AI economics / measurement

How exposed is an industry to AI?

From occupations and wages to comparable industry measures.

Mario Diez Martínez · January 2026

“AI exposed” is not a primitive: exposed through which jobs, weighted by what, and mapped to which industry definition?

This project treats exposure as a measurement problem. It combines Felten, Raj, and Seamans exposure scores with BLS employment and wage data, then maps NAICS industries into Fama-French portfolios through an auditable SIC crosswalk.

The output is a reusable index for later work on employment, prices, investment, and markets. It is not evidence that AI has already caused those outcomes.

43 / 49Fama-French portfolios with valid coverage

0.968Correlation between 2023 AIIE and ExpWB

2Independent exposure constructions

2021 / 2023OEWS measurement vintages

01 / Construction

Occupation to portfolio

01
Start with occupation- and industry-level AI exposure from Felten, Raj, and Seamans.
02
Merge BLS employment and mean wages by occupation-industry cell.
03
Build equal- and wage-bill-weighted exposure inside each NAICS industry.
04
Crosswalk NAICS to SIC and Fama-French 49; compare equal and BEA value-added weights.

02 / Measure

Economic weight inside the workforce

The bottom-up measure gives more influence to exposed occupations that account for more of an industry's wage bill.

{ExpWB}_{i} = \frac{\sum_{o} {AIOE}_{o} L_{i o} w_{i o}}{\sum_{o} L_{i o} w_{i o}}

(1)

I keep a direct industry score, AIIE, alongside ExpWB. Agreement between them is informative because they reach the industry from different directions.

03 / Aggregation

NAICS industries into FF49 portfolios

Each NAICS-4 industry is assigned to the modal Fama-French portfolio implied by its SIC crosswalk. Portfolio exposure is then calculated two ways: an equal-weighted mean and a BEA value-added-weighted mean.

X_{p}^{VW} = \frac{\sum_{i \in p} X_{i} {VA}_{i}}{\sum_{i \in p} {VA}_{i}}

(2)

z_{p} = \frac{X_{p} - \bar{X}}{sd (X)}

(3)

Every published variant records the source vintage, aggregation rule, and whether standardization was applied. Six FF49 portfolios are excluded because the available SIC-NAICS mapping produces no valid NAICS-4 match.

Horizontal bar chart showing the five highest and five lowest AI-exposure Fama-French industries in 2023. — Figure 01 Direct 2023 exposure is highest in insurance, software, banks, trading, and telecom; it is lowest in coal, agriculture, construction materials, construction, and mining. Scores are standardized across covered portfolios.

Scatter plot comparing direct industry AI exposure with wage-bill-weighted occupational exposure across 43 Fama-French industries. — Figure 02 The direct AIIE and bottom-up ExpWB measures are strongly aligned across 43 portfolios (r = 0.968), despite different construction paths.

04 / Use and limits

An input, not an outcome

The index can be merged into labour, investment, price, or market panels without rebuilding the classification work each time. Both equal-weighted and BEA value-added-weighted versions are published, together with mapping audits.

Exposure is not adoption, productivity, substitution, or causality. Six Fama-French portfolios do not pass current coverage checks, and crosswalk choices can matter at the edges. The public tables keep those limits visible.

05 / Tools

Tools and sources

PythonpandasNumPySciPyBLS OEWSBEA APIFelten exposure dataNAICSSICFama-French 49MatplotlibYAML

The repository contains construction code, mapping audits, and aggregate exposure measures. Source licenses and API credentials remain outside the release.