Developers who use generative-AI tools frequently report notable gains in productivity and code quality, with usage frequency the strongest predictor of perceived benefit. Adoption spreads from Enthusiasts to Pragmatists while security worries and a stalled market for AI testing tools slow broader diffusion.

Main Finding

Broader and more frequent use of AI-assisted development tools is associated with higher developer-reported productivity, higher perceived code quality, and stronger intent to increase future AI use. Usage and perceived benefits are larger for coding than for testing, and three developer segments (Enthusiasts, Pragmatists, Cautious) differ sharply in adoption, perceptions, and organizational policy presence. Ease of integration and current frequency of use positively predict adoption intent, while security concerns reduce it.

Key Points

• Usage patterns

  • 67% of respondents use AI coding tools “often” or “always.”
  • Median number of AI coding activities = 5 (out of 11); median testing activities = 2 — evidence of a “testing gap.”
  • Most common coding use case: code completion (67%). Most common testing use case: test-case generation (72%).

• Perceived productivity and quality

  • Developers report larger productivity gains for coding than testing (coding mean on 1–6 scale = 3.92, median = 4; testing mean = 3.14, median = 2).
  • ~60% report >3 hours saved for coding; ~48% report >3 hours saved for testing.
  • ~74% report improvements in code quality; ~68% report improvements for testing contexts.
  • Perceived productivity and perceived code quality are positively correlated (developers perceive little trade-off between speed and quality).

• Statistical associations (selected)

  • AI Coding Tool Index — perceived productivity (coding): Spearman ρ = 0.347 (p < 0.001).
  • Frequency of AI coding use — perceived productivity (coding): ρ = 0.458 (p < 0.001).
  • AI Testing Tool Index — perceived productivity (testing): ρ = 0.357 (p < 0.001); frequency ρ = 0.479 (p < 0.001).
  • AI Coding Tool Index — Perceived Quality Index (PQI): ρ ≈ 0.236 (p ≈ 0.004); AI Testing Tool Index — PQI: ρ ≈ 0.289 (p ≈ 0.0004).
  • Adoption intent correlates positively with perceived productivity, PQI, and breadth of AI usage (stronger for testing-related measures).

• Predictors of adoption intent (linear regression)

  • Model R² = 0.295 (p < 0.001).
  • Significant positive predictors: frequency of AI testing tool use (β = 0.38, p < 0.001), ease of integration (β = 0.29, p = 0.001).
  • Significant negative predictor: security concerns (β = -0.17, p = 0.001).
  • Cost, PQI, organizational policy, and developer interest were not significant in this model.

• Developer segments (k = 3)

  • Enthusiasts (n = 56): high usage breadth, high PQI, high intent. 59% report organizational AI coding policy.
  • Pragmatists (n = 53): moderate usage and outlook; intermediate attitudes; 26% report policy.
  • Cautious (n = 38): low usage, lower perceived benefits, low intent. 5.3% report policy.
  • Cluster membership is significantly associated with presence of organizational AI policies (χ² = 45.59, p < 0.001).

• Future skills & concerns

  • Developers rate orchestration, data management, prompt engineering, and model selection as important future skills.
  • Reported challenges: model hallucinations, security risks, and increased testing complexity.

Data & Methods

• Sample

  • N = 147 professional software developers.
  • Survey distributed Oct–Nov 2025 via the author’s channels; voluntary, anonymous responses.
  • Respondents varied by role, experience, and geography (detailed demographics in supplement).

• Instrument

  • 34-question structured survey (mix of single-item and matrix items).
  • Most perceptual items on 5-point Likert scales; some productivity measured on a 1–6 scale.
  • Full instrument and cleaned dataset provided in supplementary material.

• Key composite measures (indices)

  • Intent to Increase Usage Index (dependent variable)
  • Perceived Quality Index (PQI)
  • AI Coding Tool Index (breadth of coding tool usage)
  • AI Testing Tool Index (breadth of testing tool usage)
  • Mean Concern Index (challenges/risks)
  • Internal consistency checked (Cronbach’s alpha; adequate for exploratory work).

• Analytic approach

  • Descriptive statistics for usage patterns.
  • Spearman rank correlations for ordinal associations.
  • Linear regression predicting Intent to Increase Usage Index (controls and contextual factors included).
  • PCA for dimensionality reduction followed by k-means clustering for segmentation (k chosen = 3 based on elbow/silhouette and interpretability).
  • Significance threshold α < 0.05. Results are associational (cross-sectional, perception-based).

• Limitations called out by the authors

  • Perception-based measures (no objective productivity or code-quality metrics).
  • Cross-sectional sample of modest size (147) and convenience distribution—potential selection and reporting biases.
  • Associations cannot be interpreted causally.

Implications for AI Economics

• Measured productivity gains and adoption patterns

  • If subjective productivity gains reflect true output gains, broad AI tool adoption could raise developer labor productivity — with implications for firm-level output, project completion times, and developer hours per unit of software.
  • Heterogeneous adoption (Enthusiasts vs Cautious) implies uneven productivity gains across firms, teams, and sectors, complicating economy-wide aggregation of AI effects.

• Complementarities and organizational adoption

  • Ease of workflow integration is a strong predictor of intention to adopt; complements (tooling, integration, policies, training) matter for realized productivity gains. Firms that invest in integration and governance may capture more of the benefits.
  • The association between organizational policy presence and high-adoption clusters suggests coordination and policy as complements that accelerate diffusion — relevant for firms considering governance investments.

• Skills, labor demand, and reallocation

  • High importance placed on orchestration, data management, and prompt engineering suggests shifting skill demand toward AI-complementary tasks. Human capital investment (retraining) may increase returns to developers who acquire these skills and reduce returns for those who do not.
  • Heterogeneity in developer outlook implies variable wage-market responses; early-adopting firms may command productivity premiums.

• Quality, testing gap, and risk externalities

  • The “testing gap” — lower AI usage and smaller perceived gains in testing — raises risks if AI is primarily accelerating coding without equally improving or validating quality. Potential externalities: faster deployment but higher latent defects if testing adoption lags.
  • Security concerns negatively affect adoption intent; unresolved security/IP risks can suppress realized economic benefits and increase compliance/insurance costs.

• Measurement and policy caution

  • Reliance on perception-based gains risks overestimation of AI’s economic impact. Objective, artifact-level and firm-level causal studies are necessary to quantify true productivity, error rates, and downstream costs.
  • Public policy and firm governance (IP rules, security standards, disclosure) will shape adoption dynamics, competition, and possibly market structure (firms that manage AI risk effectively may gain relative advantage).

• Research directions for AI economics

  • Estimate causal effects of AI tools on developer output, bug rates, and time-to-market using experimental or longitudinal designs.
  • Model heterogeneous adoption and uptake costs across firms, and equilibrium effects on wages and employment in software labor markets.
  • Study complementarities between AI tooling, organizational policy, and training investments to assess returns to firms and social welfare.
  • Quantify potential externalities from under-adopted testing practices and design interventions (standards, audits) to mitigate quality/security risks.

Summary note: the study provides useful early evidence on developer perceptions and adoption correlates but is limited by self-reported, cross-sectional data. For robust AI-economics conclusions about productivity, quality, and labor-market impacts, follow-up studies using objective measures and causal identification are needed.