AI.

The chart depicts two income distributions: the lighter line represents the pre-AI world, and the darker line shows the post-AI scenario. All workers below the AI level use the AI to help solve problems and all earn the same income, since they all handle problems up to zAI. Similar to the empirical findings in Brynjolfsson et al., the less skilled workers benefit the most from the AI- they are there to feed problems to the AI, and all earn the same, as they all effectively have the same ability- the one of the AI they use. All individuals who have problem solving ability above the AI level are specialized problem solvers and also benefit, as they can now use AI to reduce the time they spend on simple problems and allocate more mental resources to complex problem-solving, hence cheaply leveraging their expertise over a large number of problems. 

When the AI can work alone

Now switch to the case where AI is autonomous, it can solve problems without humans. For example, consider the AI agents currently under development (“agentic AI”). By adding long-term memory to AI systems capable of reasoning, agents will be able to function as ‘remote workers on demand’. Models like o1 are already capable of complex problem solving, but require you to provide context and background in  every conversation.  With long-term memory, the AI could remember and work on tasks across sessions. Add to that the ability to interact directly with the web the way humans do — which Anthropic just announced — and improvements to multi-step planning, and you have a system capable of autonomously working on and solving problems.  

To study the changes in income and organization in this case we continue to assume that every human can use AI if she wants. However, AI is now more expensive, since it can solve problems on its own, and some humans might decide using AI it’s not worth the price.  Additionally there continues to be a fundamental scarcity in the economy: there are always more problems to solve than available resources (compute plus human time). 

The two charts below show the earnings distribution with autonomous AI systems at different levels of capability. The axes remain the same as before: skill level on the horizontal axis and earnings on the vertical axis. The left-hand graph illustrates  the case where AI can solve as many problems as a pre-AI worker.  This reduces demand for workers, lowers their wages, and pushes top workers to become problem solvers who supervise low level workers –hence the  shift moves (some) humans from routine tasks to knowledge-intensive roles. AI clearly benefits top solvers who can use AI to solve their routine problems (think of getting a set of AI paralegals). Finally, even though AI has an absolute advantage over a segment of the population, there is no unemployment. This occurs because there are more problems to solve than resources available.

In the right-hand graph, AI is much “smarter” – it can solve problems as well as a pre-AI solver, reducing the need for human problem solvers and pushing the least skilled back into routine work. This improves the quality of lower-tier solvers and enhances the productivity of less skilled workers by matching them with better solvers or AI. Thus, while AI displacing lower-skilled tasks moves workers into more complex roles, AI replacing higher-skilled solvers can lead to an improvement in workers earnings. The very best problem solvers still do benefit from having top workers (previously problem solvers) work for them.

The difference between the autonomous and the not autonomous scenarios is very sharp. In both cases, total labor income increases. But when AI is not autonomous (as in the first chart) all humans always benefit from AI. When AI can work alone, AI pushes down wages whose skills are close to what AI can do– there are winners and losers. 

How can we quantify the impact of AI on GDP and labor income?

Can we use the model to generate actual quantitative estimates? In principle, yes. To do that, we would need to know the distribution of cognitive/problem solving ability in the population–which in the illustrations we have used, for simplicity,  were based on a uniform distribution; we would need an estimate of how much compute /machine time is going to be deployed in human equivalents — in the examples above we assume that machine time deployed, compared to total human brain power, is at least twice the amount of human time; and how good will AI be — here the crucial parameter in the model is the share of problems that AI can solve on its own –min the examples above, AI has the same cognitive ability as a 25 percentile worker. Without knowing how much compute we have, and how much cognitive capacity it adds to our economy, the calculation of GDP gains is impossible.  

We can then use the model with our preferred scenarios to produce an estimate of the value of AI. For instance, if we had compute equivalent to twice the processing ability of the population of a country and with the same problem solving capacity as the 25% percentile human, the model yields a prediction of an increase in labor income of 50% in the case where AI must be paired with humans to solve problems, and by 20% when it can work on its own (“agentic” AI).  GDP increases by  over 50% in the first case and almost 70% in the second.  In the autonomous case, a large share of the gains go to the capital owners, who can use the AI to deal with problems without any human intervention – their reward depends on total compute available. These large numbers are not surprising: After all, the rise of AI implies that the vast amounts of computing power we have available can significantly increase our collective problem-solving capacity.

Why do economists struggle to find the effects of AI?

There are two main reasons why the effects usually found are small. 

First, keeping the structure of work unchanged, it is difficult to find big changes. Electricity would not have done much for the economy before machines were introduced — if you offered electricity to a household before the existence of washing machines or lightbulbs, what would they do with it? Similarly, the first use of petroleum was as kerosine for lighting, replacing whale oil. It would take the invention of cars before gasoline (previously treated as a waste product) found its true use in mobility. Doing a bit better on the existing tasks is always going to have a small effect  — the big gain comes when we reorganize the economy, 

Second, and more broadly, changes in the cost of acquiring and communicating knowledge are generally not part of the economics toolbox — economists tend to assume agents with unlimited rationality. Kenneth Arrow’s The Limits of Organization (1974) points out that the competitive equilibrium requires vast amounts of information–individuals must know a vast amount of prices of contingent commodities for each state of the world. The cost of acquiring information is a consequence of individuals’ limited capacity to process it, the irreversible nature of investing in knowledge, and the uncertainty of its future value. Organizations mitigate these constraints by acquiring and processing more information than individuals, using internal languages and hierarchies to optimize knowledge flow. Where price systems fail, technologies like artificial intelligence offer significant potential to enhance decision-making in organizations and that can lead to large increases in productivity.

Third, AI can increase welfare without increasing GDP, as more individuals solve problems by themselves. Imagine we give consumers better problem solving technology — an AI. Suppose this is sufficiently better than the consumer rarely needs to ask questions — or suppose that in the limit does not ask them. Then AI is vastly increasing welfare, by solving lots of problems — but GDP is going down, because there are no "market" transactions — individuals solve the problem themselves. No one is calling the repairman, because they can repair the problem themselves.

In my view, economists vastly underestimate the impact of AI and its ability to transform productivity by focusing on task replacement. Task-based models do provide useful insights into how automation displaces labor in specific industries. But they overlook how AI can reshape organizational structures and knowledge work throughout the economy. AI is not just another automation tool — it fundamentally alters the constraints on cognitive ability, time, and coordination that have historically limited human productivity.

By enabling scalable application of knowledge, AI breaks out of the human limitations of time and bounded rationality. It allows for a large increase in the available knowledge that we can apply to solve problems, and creates enormous efficiencies in resource allocation. This ability to scale knowledge across domains offers the possibility of large productivity gains.

Biographical note

If you are interested in the theory of “Knowledge hierarchies” discussed here, you can read a survey my coauthor Esteban Rossi-Hansberg and myself wrote reviewing the main ideas: "Knowledge-based hierarchies: Using organizations to understand the economy" in the  Annual Review of Economics.  The theory was introduced in my JPE 2000 paper, although the work with heterogeneous individuals that serves as foundation for what we discuss here originates from my work with Esteban Rossi-Hansberg in the AER in 2004 and QJE in 2006; with Pol Antràs in the QJE in 2006 as well, and with  Willie Fuchs and Luis Rayo in the REStud in 2015. With Thomas N.Hubbard I wrote a series of papers between 2005 and 2018 documenting the organization of knoweldge (including the associate-partner hierarchies and the organization of specialization) in the law. You can find all of these papers on my webpage.

References

Acemoglu, Daron. The Simple Macroeconomics of AI. No. National Bureau of Economic Research, 32487. 2024.

Brynjolfsson, Erik, Danielle Li, and Lindsey R. Raymond. Generative AI at work. No. w31161. National Bureau of Economic Research, 2023.

Garicano, Luis, and Esteban Rossi-Hansberg. "Knowledge-based hierarchies: Using organizations to understand the economy." Annual Review of Economics 7, no. 1 (2015): 1-30.

Ide, Enrique, and Eduard Talamas (2024).  “Artificial Intelligence in the Knowledge Economy.”https://doi.org/10.48550/arXiv.2312.05481.

1. For instance, in a 2024 paper, Daron Acemoglu expects 0.55% growth in TFP over the coming decade.

2. For a general introduction to this literature, see the review in Garicano And Rossi-Hansberg (2015).

3. Documented in “Learning about the nature of production from equilibrium assignment patterns" by Luis Garicano and Thomas N. Hubbard, Journal of Economic Behavior & Organization (2012).

4. The model builds on my work with Pol Antràs and Esteban Rossi Hansberg to study outsourcing, extended in a paper with Willie Fuchs and Luis Rayo Esteban Rossi-Hansberg and I analyzed a general case with multiple layers in our QJE paper. See the bibliographic note.