AI and Human-Capital Risk
Research in progress on whether firm-level AI exposure can help build an equity portfolio that partially hedges AI-related human-capital risk.
At a glance
- 2026-NOW
- RESEARCH / DATA / PORTFOLIO DESIGN
- AI EXPOSURE / LABOR ECONOMICS / EQUITY PORTFOLIOS
- The data pipeline and portfolio design are in progress. No empirical findings or hedge performance exist yet, and none are claimed.
The question
People hold much of their economic wealth in future labor income. If AI changes the demand for occupations unevenly, households become exposed through their human capital as well as their portfolios.
The research question is:
Can firm-level AI exposure be used to construct an equity portfolio that partially hedges AI-related human-capital risk?
“Partially” carries weight. A tradeable equity portfolio cannot reproduce any individual worker's employment risk. The test is whether systematic differences in firms' AI exposure offset part of it usefully.
The approach
The research connects several levels that do not arrive in one dataset:
- Occupational exposure to AI.
- Employment and wage composition across occupations.
- A mapping from occupations toward firm-level exposure.
- Equity portfolio construction using those exposure measures.
- Evaluation of whether the portfolio behaves like a partial hedge for AI-related human-capital risk.
The inputs include occupational AI exposure measures, separate language-model and image-generation exposure data, OEWS employment data and SOC occupation crosswalks. The hard part is preserving a defensible connection between occupations, firms and a tradeable portfolio while estimating exposure at each level.
Methodology
Start from measures that estimate how strongly tasks within an occupation may be affected by AI capabilities. Exposure is not job loss or productivity gain; it marks where technological capability may matter.
Use occupational employment and wage data to see how labor and income are distributed across occupations. This supplies the human-capital side of the problem.
Translate occupational exposure toward firms using their labor composition or the best available mapping. A poor mapping here makes every later number look precise and mean little.
Sort or weight equities using the firm-level exposure measure. The design has to separate the intended exposure measure from unrelated firm characteristics as far as the data allows.
Test whether the resulting portfolio offsets part of the labor-income risk associated with AI exposure. The evaluation has to separate a plausible hedge relationship from a coincidental return pattern.
My role
I developed the research question and assembled the pipeline that makes it testable: collecting and reconciling the exposure datasets, occupational employment data and SOC crosswalks; designing the firm-level measure; and specifying how that measure enters portfolio construction.
The finance background shapes the question. The data work forces precision about what “AI exposure” actually represents at each stage.
Decisions
Separate exposure from outcome
An occupation can be exposed to AI without employment rising or falling because of it. The research refuses to quietly turn a capability measure into a displacement forecast.
Keep the crosswalk visible
Occupation classifications change, so crosswalking SOC codes decides which workers and wages enter each exposure measure. It is methodology, not clerical work.
Validate the measure before building the portfolio
No portfolio method repairs a weak firm-level measure. The exposure construction gets inspected before any optimization runs.
Claim only a partial hedge
Human capital is individual, illiquid and moved by more than public-equity returns. The portfolio is framed as a possible partial hedge, never as insurance against technological change.
What exists
The research work includes:
- Occupational AI exposure data
- Language-model and image-generation exposure measures
- OEWS employment and wage data
- SOC classification crosswalks
- A planned labor-risk indicator
- A firm-exposure construction problem, still open
- An equity portfolio design and evaluation framework
The case study stops at methodology. There are no final estimates, charts or performance figures to show yet.
What it is teaching me
One distinction keeps proving itself: a useful financial variable is not the same as an interesting social measure.
To reach a portfolio, AI exposure has to survive several translations: tasks to occupations, occupations to workers, workers to firms, firms to returns. Each step introduces assumptions, and the final result cannot be more defensible than that chain.
The portfolio is only as defensible as the occupation-to-firm mapping behind it.