![Author ORCID: We display the ORCID iD icon alongside authors names on our website to acknowledge that the ORCiD has been authenticated when entered by the user. To view the users ORCiD record click the icon. [opens in a new tab]](https://www.cambridge.org/engage/assets/public/coe/logo/orcid.png){kind=logo}
Abstract
India has low aggregate AI exposure, ranking 123rd of 149 countries, yet exposure varies
sharply across its firms. The IT-BPM industry, which employs roughly 5.4 million workers in
back-office, coding, and software roles, holds a disproportionate share of jobs highly exposed
to large language models (LLMs). I ask how more AI-exposed Indian firms in this sector
adjusted hiring and productivity around two technology shocks: the cloud and deep-learning
wave of 2015-16, and the LLM shock of 2022. I build an unbalanced panel of 13 publicly listed
IT-services firms over FY2010-FY2025 (180 observations) and measure each firm’s exposure
as the weighted average of the Eloundou et al. (2024) occupational scores across its business
lines. I then estimate a continuous-treatment difference-in-differences model with firm and
year fixed effects. After the 2022 shock, more-exposed firms slowed net hiring (β2 = −0.004
log points per standard deviation of exposure) and raised labor productivity (β2 = +0.075),
though both full-panel estimates are imprecise, pointing to augmentation rather than displacement. The raw data are sharper: high-exposure firms cut annual net hiring from 8.3% to 3.1%,
while low-exposure engineering-R&D firms held near 7.7%. The hiring effect strengthens and
turns significant on the higher-quality subsample (β2 = −0.039, p < 0.05), and the earlier 2016
wave shows no such effect. These are, to my knowledge, the first firm-level tests of these two
waves of AI/ML adoption in the sector most exposed to AI globally.
Supplementary materials
Title
India IT Services Dataset Merged
Description
One Excel workbook consolidating all project CSVs, each on its own sheet, with a README index. Two sheets feed the paper: panel_long, the unbalanced firm-level panel of 13 listed Indian IT firms across FY2010–2025 (180 firm-years, carrying headcount, US-dollar revenue, an Eloundou exposure beta, business segment, and a documented/reported/estimated quality flag per observation); and eloundou_exposure, the 488 O*NET-SOC occupation LLM-exposure scores from Eloundou et al. (2024) that anchor the firm betas and Figure 4. Five auxiliary sheets contain source material: a public-only panel variant with per-cell quality flags; World Bank WDI macro controls; annual and quarterly headcount seed tables with source and reliability notes; and a latest-year firm snapshot. The README sheet labels each sheet "used in paper" or "auxiliary/source," keeping the analysis inputs separable from the raw build materials.
Actions
Title
India IT Services Python Scripts
Description
A single runnable file combining the two-stage pipeline. Part 1 (formerly build_long_panel.py) assembles the panel from hardcoded year-end headcounts and revenues for 13 firms, tags each with an exposure beta and a quality flag, and writes to india_it_panel_long.csv. Part 2 (formerly estimate_long.py) reads that CSV and runs the econometrics in pure Python, with no Stata dependency: it builds log employment, labor productivity, and a firm-standardized exposure index, defines the 2016/2018/2023 breaks and their exposure interactions, and estimates two-way (firm and year) fixed-effects OLS with CR1 cluster-robust standard errors clustered by firm. It prints the main two-break DiD for hiring and productivity, robustness cuts, descriptive hiring means by exposure group, and summary statistics. Run top-to-bottom; it reproduces the paper's numbers (180 observations, N=167 for net hiring). The two source scripts remain separately in the folder.
Actions
Title
India IT Services Stata Code
Description
The Stata counterpart to the Python pipeline, written to replication standards for referees. A global SIMULATE toggle runs it end-to-end on a planted synthetic panel, confirming the estimator recovers a known effect, or on the assembled real data; all toggles and style constants are globals, so the file survives being run in pieces. It builds the same outcomes and firm-standardized exposure measures, defines post-2018 and post-2022 exposure interactions, and estimates the main models with reghdfe, absorbing firm and year fixed effects and clustering by firm, with an inline wild-cluster bootstrap (boottest) for the few clusters. It adds an event-study pre-trend test and a robustness battery (dropping COVID years, Felten–Raj–Seamans exposure, dropping TCS, a levels outcome), then exports seven publication figures and formatted esttab tables. Every graph is wrapped in noisy capture, so a failing figure never halts the run.
Actions