Research, in Motion.

AI Exposure and Firm Outcomes

This study separates theoretical AI capability, observed industry usage, and firm engagement measured from public SEC filings. Event-study and panel tests ask whether that firm-level signal predicts market reactions or later operating outcomes, while specification changes expose how fragile a pricing story would be.

Optimised for desktop. Complete on smaller screens.

Answer first

A firm-level AI measure built from 10-K text produces no robust stock-price reaction to ChatGPT and does not predict later revenue per employee or operating margin.

Boundary. Industry capability and observed usage correlate at 0.90 after aggregation, while event-study conclusions remain sensitive to pre-trends and the choice of risk model.

0.90Industry capability–usage correlation
0.20Firm–industry correlation
p = 0.98Operating-margin test
Question01 / 04
01Question

Research premise

Turn public text into an empirical test

Public filing language becomes useful only when it can be compressed into a measurable signal and survive tests against observable outcomes.

Text to testPublic language → empirical signal → outcomes
Public filing language is compressed into one pre-ChatGPT firm-level engagement measure. The same signal is then tested against market pricing, productivity, and profitability; none produces a robust predictive result.

Does a firm's exposure to artificial intelligence show up in its stock returns or its operating fundamentals?

What the design can test

Whether a public, pre-ChatGPT firm-level signal is associated with market reactions or later operating outcomes.

What it cannot establish

Whether AI caused a particular return, productivity change, or management decision.

02Measurement & design

Measurement

Three measures, two levels of analysis

Capability, observed usage, and firm engagement describe different layers of the same phenomenon.

Occupation scores describe theoretical capability. Public usage data describes what industries appear to be doing. SEC 10-K language supplies a distinct firm-level engagement measure before the ChatGPT release.

Measurement notebook

Schematic marks, not observations

Field notebook comparing capability, usage, and firm-level AI measurementsTriangles represent theoretical capability, circles represent observed usage, and squares represent firm engagement measured from public SEC 10-K text. The mark positions are schematic, not plotted observations. Industry capability and usage correlate at 0.90 after aggregation, while the firm measure correlates only 0.20 with industry capability. The firm measure produces no robust stock-price effect and does not predict later productivity or operating margin.
CapabilityOccupation scores aggregated to industries
Observed usagePublic AI-usage data aggregated to industries
Firm engagementAI intensity measured from public SEC 10-K text
Distinct exposure is not predictive success. The firm measure has only a weak 0.20 correlation with industry capability. It produces no robust stock-price reaction to ChatGPT and does not predict later revenue per employee or operating margin. The 0.90 industry correlation also means capability and usage cannot be separated after aggregation. Marks are schematic, not plotted observations. Review the public source.

Identification consequenceAt industry level, capability and usage move together at r = 0.90. The firm measure relates much more weakly to industry capability at r = 0.20, making it distinct enough to test, but not automatically useful for prediction.

03Measurement & design

Research design

Test the signal at three checkpoints

The analysis moves from measurement validity to market pricing and then to operating fundamentals.

STEP 01

10-K engagement measure

Public pre-ChatGPT SEC 10-K text is converted into a firm-level AI engagement measure, distinct from occupation-based capability and observed industry usage.

STEP 02

Event-study specifications

ChatGPT-window abnormal returns are re-estimated under a single index, a market model, and Fama–French five factors so sign changes remain visible.

STEP 03

Firm-outcome panel

A 2017–2024 panel with firm and year fixed effects tests subsequent revenue per employee and operating margin rather than inferring outcomes from attention.

04Evidence

Market evidence

The pricing result changes with the risk model

A market response should survive reasonable choices about how abnormal returns are measured.

Published evidence02 / 03
Coefficient plot showing the estimated ChatGPT-window return changes sign across three risk models and is not significant at five percent

Coefficient plot showing the estimated ChatGPT-window return changes sign across three risk models and is not significant at five percent. Public AI Exposure working paper.

Text equivalent

ChatGPT-window abnormal return per one standard deviation of firm AI engagement: single index −0.11% (p = 0.58), market model +0.43% (p = 0.08), and Fama-French five-factor −0.35% (p = 0.07). No estimate is significant at five percent, and the sign changes with the risk model.

Why this mattersThe coefficient changes sign across the single-index, market, and Fama–French specifications. No estimate clears the five-percent threshold, so one attractive specification cannot carry the conclusion.

05Evidence

Operating evidence

The signal does not travel into fundamentals

Attention in filings is only economically meaningful if it anticipates something observable in firm performance.

Published evidence03 / 03
Two outcome cards reporting no predictive effect of firm AI engagement on revenue per employee or operating margin from 2017 to 2024

Two outcome cards reporting no predictive effect of firm AI engagement on revenue per employee or operating margin from 2017 to 2024. Public AI Exposure working paper.

Text equivalent

In a 2017–2024 panel with firm and year fixed effects, firm AI engagement does not predict revenue per employee (coefficient 0.011, p = 0.30) or operating margin (coefficient 0.000, p = 0.98). Flat pre-trends support the null interpretation.

In the 2017–2024 fixed-effects panel, firm engagement does not predict revenue per employee or operating margin. Flat pre-trends make the null result more credible; they do not turn it into proof that AI never matters.

06Interpretation & record

Interpretation

Distinct measurement is not predictive success

The contribution is a disciplined separation between a measurable construct and a predictive claim.

A firm-level AI measure built from 10-K text produces no robust stock-price reaction to ChatGPT and does not predict later revenue per employee or operating margin.

Verified metrics

  • 0.90
    Industry capability–usage correlation

    After four-digit NAICS aggregation, capability and observed usage are too closely aligned to identify separate industry stories.

    Verify in public source
  • 0.20
    Firm–industry correlation

    The 10-K engagement measure has only a weak relationship with industry capability, supporting its use as a distinct firm-level signal.

    Verify in public source
  • p = 0.98
    Operating-margin test

    The fixed-effects panel reports a near-zero coefficient for later operating margin, consistent with no predictive effect.

    Verify in public source
What the design can test

The public firm-level measure is empirically distinct, but shows no robust pricing or operating-performance signal in these designs.

What it cannot establish

Industry capability and observed usage correlate at 0.90 after aggregation, while event-study conclusions remain sensitive to pre-trends and the choice of risk model.

07Interpretation & record

Public record

Evidence archive

Every number shown above can be traced to a public repository or archived research output.

Public evidence registerReproducible sources only
  1. 01
    Comparison showing a 0.90 industry capability-usage correlation and a 0.20 firm-engagement correlation with industry capability

    Industry AI capability and observed usage correlate at 0.90 after four-digit NAICS aggregation, so they cannot be separated as industry measures. Firm engagement measured from pre-ChatGPT public 10-K text correlates only 0.20 with industry capability and is a distinct firm-level signal.

    Verify
  2. 02
    Coefficient plot showing the estimated ChatGPT-window return changes sign across three risk models and is not significant at five percent

    ChatGPT-window abnormal return per one standard deviation of firm AI engagement: single index −0.11% (p = 0.58), market model +0.43% (p = 0.08), and Fama-French five-factor −0.35% (p = 0.07). No estimate is significant at five percent, and the sign changes with the risk model.

    Verify
  3. 03
    Two outcome cards reporting no predictive effect of firm AI engagement on revenue per employee or operating margin from 2017 to 2024

    In a 2017–2024 panel with firm and year fixed effects, firm AI engagement does not predict revenue per employee (coefficient 0.011, p = 0.30) or operating margin (coefficient 0.000, p = 0.98). Flat pre-trends support the null interpretation.

    Verify

Public sources and repositories