DORA — 2025 State of AI-assisted Software Development
TL;DR
DORA’s annual survey, this year focused entirely on AI: ~5,000 technology professionals worldwide plus 100+ hours of qualitative data. It is the largest-sample source in this ingest, and the one whose central finding is a split, not a direction.
Adoption is effectively universal. 90% of respondents use AI at work, up from 76% in 2024. 90% of organisations have adopted at least one platform. Trust has not kept pace: 70% report trusting AI-generated code, meaning 30% report little or none — a striking number given near-total adoption. People are using a tool a third of them do not trust.
The central thesis: “AI doesn’t fix a team; it amplifies what’s already there.” Strong teams get more efficient; struggling teams find AI surfaces the problems they already had. This is the report’s answer to why AI-adoption studies produce such scattered effect sizes — the moderator is the delivery system, not the tool.
The throughput/stability split — the finding that matters.
| Relationship with AI adoption | 2024 | 2025 |
|---|---|---|
| Software delivery throughput | negative (~1.5% drop per 25% adoption rise) | positive |
| Software delivery stability | negative (~7.2% drop per 25% adoption rise) | still negative |
“AI adoption does continue to have a negative relationship with software delivery stability.”
The mechanism, in the report’s own words: “AI accelerates software development, but that acceleration can expose weaknesses downstream. Without robust control systems, like strong automated testing, mature version control practices, and fast feedback loops, an increase in change volume leads to instability.”
Read that carefully, because it relocates the problem. Throughput turning positive year-over-year means the authoring constraint has been relieved. Stability staying negative means the absorptive capacity of the delivery system — review, testing, rollback, observability — has not moved with it. That is the same structural claim the practitioner material arrives at independently: Claire Vo builds a risk-scoring reviewer because review, not authoring, is now the bottleneck, and Carson gates 40 daily PRs on a two-loop automated review plus a video he watches himself. DORA’s prescription and their practice are the same prescription.
Seven team profiles. Cluster analysis yields seven archetypes, including “Foundational challenges” (survival mode — low performance, high instability, high burnout and friction) and “Harmonious high achievers” (strong across well-being, product outcomes, and delivery). The profiles are the operational form of the amplifier thesis: the same AI intervention lands differently depending on which cluster you are in.
The DORA AI Capabilities Model names seven capabilities that amplify AI’s positive impact: clarify and socialise AI policies; connect AI to internal context; prioritise foundational practices; fortify your safety nets; invest in your internal platform; focus on end-users; plus one further capability in the model visual. Two framing insights sit behind them: user-centricity is a prerequisite (AI is most useful pointed at a clear problem), and platform engineering is the foundation — internal platform quality correlates directly with the ability to unlock AI value.
Two other sources sit directly underneath these two numbers. For throughput, Cui et al. give a pre-registered causal magnitude — +26.08% completed tasks across 4,867 developers — for the half of the story that turned positive. For stability, Liu et al. show what is physically accumulating in the repositories: 484,366 issues introduced across 302.6k AI-authored commits, 22.7% still present at the latest revision. DORA says the delivery system destabilises; Liu et al. say what is piling up inside it.
Dynamic-capabilities reading
digital-transforming/improving-digital-maturity— the capabilities model is a maturity ladder, and the report’s claim is that position on it, not tool choice, determines the return.digital-transforming/redesigning-internal-structures— the platform-engineering finding is a structural prescription: build the internal platform before expecting AI leverage.contextual/internal-enablers— safety nets (automated testing, mature version control, fast feedback) are named as the conditions under which higher change volume does not become instability.contextual/internal-barriers— the seven profiles, especially “Foundational challenges,” describe the barrier state directly: AI makes an already-unstable delivery system worse.
Linked entities and concepts
- Entities: Google, DORA
- Concepts: ai-coding-productivity-evidence, agentic-pull-requests, enterprise-ai-adoption, agentic-engineering, micro-productivity-trap
- Dangling (single-source mention, deferred): Nathen Harvey, Derek DeBellis
Scope and reliability
The announcement blog post only — the full report (including the complete seven-profile chapter and the seventh AI capability) was not converted. Cross-sectional self-report survey: the adoption, trust and perceived-productivity numbers are beliefs, and METR’s RCT establishes that developer beliefs about AI productivity are systematically and substantially wrong. The delivery metrics (throughput, stability) are behavioural and more trustworthy. Published by Google Cloud, whose commercial interest runs toward both AI adoption and platform engineering; the stability finding cuts against that interest, which is a point in its favour. The 2024 baseline figures (1.5% / 7.2% per 25% adoption) come from the prior year’s report and are carried here as context, not re-verified.