---
title: "\"Extending 'GPTs Are GPTs' to Firms\" is now out in AEA Papers & Proceedings. Lots of talk about..."
canonical_url: "https://typefully.com/t/gsUTw14"
markdown_url: "https://typefully.com/t/gsUTw14.md"
agent_guide_url: "https://typefully.com/ai-agents.md"
post_delimiter: "***"
robots: noindex
source: typefully_shared_draft
share_id: "gsUTw14"
---

"Extending 'GPTs Are GPTs' to Firms" is now out in AEA Papers & Proceedings. 

Lots of talk about AI's impact on employment this week. One way AI will influence labor demand is through firm-level impacts. We built initial descriptive statistics that can shed light on these impacts.

Using data on a sample of 7,894 publicly traded firms in the US, we estimate that the average firm has about 17% of its workers' tasks exposed to LLMs alone, and about 47% exposed if you assume partial integration of LLMs into complementary software workers use in their jobs.

![Screenshot 2025-05-31 at 12.59.28 PM.png](https://d3e0luujhwn38u.cloudfront.net/original/img/original/157010/0ac6657b-7e7c-4ed6-b96d-e18e7e63904d.png)

***

We used the occupation-level exposure scores from our earlier GPTs are GPTs paper (ungated preprint version here: https://arxiv.org/abs/2303.10130) and mapped them to @RevelioLabs data on the workforce composition of 7,894 publicly traded firms.

***

This allows us to estimate firm-level exposure scores by simply multiplying within-firm employment counts in each occupation by that occupation's share of tasks exposed to LLMs at different levels.

***

As in GPTs are GPTs, we use three different levels of exposure: E1 (tasks where a worker could plausibly get a doubling of productivity by using LLMs alone), E2 (tasks that can expect a doubling of productivity by using LLMs once fully integrated with complementary software a worker may use), and E0 (Not exposed). At the occupation level we then calculate our primary exposure measure of E1+0.5*E2 to capture exposure if you assume only partial integration of LLMs with complementary software workers might use in their jobs today or in the future.

***

When aggregated up to the firm-level, we find that the average firm in our sample has 17% of their tasks exposed to LLMs alone (E1), and that share jumps up to 47% assuming partial integration (E1+0.5*E2), and 77% if you assume full integration of LLMs into complementary software (E1+E2).

***

Tech firms and those with a higher share of AI skills in their workforce are among the most highly exposed. Larger firms are on average more exposed to AI than smaller firms (corroborating trends in reported adoption from survey data, e.g. https://www.nber.org/papers/w32319).

![Screenshot 2025-06-01 at 9.42.05 PM.png](https://d3e0luujhwn38u.cloudfront.net/original/img/original/157010/88fdc512-936b-4435-9801-5862b4e4f824.png)

***

There are wide gaps between firm-level exposure estimates and reported firm-level adoption, however, suggesting that a large share of AI adoption is bottom-up, and may not be captured by surveys of managers or executives. We also see this in discrepancies between firm-level adoption surveys (see link above) and individual-level surveys, like those run by Bick, Blandin and @ProfDavidDeming (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4964384)

***

Will aim to post a few more follow-up pieces of analysis that didn't make it into the P&P on cross-industry differences and what this analysis might suggest about superstar effects from AI adoption in the next few weeks. 

Full (short) paper here: https://www.aeaweb.org/articles?id=10.1257/pandp.20251045

***

Shoutout to the team! @danielrock 
@EconoBen @ThankYourNiceAI 
@manlikemishap