After the 2022 LLM shock, India's most AI-exposed IT-services firms reduced hiring sharply while raising output per worker, consistent with augmentation rather than mass job loss; the full-sample estimates are noisy but strengthen on a higher-quality subsample and show no analogous response to the 2016 deep-learning/cloud wave.

Main Finding

In India’s IT‑services sector, firms with greater pre‑existing occupational exposure to LLMs slowed net hiring and experienced faster growth in revenue per employee after the 2022 LLM wave. Point estimates are consistent with technology augmentation and resource reallocation rather than outright mass displacement, but precision is limited by a small firm sample and some data-quality constraints.

Key Points

• Sample and scope: Unbalanced panel of 13 publicly listed Indian IT‑services and engineering‑R&D firms, FY2010–FY2025 (180 firm‑years; 167 for net‑hiring analyses).
• Exposure measure: Firm‑level LLM exposure constructed as a shift‑share of occupational exposure scores from Eloundou et al. (2024) (weighted by firm business lines), standardized to mean 0 and SD 1.
• Two “shocks” tested:
  • Deep‑learning/cloud wave: FY2016 (robustness: alternative 2018).
  • LLM wave: FY2023 (post‑ChatGPT).
• Outcomes:
  • Net hiring (∆ ln employees).
  • Labor productivity (ln revenue per employee).
• Raw patterns:
  • High‑exposure IT‑services firms: annual net hiring fell from 8.3% (2016–2022) to 3.1% post‑LLM.
  • Low‑exposure engineering‑R&D firms: hiring stayed near 7.7% post‑LLM.
  • Revenue per employee for high‑exposure firms pulled away after FY2023.
• Main regression estimates (continuous‑treatment DiD with firm and year fixed effects):
  • Exposure × Post‑2022 (β2): net hiring ≈ −0.004 log points per SD of exposure; ln(revenue/employee) ≈ +0.075.
  • Exposure × Post‑2016 (β1): near zero.
  • Higher‑quality (reported‑only) subsample: hiring effect strengthens (β2 ≈ −0.039, p < 0.05).
• Event study: pre‑trends largely flat; shifts concentrated in FY2023 onward; no comparable movement at the 2016 (or 2018) break.
• Robustness & caveats:
  • Some firm‑year observations (64/180) are extrapolated estimates; results excluding extrapolations are stronger.
  • Small number of clusters (13 firms) makes standard cluster‑robust inference unreliable; authors used CR1 finite‑sample correction and advise wild‑cluster bootstrap for definitive inference.
  • Point estimates point toward augmentation (productivity up while hiring slows), but full‑panel coefficients are imprecise and should be treated as indicative.

Data & Methods

• Data sources: annual reports, SEC filings for dual‑listed firms, business‑intelligence data; revenues reported in USD; employment = consolidated year‑end FTE.
• Exposure construction:
  • yi = Σk hk,y * ELk, where hk,y = share of firm i’s workforce in occupation k and ELk = Eloundou et al. (2024) LLM exposure score for occupation k.
  • Resulting firm exposure standardized (across firms and time).
• Empirical model:
  • Continuous‑treatment difference‑in‑differences:
    • Yit = β1 (Ei × Post2016t) + β2 (Ei × Post2022t) + γ′Xit + μi + λt + εit
    • Ei = standardized firm exposure; Post2016t = 1 for fiscal years ≥2016; Post2022t = 1 for fiscal years ≥2023 (fiscal year convention: FY2023 starts after March 31, 2023).
  • Fixed effects: firm (μi) and year (λt).
  • Outcomes: net hiring (∆ ln employees), ln(revenue/employee).
• Identification relies on:
  • Exogeneity of baseline exposure (not endogenously chosen in response to shocks).
  • Parallel pre‑trends in outcomes across exposure levels (tested via event study).
  • Controlling common time shocks via year fixed effects.
• Sensitivity checks:
  • Alternative break date (2018), dropping the largest firm (TCS), excluding COVID years, and keeping only reported (non‑extrapolated) firm‑years.
  • Authors note small‑cluster inference issues and recommend wild‑cluster bootstrap.

Implications for AI Economics

• Sectoral heterogeneity matters: Aggregate unemployment or country‑level statistics can mask important sector/firm‑level reallocation—high‑exposure firms in a globally exposed sector can show distinct hiring and productivity responses even when national exposure is low.
• Evidence toward augmentation: Simultaneous rise in revenue per employee and slowed hiring among more LLM‑exposed firms is consistent with LLMs augmenting worker productivity and enabling reallocation (fewer new hires, higher output per worker) rather than immediate mass layoffs.
• Timing and technology specificity: The effects are concentrated after the 2022 LLM wave, not the earlier deep‑learning/cloud rollout, indicating that different AI generations (LLMs vs prior ML/cloud tooling) can have materially different labor impacts.
• Policy relevance: Adjustment policies should focus on task‑ and firm‑level transitions (retraining, redeployment, support for firms adopting LLMs) rather than only macro unemployment insurance—because impacts appear to be redistribution/augmentation within sectors rather than economy‑wide job destruction.
• Measurement and methodological takeaways:
  • Firm‑level, occupation‑weighted exposure indices are a useful tool to detect heterogeneous impacts of AI.
  • Small‑sample inference in firm‑level studies needs careful treatment (wild‑cluster bootstrap, more clusters).
  • Future work should expand the firm universe, include wage and composition outcomes, obtain non‑extrapolated employment measures, use task‑level adoption indicators, and track longer post‑LLM horizons to distinguish short‑run hiring delays from permanent displacement.

Overall, the paper provides early firm‑level evidence from the industry most exposed to LLMs that the LLM shock is associated with a slowdown in hiring and a rise in per‑worker revenue at more‑exposed firms—consistent with augmentation and reallocation—while underscoring the need for larger samples and careful inference to draw definitive causal conclusions.