In a new working paper, Benjamin Verschuere and Angus Cameron argue that the wide dispersion in economists’ forecasts for the impact of artificial intelligence on the economy stems from two gaps. The first is that estimates for growth, jobs, and prices are each built in isolation, with no single framework to reconcile them. The second is that models fixate on AI’s current capabilities, rather than on how fast it spreads and how much of a given job it can eventually reach. The authors build a unified framework that predicts roughly $2 trillion in long-run output gains, the loss of about 20 million American jobs, and falling prices.
Ask three experts how artificial intelligence will change the economy and you will get three irreconcilable answers. Daron Acemoglu’s careful work, the most skeptical of the serious estimates, puts the cumulative productivity gain from AI at under one percent over a decade, and does not forecast a large output boost at all. Analysts from Goldman Sachs have floated a $7 trillion boost to global output over the same time span. Analysts from McKinsey speak of trillions in annual value. Equally on the labor front, forecasts are divided on how AI will displace workers, with even some arguing that it will increase hiring.
When forecasts diverge so widely, the fault rarely lies with the data. In this case, it lies in the fact that each estimate rests on its own model and assumptions.
What has been missing is a unified framework, one that starts from what actually governs AI’s economic impact: how much of a job it can ultimately automate and how fast it diffuses. Ours takes both into account, deriving forecasts for GDP, prices, jobs, and the fiscal balance within a single structure, and treating AI not as one innovation wave but as several, each interacting differently with different macroeconomic variables. We find that the field has underestimated both job losses and disinflation, that output forecasts are scattered while ours sits on the larger side but coheres with our other numbers, and that our framework, unlike others, can quantify the fiscal impact.
Our model
Our modeling starts at the task level. AI automates tasks, not whole jobs, and most jobs are a chain of tasks, some of which no machine can perform. The first question is, therefore, how much of a given job can AI actually perform?
To answer this question, we develop a metric called “workflow completeness,” which estimates the share of a job’s tasks that AI can complete end to end. We combine this with Amdahl’s Law, which holds that the acceleration of any process is capped by the part that cannot be accelerated. If a fifth of a job must remain human, then however fast AI handles the remainder, the workflow can move no faster than its human component. A radiologist’s role of flagging anomalies can be automated; the final diagnosis, the liability, and the conversation with the patient cannot. Our term “workflow completeness” captures that ceiling.
Aggregating tasks to jobs, then to the industry level, we find that AI can automate about 12% of the construction sector, about 68% of information, technology, and finance, and 75% of customer service. This is the structural limit toward which AI is converging, and it does not rise as AI models improve, because what holds them down is the physical, regulatory, and relationship-bound work that better software does not touch. Our measures align closely with academic indices grading industry exposure to AI replacement, such as the AI Occupational Exposure index of Edward Felten, Manav Raj, and Robert Seamans and the task-based measure of Tyna Eloundou and co-authors.
Next, we measure diffusion. Here lies a common error among economists, who tend to focus just on what technology is capable of, rather than other factors such as occupational and social frictions to adoption. Electricity was a technology capable of powering factories in 1882. Computers were capable of simplifying an array of tasks in the 1950s. Neither registered in the productivity statistics for decades. In every previous general-purpose technology, the constraint was never the raw capability of the technology but the speed with which firms and workers reorganized themselves around it. Factories had to be physically rebuilt around electric motors before the gains appeared. Firms in the 1990s had to redesign their workflows around computers, not merely purchase them, before output responded.
Rather than ask how capable AI is, we trace how it actually advances toward its ceiling. We decompose that progress into four distinct rates. The innovation rate is the speed at which AI becomes technically able to perform a task. It commands the most attention and bears least on near-term outcomes. The deployment rate is the speed at which firms place the tools in workers’ hands. The adoption rate is the speed at which workers and organizations redesign their work so that the tools yield real gains, which is a very different matter from mere access. This is the rate that drives productivity. The displacement rate is the speed at which AI shifts from assisting workers to replacing them. This is the rate that drives job losses.
Deployment today runs well ahead of adoption: the tools are ubiquitous, but most organizations have yet to rebuild their work around them, so the productivity gains remain largely uncaptured. Displacement, the conversion of those gains into actual headcount reduction, is slower still, and has barely begun. This is why the technology can feel pervasive while aggregate macroeconomic statistics may look quiet.
Taken together, these four rates and workflow completeness constitute a single engine measuring how AI will impact growth, jobs, prices, and AI-industry revenue.
How AI has already changed the US economy
Our measure of workflow completeness shows that AI has already left its mark on the United States sectors most exposed to it. First, the most exposed sectors are already 11 percentage points more productive than they were before 2022, when OpenAI launched the first widely available AI model. Our measure of AI exposure explains roughly 70 percent of the cross-sector variation in productivity gains, some $642 billion in output above where the pre-2022 trend would have put it.
Second, jobs in the most AI-exposed sectors run roughly 2.9 million below their pre-pandemic trajectory, a shortfall driven mainly by suppressed hiring rather than layoffs. Third, wages in the most exposed sectors have grown slower than productivity, and the size of that gap scales almost exactly with how many tasks in a sector can be automated. The more tasks, the higher the productivity and the slower the wage growth. The gap is widest in the most exposed sectors, information and technology and finance, and negligible in the least exposed, such as construction.
The economy to come
Our model forecasts that AI will add about $1.5 trillion to U.S. output within the next five years and $2 trillion over the longer term. The bulk arrives early because diffusion follows an S-curve: fastest in the middle phase and tapering as sectors approach their automation ceilings.
Growth will come at a cost, with AI expected to eliminate around 20 million jobs over the longer term (roughly 19 to 21 million). This is above the 10 to 12 million implied by Goldman’s estimate that six to seven percent of jobs in the U.S. will be eliminated. The difference is due to methods: we build the estimate from the bottom up, sector by sector, rather than applying a single coarser top-down share across the economy.
On prices, AI is disinflationary. As firms pass productivity gains on to consumers, we project a permanent reduction in the overall price level of around seven to eight percent plus a separate ongoing demand-side drag on inflation of roughly 0.7 to 0.9 percentage points a year. This is larger than the disinflation that followed China’s integration into global trade: China’s entry into the World Trade Organization in 2001 lowered U.S. manufacturing prices by roughly 7.6 percent from 2000 to 2006. However, manufactured goods are only about a quarter of the American consumption basket, so the effect on the overall price level was closer to two percent, roughly a third of the disinflation we project from AI.
On the AI industry itself, the framework points to a revenue pool of around $500-$700 billion a year over the longer term, the flip side of the productivity gains and the displaced jobs.
And on fiscal capacity, the wealth surplus that AI creates, the framework links the pieces in a way that single-issue models miss. The fiscal balance proves to be a function of how fast AI completes a task, the faster the speedup the larger the fiscal capacity. We project that AI generates enough fiscal capacity to fund support for the displaced through existing channels without new programs.
The central policy implication of our framework is that the transition is, on balance, positive. The $2 trillion in productivity gains more than offsets the cost of displaced workers and delivers lower prices across the economy, while the resulting fiscal capacity is sufficient to support the displaced.
The challenge, then, is not whether the gains are large enough, but where they land. Here lies another benefit of our approach. The framework tracks not merely the size of the surplus but also its path. As we show, the surplus is not yet reaching consumers. Instead, it is being captured by companies selling AI products and those firms buying them. The gap between productivity gains and wages in the most AI-exposed sectors illustrates this development.
Competition is what erodes the concentration of welfare gains to a few producers and employers. As rival firms adopt the same tools and engage in competition, the productivity gains convert from profits to lower prices, and the surplus passes to consumers. The policy task therefore cannot simply be “let AI run.”
First, the government must make sure markets selling AI models, AI agents, and AI-adjacent services like computer chips and cloud computing remain open and competitive. Second, the government must lower the barriers that slow employers from hiring displaced workers, so labor can flow to new demand rather than leaving workers stranded and the gains with capital.
It is essential that markets be enabled to absorb displaced workers. If markets do not find displaced workers remunerative new jobs, AI will reduce aggregate income and thus demand. Monetary policy should not be the tool for this: AI’s disinflation is a supply-side gain, and the demand gap it opens is structural, from lost wages, not a cyclical shortfall rate cuts can close. Instead, our model shows that current fiscal channels are sufficient to support the displaced while general consumer welfare benefits via lower prices. The danger is not that AI runs too fast, but that rigid labor markets and weak competition leave the gains stranded with capital.
Our model offers another array of forecasts: some track those of other economists, some disagree, and some, like the fiscal impact, fill gaps the literature has left empty. More importantly, our model provides a single consistent framework to guide policymakers’ response to the AI revolution. Our model predicts much upheaval, but also reasons to be optimistic.
None of which is to claim the last word on any single number. The displacement forecast rests on current data and macroeconomic conditions. A recession could pull near-term displacement forward without altering the long-run picture.
However, the model is built to be recalibrated as the data accumulate. A strength of our model is its dynamism and that it takes multiple rates and dynamics into account. But its core strength is its internal consistency: because every forecast comes from the same engine, they must add up, which is also what exposes rival projections that quietly do not, claiming large productivity gains without saying where the displaced wages or the lower prices go.
The full working paper is available here and a short version here.
Authors’ Disclosures: Benjamin Verschuere and Angus Cameron are employed by Liminal Capital LLC, a CFTC-registered commodity trading advisor and NFA member. Liminal trades futures for client accounts and does not hold AI-company equities in those accounts. The firm’s proprietary account may hold positions in AI companies, and Liminal, its principals, and the authors could therefore benefit if AI-related assets appreciate. This research uses only publicly available data and recommends no asset, trade, or investment position
Articles represent the opinions of their writers, not necessarily those of the University of Chicago, the Booth School of Business, or its faculty.
Subscribe here for ProMarket’s weekly newsletter, Special Interest, to stay up to date on ProMarket’s coverage of the political economy and other content from the Stigler Center.
