Generative AI and Firm Values

Is generative AI already priced in public equity markets?

Andrea L. Eisfeldt, Gregor Schubert, and Miao Ben Zhang study how generative AI affects firm value as a labour-saving technology in « Generative AI and Firm Values ».

They ask whether stock prices around ChatGPT’s release reflect differences in firms’ workforce exposure to generative AI, and which mechanisms can explain the repricing in capital markets.

They combine this exposure with event-study returns, analyst forecasts, profitability, job postings, and wage data to interpret market pricing to conclude:

• Out of 19,265 analysed tasks, 14% are directly exposed and 22% indirectly exposed. Occupations average exposure 23.4%, and firm exposure averages 35.4% across 2,518 firms.
• Markets reprice exposure immediately after the release of ChatGPT: the « Artificial-Minus-Human » (AMH) portfolio built by the authors earns 0.45% per day over the following two weeks, with a market-adjusted alpha of 0.44% and a five-factor alpha of 0.35%.
• The cumulative repricing is large: the same AMH portfolio earns about 5% over the two post-release weeks, consistent with a sizeable valuation wedge between firms with high versus low workforce exposure.
• The ability to handle data assets within the workforce amplifies returns more than firm’s access to data, with a regression coefficient 56 times higher at the 5% significance level.
• Cash-flow expectations move with exposure, with a slight positive relationship with EPS forecasts and gross profitability, and a large one on long-term earnings growth.
• A one standard deviation increase in exposure is associated with 8% fewer job postings and 0.6% lower hourly wages post-ChatGPT, with effects 66% and 97% stronger when exposure comes from core tasks.

This research suggests exposure-driven valuation gains come with workforce reallocation and wage pressure, with the labour and data channels identified as immediately material factors.

Analysts can combine workforce exposure with disclosed data-asset reliance to stress-test both the upside (efficiency) and the downside (execution, workforce, privacy, and governance) of AI adoption.

As a limitation, the chosen 2-week time window balances investors’ digestion of complex information against contamination from other events but can be challenged.

The authors also relied on ChatGPT to classify tasks statements: the ability of an LLM to do so can be considered a joint hypothesis that mitigates the study’s results.