The AI Measurement Problem: When Markets Price What Firms Can’t Easily Disclose

The AI measurement problem

Tania Babina, in this paper, analyzes how should researchers measure firms’ AI efforts, and how does that measurement choice shape what we can conclude about AI’s economic impact on firms, workers, and markets.

The author argues that “AI data” is not a single object, and that the same research question (e.g., “How has AI affected employment?”) can yield different, seemingly conflicting answers depending on whether the researcher uses firm-investment data, occupational task-exposure data, disclosure-based data, or usage data. Babina’s  goal is to give researchers a framework for choosing the right measure for their question.

Babina organizes AI datasets along six dimensions: invention vs. use, internal vs. outsourced capability, breadth vs. depth, coverage of all vs. specific AI applications, actual vs. perceived capability, and ex-ante vs. ex-post measurement. She then proposes quality criteria for evaluating AI datasets: coverage/representativeness, comprehensiveness, granularity, timing/panel structure, construct validity, manipulability.

Findings

• AI adoption is highly uneven, concentrated among larger firms with more high-skilled workers.
• AI investment correlates with higher sales growth and valuation, though effects often lag.
• Productivity gains are harder to detect in aggregate firm-level data than in narrow task-level settings.
• Labor effects are nuanced: AI substitutes for specific tasks while firm-level employment often still grows net.
• The financial sector is a particularly informative setting, since AI affects credit allocation, investor behavior, trading, and market efficiency simultaneously.

Understanding Firms’ AI Efforts and Their Economic Impact
Author: Tania Babina
From: University of Maryland

Letting markets measure AI: Evidence from 380 trillion tokens

Nicola Borri, Aleh Tsyvinski and Yukun Liu, in this paper, analyze how does equity markets price firms’ exposure to AI, and what is the magnitude, drivers, and scope of the resulting “AI Premium.”

The authors use a licensed proprietary dataset from OpenRouter, Inc. covering 380 trillion tokens of realized AI consumption across more than 400 large language models from January 2024 through April 2026 i.e. about 2% of global monthly AI token usage. Unlike prior task-based or usage-based approaches, their measures are market-implied, built from stock-price comovement rather than surveys or technical automatability scores. The methodology proceeds in three steps:

1. AI Factor construction: They build a weekly AI Factor as the first principal component of standardized log growth in three series namely, total tokens, dollar spending, and distinct users, which explains 56.5% of joint variance.
2. AI Beta estimation: Each firm’s AI exposure is estimated via 13-week rolling regressions of weekly stock returns on the AI Factor, controlling for market returns, updated weekly to avoid look-ahead bias.
3. Portfolio sorts and Fama-MacBeth regressions: Firms are sorted into AI-beta quintiles weekly, with returns compared across high vs. low exposure, and cross-sectional regressions test the AI premium controlling for size, value, profitability, investment, momentum, reversal, leverage, and accruals.

Findings

• A value-weighted long-short strategy (high minus low AI-beta) earns 64.1 basis points per week, or around 56 bps after Fama-French five-factor and momentum adjustment. This is a large, statistically robust AI premium. 
• The premium survives controls for tech-sector and AI-themed equity returns , indicating it reflects genuine AI-consumption exposure rather than a broader tech rally or AI “hype”.
• The premium concentrates on the intensive, frontier margin of AI use i.e. closed-source models, paid/seasoned users, and long prompts, rather than casual or open-weight use. 
• Geographically, the premium is significant in developed markets but statistically insignificant in emerging markets, including China. 
• At the labor/skills level, occupations with more nonroutine interactive content (communication, persuasion, instruction) show higher market-implied AI exposure, while analytical/scientific/operations-control skills show lower exposure . 
• The paper finds early evidence of a positive “agentic premium,” as tool-invoking agentic AI requests rose from near zero in 2024 to roughly half of total tokens by the sample’s end. 

The authors interpret the positive premium through a “transition-risk” lens: AI reallocates economic rents across firms and sectors, and investors demand compensation for holding firms whose fortunes are most tied to this systematic reallocation.

AI Premium
Authors: Nicola Borri, Aleh Tsyvinski, Yukun Liu
From: Luiss University, Yale University, University of Rochester