A single firm's AI coding push coincided with a doubling of developers' merged pull requests to 2.09x pre-mandate levels by April 2026; evidence links most of the gain to voluntary adoption and accumulated use, and code-review work shifted heavily toward automation without higher revert rates.

Main Finding

In a longitudinal, instrumented case study of a documented enterprise “2×” AI mandate (mid-2025) at a mid-sized, AI-forward software firm, per-capita merged pull-request (PR) throughput roughly doubled by April 2026 (2.09× the pre-mandate baseline). A staggered difference‑in‑differences analysis attributes most of that within‑developer gain to AI tool adoption and a dose‑response that grows with accumulated use (about a 1.5× within‑developer gain attributable to adoption+use, reaching ~2× after ≈9 months of tool use). The mandate acted mainly as a catalyst for adoption rather than an instantaneous productivity boost. Adoption also restructured review: per-reviewer load roughly doubled and automated review became dominant, while coarse short‑horizon quality indicators (merge and revert rates) remained roughly stable.

Key Points

• Sample and outcome
  • Developer-month panel: 802 developers (January 2024–April 2026).
  • PR-level data: 196,212 non-bot PRs across 364 repositories.
  • Primary metric: merged PRs per engineer per month (the firm’s own “progress” metric).
• Magnitude & timing
  • Organization-level per-capita PR throughput ≈2.09× pre-mandate by Apr 2026.
  • Within-developer decomposition: an immediate adoption jump plus a growing accumulated-use effect → combined ≈1.5× (robust to organization-wide monthly shocks); reaches ~2× by nine months on the tool.
• Heterogeneity
  • Gains were broadly shared across seniority (from ICs to principals).
  • Gains concentrated in newer repositories; legacy repositories saw little effect.
  • The observed gain could not be cleanly separated by model-generation (Sonnet 4.5, Opus 4.5, Opus 4.6) within the rollout.
• Organizational/process effects
  • Automated review (bots, CI, policy tools) accounted for a large and rising share of review events (~38% of review rows).
  • Per-reviewer human review load roughly doubled as AI-authored PR volume increased.
  • Merge and revert rates showed little systematic change over the study window.
• Cautions/limitations
  • Adoption and usage intensity were not randomized; early adopters already had higher pre-adoption throughput.
  • The firm is a near‑ideal, permissive environment (unlimited seats/spend), so results represent an upper-envelope case rather than an industry average.
  • The “created-by-ai” PR label is an internal convention and not independently validated; per-PR comparisons are treated as correlational.

Data & Methods

• Data sources
  • Full internal PR and review history (titles, authors, timestamps, sizes, review events, comments, labels).
  • AI-tool telemetry (per-developer, per-month lines suggested/written, token spend, acceptance rates) from Cursor and Claude Code; additional “Other” tools inferred from PR labels.
  • HR role data for ~91% of developers (mapped to seniority ladder).
• Constructed datasets
  • Developer-month panel (802 devs; estimation sample restricted to 564 devs with ≥3 active months: 451 adopters, 113 never-adopters).
  • PR dataset with human/automated review classification; ~30.2% of PRs labeled “created-by-ai.”
• Key measurement choices
  • Human review identified by excluding flagged bot/service accounts and templated bot messages; pickup/review-lead anchored at first non-author human review or PR-thread comment.
  • Time-to-merge decomposed into coding lead, pickup, review lead, and total cycle time.
• Identification / empirical strategy
  • Staggered difference‑in‑differences (DiD) aligning each developer on their first observed AI-tool use (adoption), with developer fixed effects to compare each developer to their own pre-adoption baseline.
  • Baseline regression: log(PRs_it) = α_i + β1 * Adopted_it + β2 * PostMandate_t + ε_it; preferred specifications replace PostMandate with calendar-month fixed effects to absorb organization-wide shocks.
  • Dose–response analysis for accumulated use (growing returns with experience).
  • Robustness: event studies, clustered SEs by developer, Poisson models (to retain zeros) — conclusions robust across specifications.
• Major threats addressed
  • Time-varying confounding: never-adopters (113 devs) anchor organization-wide calendar trend; month fixed effects absorb shared shocks.
  • Pre-trend checks show flat pre-adoption trajectories for adopters, supporting parallel trends assumption for within-developer inference.
• Limitations explicitly noted by authors
  • Non-random adoption/use implies findings “strongly implicate” an adoption-and-use channel but should not be read as exact causal magnitude.
  • Internal PR labeling and tool-rollout timing limit precise separation of model-generation effects.

Implications for AI Economics

• Mandates can produce large gains, but via adoption + accumulated human investment, not instantaneous model magic
  • The study shows that enterprise-level “doubling” is attainable in a favorable environment, but the primary mechanism is developer adoption and learning (complementary human investment), reinforcing the literature on complementary investments for general-purpose technologies.
  • Policy/management implication: mandates alone are insufficient; organizations must enable adoption, provide time/training, and allow experience accumulation to realize large returns.
• Redistribution of work and hidden costs matter
  • Productivity gains in output volume shifted work downstream (review) rather than eliminating it: per-reviewer human load rose and automated tooling ate a larger share of review.
  • Short-horizon, coarse quality proxies (merge/revert rates) were unchanged, but longer-term costs—maintenance burden, technical debt, knowledge dilution, and cognitive/intent debt—remain plausible and were not fully captured; economists and firms should value both throughput and downstream/maintenance externalities.
• Heterogeneous returns emphasize where value accrues
  • Returns concentrated in newer code suggests product-stage and legacy-asset considerations are critical: AI productivity gains may favor greenfield development and accelerate new feature throughput, while leaving legacy maintenance relatively unaffected.
  • Broad sharing by seniority implies mandate effects are not limited to a narrow senior cohort, but evaluators should check for concentration of review/verification burden among experienced engineers.
• Measurement and evaluation advice
  • Firms and researchers should combine tool-telemetry with repo data and decompose delivery pipelines (coding vs. pickup vs. review) to detect where bottlenecks or backlogs form.
  • Longitudinal designs and within-worker comparisons (staggered DiD) are necessary to separate adoption/experience effects from calendar shocks and hiring/seasonality.
• External validity and policy caution
  • The studied firm was unusually permissive and resourced; generalizing to firms with seat/token limits, stricter governance, or weaker AI culture is risky. Policymakers and researchers should treat observed doublings as an upper bound achievable under favorable complementarity and governance.
• Research agendas
  • Better causal identification (randomized rollout of tools or training) could quantify the marginal value of enablement and training.
  • Longer-horizon studies are needed to assess maintenance costs, code quality over time, and whether review burdens produce attrition among reviewers or induce process redesigns that affect innovation capacity.

Summary takeaway: In a near-ideal enterprise environment, a documented AI mandate catalyzed adoption that, with accumulated use, produced very large measured throughput gains and reorganized review toward automation. The core economic mechanism appears to be complementary human adoption and experience—so mandates without enablement and review-capacity planning are unlikely to deliver comparable outcomes or sustainable benefits.