AI-assisted dev productivity 2026: published evidence ranges from 21-28% speedup (GovTech) to 19% slowdown (METR RCT) — honest ceiling is 1.3-1.7× for fluent users

Published research range — three primary sources:

• GovTech Singapore GitHub Copilot study (arXiv:2409.17434): "coding/tasks speed increased by 21-28%".
• Longitudinal arXiv:2509.19708 (100+ developers): "approximately 30-40% of code shipped to production through this tool accounts for overall 28% increase in code shipment volume".
• METR 2025 RCT (arXiv:2507.09089, n=16 experienced OSS developers): "allowing AI actually increases completion time by 19%—AI tooling slowed developers down," despite developers forecasting a 24% reduction.

Confidence: Moderate. The field is moving fast and results are workload-dependent.

Translation:

• Fluent users with skill-library discipline → meaningful (20-30%) lifts.
• Naive power-users on unfamiliar codebases → can be slower than baseline.
• Honest ceiling for a 2-person Candid Creative shop: ship 1.3-1.7× the marketing-site work a 2024-equivalent shop could once the team is fluent.
• Cap is set by client communication, discovery, and design — not coding.

Where AI benefits most (high confidence):

• Greenfield Astro/Tailwind scaffolding.
• Boilerplate refactors (renaming, type-narrowing, API client updates).
• Framework migrations (Tailwind v3→v4, Astro 4→5→6, Next.js 14→15→16).
• Writing tests for existing code.
• One-off scripts and CMS schemas.

Where it benefits least (moderate confidence):

• Performance optimization (requires human judgment about what to measure).
• Accessibility decisions.
• Information architecture and content strategy.
• Anything requiring institutional/client context the model doesn't have.

Push back on hype: "Vibe coding" entire production marketing sites without review produces fragile code. The shops that win in 2026 use AI to accelerate the boring and keep humans on the judgment.