Firms that report adopting AI in ten sub‑Saharan African countries show higher labour productivity and faster sales growth, though gains vary sharply by country and industry; differences align with technological readiness and industrial structure, suggesting targeted upskilling and regulatory support are needed to realize wider benefits.

Main Finding

Using World Bank Enterprise Survey firm-level data (2007–2024) from 10 sub‑Saharan African countries, the paper finds that AI adoption is associated with higher firm sales growth and—depending on estimation—positive or weakly positive effects on labour productivity. Results are heterogeneous by firm size: AI is positively associated with sales growth (overall and notably for small firms), but high-dimensional fixed‑effects estimates show a significant negative association between AI and labour productivity for micro firms and an overall insignificantly positive effect for the full sample. Information asymmetry (an IA index) is positively associated with labour productivity; firm age is consistently negatively associated with both productivity and sales.

Key Points

• Data and scope
  • Source: World Bank Enterprise Surveys (WBES), 10 countries: Central African Republic, Ethiopia, Ghana, Kenya, Lesotho, Nigeria, Rwanda, Sierra Leone, South Africa, Tanzania.
  • Period: 2007–2024; pooled panel sample (dependent-variable observations ~11.8k).
• Outcome variables
  • Labour productivity growth: computed from sales per worker changes over a three‑year window.
  • Sales growth: computed as change in sales (current vs three years prior) converted to dollars.
• Key regressors and indices
  • AI index: constructed via multiple correspondence analysis (MCA) from innovation and organisational questions (new/improved products, organisational changes, time for idea development, R&D spending).
  • IA index (information asymmetry): proxied via international certification/public listing/external audit indicators.
  • Controls: firm age, firm size (micro/small/medium/large dummies), foreign ownership, industry dummies, year fixed effects.
• Estimation strategy
  • Feasible generalized least squares (FGLS) to address heteroscedasticity/autocorrelation.
  • Robust OLS for baseline estimates.
  • High-dimensional fixed effects (HDFE) linear regression (robust SEs) as a robustness check to absorb unobserved heterogeneity.
• Main empirical results
  • Robust OLS / FGLS: AI positively and significantly associated with labour productivity (small coefficient; significance at conventional levels in pooled model) and with sales growth (stronger and significant).
  • HDFE: AI effect on labour productivity becomes statistically insignificant for the pooled sample; micro firms show a significant negative association (AI coefficient ≈ -0.151, p<0.1).
  • Heterogeneity:
    • Sales growth: AI has a strong positive association overall; particularly positive and significant for small firms.
    • Labour productivity: positive/weak in pooled OLS but negative for micro firms in HDFE; other size categories show mixed/insignificant effects.
  • Firm age: consistently negative for productivity and sales (older firms less able to keep pace with AI-driven change).
  • IA index: positively associated with labour productivity in many specifications.
  • Foreign ownership: mixed effects (negative on labour productivity in one specification).
• Robustness and diagnostics
  • Variance inflation factors low (mean VIF ≈1.08) — multicollinearity unlikely.
  • R-squared varies substantially across size panels; model fit better in some subgroups.

Data & Methods

• Data source: World Bank Enterprise Surveys (publicly available).
• Sample: Panel assembled for 10 SSA countries, standardized survey questions across waves (2007–2024).
• Construction of indices:
  • AI index: MCA on multiple innovation/organisational questions; transformed to index (MCA via Stata v15).
  • IA index: constructed from certification/public listing/external audit indicators.
• Outcome calculations:
  • Labour productivity growth = (sales-per-worker current / sales-per-worker 3 years prior)^(1/(Δyears)) − 1 (method follows prior literature cited).
  • Sales growth computed analogously using current and three-year-ago sales (converted to USD).
• Estimators:
  • FGLS to handle heteroscedasticity/autocorrelation where error structure uncertain.
  • Robust OLS with year dummies and cluster-robust standard errors.
  • HDFE linear regression to absorb high-dimensional fixed effects (country/industry/other unobservables) and test robustness; standard errors corrected per cited method.
• Limitations explicitly noted by authors:
  • Dependence on WBES survey items (mix of subjective and objective responses).
  • Index-based AI measure (innovation proxy) rather than direct measures of specific AI systems.
  • No instrumental-variable (IV) or causal identification strategy implemented (unlike some cited 2SLS/IV studies), so estimates are associational.

Implications for AI Economics

• Policy implications (authors’ emphasis)
  • Targeted upskilling and workforce development are needed to help workers adapt and capture productivity gains, especially in countries/industries lagging in technological readiness.
  • Strengthen digital infrastructure, data ecosystems, and affordable access to AI tools to reduce cross‑country heterogeneity in AI benefits.
  • Support for SMEs: micro firms may face displacement or adoption costs that reduce measured labour productivity—policy should focus on subsidised training, accessible tech support, and adoption pathways that preserve employment.
  • Regulatory and governance frameworks to manage labour displacement risks and encourage responsible AI diffusion.
• Implications for researchers / future work
  • Need for causal identification (IVs, difference‑in‑differences, firm fixed‑effects with adoption timing) to separate selection into AI use from productivity effects.
  • Disaggregate AI measurement: move beyond innovation indices to direct measures (types of AI tools, automation intensity, digital platforms) to clarify mechanisms.
  • Country- and sector‑specific analyses: effects vary by industrial structure and technological readiness; richer country-level covariates and interaction analyses would help.
  • Investigate short‑ vs long‑run effects: micro firms' negative productivity association suggests initial adoption costs/skill mismatches that may reverse over time.
• Managerial implications
  • Firms should pair AI adoption with workforce training and organisational change to realize productivity gains and convert them into sustainable sales growth.
  • Small firms can capture sales benefits from AI (marketing, customer analytics), but micro firms may need targeted support to translate AI into productivity improvements.

Overall contribution: the paper provides one of the first multi‑country empirical assessments of firm‑level AI (proxied via innovation/organisational indices) and its association with labour productivity and sales in selected sub‑Saharan African countries, highlighting positive links to sales and nuanced, size‑dependent effects on productivity that call for policy and research attention to heterogeneity and causal identification.