Where Will AI-Driven Productivity Gains Diffuse First?

---

---

Multiple research groups have produced estimates of AI applicability or exposure by occupation. We use Microsoft's measure as our primary signal but document alternatives for transparency and comparison.

Primary: Microsoft AI Applicability (This Study)

Alternative Measures

AI Applicability

The AI applicability score aggregates task-level AI assistance rates weighted by task scope. Values range from 0 to ~0.5, with most occupations below 0.25.

ai_applicability = Σ(task_assistance_rate × task_scope_weight)

---

Task Portability (Effective Number of Sectors)

Portability measures how evenly an occupation is distributed across industry sectors. We use the exponential of Shannon entropy (Hill number of order 1):

eff_sectors = exp(-Σ(pₛ × log(pₛ)))

where pₛ = w_cellₛ / Σ(w_cell) for sector s

Interpretation: An occupation with eff_sectors = 12 is distributed as if evenly spread across 12 sectors. Higher values indicate greater portability.

---

Within-SOC Specialization Spread (WSS)

WSS captures heterogeneity in sector-specificity within an occupation. We compute the interquartile range of Specialization Index values:

WSSᵢⱺᵣ = Q₇₅(SI) - Q₂₅(SI)

High WSS indicates both sector-specific and portable titles within the occupation.

---

Empirical Bayes Shrinkage

Cells with few observations have noisy WSS estimates. We apply empirical Bayes shrinkage:

wss_shrunk = B × wss_soc_mean + (1 - B) × wss_cell_raw

B = σ²_between / (σ²_between + σ²_within / n)

Cells with n_titles_cell < 5 are flagged as low_evidence and displayed with reduced opacity.

Rationale

Raw job-posting counts reflect labor market churn (hiring activity). To weight by economic value, we incorporate BLS OEWS employment and wage data.

Allocation Method

OEWS provides national totals per SOC but not sector breakdowns. We allocate using CMap's distribution:

shareₛ = w_cellₛ / Σ(w_cell within SOC)
w_cell_empₛ = emp_soc × shareₛ
w_cell_wageₛ = emp_soc × wage_mean × shareₛ

Coverage Statistics

Occupations are classified into four regimes based on AI applicability and task portability:

Regime Distribution

Productivity Impact Formula

Following Anthropic's Economic Index methodology, we estimate productivity gains as:

productivity_gain = ai_coverage × (1 - 1/effective_speedup) × adoption_rate

effective_speedup = raw_speedup × success_rate × complementarity_discount

Parameter Calibration

Parameters are drawn from Anthropic's January 2026 Economic Index Report:

Parameter	Conservative	Moderate	Optimistic	Source
Speedup	5x	9x	12x	Anthropic benchmark tasks
Success Rate	50%	66%	70%	Task completion quality
Adoption Rate	10%	25%	50%	Workforce AI usage
Complementarity	60% (fixed)			Human-AI integration factor

Scenario Results

Cut-Point Sensitivity

Classifications depend on threshold choices. We evaluate 9 scenarios: {40th, 50th, 60th} × {40th, 50th, 60th} percentile combinations.

Classification stability by sector - the fraction of occupation-sector cells that maintain the same regime across all 9 scenarios.

---

Metric Stability

Comparison of three dispersion metrics (IQR, MAD, SD) for WSS rankings:

Kendall's τ ≥ 0.7 indicates acceptable ranking stability. IQR vs MAD shows strong agreement; IQR vs SD shows moderate agreement.

Methodological Limitations

Interpretive Caveats

• Correlation, not causation: We observe structural associations. We cannot determine whether AI adoption causes standardization or vice versa.
• Static snapshot: Data reflects 2023-2024. AI capabilities and adoption patterns are rapidly evolving.
• Complementary factors: Diffusion depends on capital availability, regulatory environment, and organizational capacity not modeled here.