A new measure finds AI could double U.S. economic output by helping businesses learn faster—or fail faster

Artificial intelligence could double U.S. economic output, according to new UC Berkeley Haas research. But potential gains vary by industry—and not in the ways one might expect.

The working paper, authored by Martin Beraja of UC Berkeley Haas and Eduard Talamàs of IESE Business School, argues that the debate over AI’s economic impact has been too narrowly focused on either modest productivity gains from job automation or the possibility of explosive growth driven by AI-powered scientific discovery. Beraja and Talamàs contend that a major source of AI’s economic gains will stem from helping companies learn faster—reaching peak productivity sooner and recognizing failing ventures before wasting years on them.

“We must think of AI as an organizational learning technology. It can help companies reach a mature and more productive state much more quickly, without spending decades learning the hard way,” said Beraja, an associate professor in the Economic Analysis & Policy Group at UC Berkeley Haas.

A new metric for measuring AI’s impact

The paper introduces a metric the researchers call VOLT, or the Value of Organizational Learning Technologies, which measures the potential increase in economic output if firms could learn faster. Using 2023 U.S. Census data on business establishments, the researchers calculated that VOLT for the American economy is approximately 2.0, meaning that AI-driven organizational learning technologies have the potential to double aggregate economic output in the long run.

Beraja and Talamàs declined to put a time frame on their estimate, however, because it is impossible to predict the pace at which AI will be adopted by all companies across all industries.

Older and wiser

The researchers’ findings build upon a well-documented phenomena in economics: mature companies are dramatically more productive than average businesses. Establishments older than 46 years are nearly three times larger than the typical U.S. firm and have expected lifespans of 69 years, compared to just 11 years for the average company.

Two things explain this gap. First, companies get better at what they do over time. They learn what their customers want, how to run operations efficiently, and what works in their industry. Employees develop expertise, managers refine their practices, and the organization accumulates a kind of institutional wisdom that makes it more productive.

Second, the weak players drop out along the way. Many businesses launch with flawed ideas, bad timing, or poor execution. They fail and exit. Those still standing after decades are not a random sample—they are the survivors, the ones that had something going for them.

Learn faster, fail faster

AI, the researchers argue, could help companies reach that mature, productive state much faster. Or it could help entrepreneurs recognize a doomed venture before wasting years on it.

This perspective led to one of the study’s surprising findings: Roughly three-quarters of the potential economic gains will come not from making firms more productive, but from extending their lifespans. When organizational learning accelerates, fewer companies spend years discovering they should have never launched, freeing resources to flow toward viable enterprises.

AI tools are already targeting both challenges, the authors note. Enterprise knowledge platforms like Glean aggregate institutional information that would otherwise remain trapped in emails and individual employees’ heads. Prediction systems like Shopify’s SimGym allow entrepreneurs to stress-test business concepts with AI-generated customers before risking real capital. Recent research has shown that AI agents now outperform experienced venture capital analysts at predicting which startups will survive.

An unexpected industry pattern

Beraja and Talamàs’ industry-by-industry analysis of AI impacts revealed a striking map of winners, as well as some surprises. Using census data on firm size and exit rates by age within 80 industries, they plotted their VOLT measure against standard scores of exposure to large language models (LLMs), which capture how easy is to apply these models to the tasks performed in those industries.

The results challenge the conventional wisdom on which industries will be most transformed by AI. It turns out that knowing how exposed an industry is to large language models tells you almost nothing about how much it stands to gain from AI as an organizational learning tool.

“There are industries where AI could be very easy to adopt, but there’s not much value in terms of accelerating organizational learning. Then there are industries where AI may be hard to adopt. But if you were to implement AI in those industries, the value would be enormous,” said Beraja.

The analysis divides the industry landscape into four zones:

The hidden gainers: Some of the largest potential gains will accrue to industries that rarely top anyone’s AI transformation list. Couriers and messengers, truck transportation, and performing arts all rank near the top on organizational learning potential yet score at or below average on LLM exposure. These sectors share a tricky organizational learning problem: young companies fail at high rates before accumulating the operational knowledge, client relationships, and experience that make mature competitors so much more productive and durable. That steep learning curve suggests AI tools can help these firms learn faster—or help entrepreneurs spot a doomed venture earlier.

Those that are already counted: A smaller group of sectors scores high on both measures, meaning AI can both automate their tasks and accelerate their learning curves. Data processing and hosting, telecommunications, and educational services all land here. These are the sectors most analysts have already identified as likely to be most affected by AI.

The middle of the pack: This group includes sectors with potentially high AI exposure but modest organizational learning potential or VOLT. These are industries where AI can do a lot of what workers do, but where the structure of company life cycles limits how much faster learning would add to their success. Publishing and credit intermediation fall here. The gains from AI in these sectors flow primarily through task automation.

Those where gains will be limited: Food services, paper manufacturing, and building materials dealers score low on both learning potential and productivity gains. Their core tasks are difficult for AI to replicate, and the life cycles of these companies don’t lend themselves to large gains from learning. A building materials store, for example, operates a relatively simple business from day one so AI isn’t likely to help much.

Consistent cross-industry measurement

Beraja and Talamàs view their research as complementary to ongoing debates about AI’s transformative potential, which have largely focused on gains from automation or AI-driven scientific discovery. Their paper offers a middle path and a consistent, cross-industry way to measure the substantial economic impact from helping organizations learn faster.

Read the paper:

The Value of Organizational Learning Technologies
By Martin Beraja and Eduard Talamàs
March 2026