The DORA AI Capabilities Model report, version 2025.1, is published by Google’s DORA research programme as a practical companion to its 2025 State of AI-assisted Software Development report, which first introduced the model. It presents the DORA AI Capabilities Model: seven capabilities, spanning technical and cultural domains, that DORA’s 2025 research found to amplify the positive impact of AI adoption on outcomes such as individual effectiveness, team and product performance, code quality, throughput and organizational performance. The report describes the model as complementary to the long-standing DORA Core Model, not a replacement for it.
The findings draw on research DORA conducted in 2025 comprising more than 100 hours of qualitative data and survey responses from nearly 5,000 technology professionals. For each capability the report explains why it matters for AI, how to improve it, common obstacles, and how to measure it — listing the DORA survey questions used and additional metrics. It closes with tools for acting on the model: team profiles from a cluster analysis, value stream mapping, and a 90-minute team prioritization workshop. The report is licensed under CC BY-NC-SA 4.0.
Findings
- The report’s central claim is that AI is an amplifier: it magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones, so the greatest returns come from investing in foundational systems rather than from the tools themselves.
- 90% of survey respondents use AI as part of their work.
- Each capability amplifies AI’s effect on particular outcomes: a clear and communicated AI stance on individual effectiveness, organizational performance and throughput, while also reducing friction; healthy data ecosystems on organizational performance; AI-accessible internal data on individual effectiveness and code quality; strong version control practices on individual effectiveness (through commit frequency) and team performance (through reliance on rollback); working in small batches on product performance, while also reducing friction; user-centric focus on team performance; and quality internal platforms on organizational performance.
- Working in small batches slightly reduces the individual-effectiveness gains teams perceive from AI. The report offers three hypotheses (the overhead of decomposing work, the friction of reviewing small chunks of machine-generated code, and tools better suited to large changes) and argues the product and friction benefits outweigh this.
- Adopting AI-assisted development tools can harm teams that lack a user-centric focus: with low user-centricity, AI adoption is associated with decreased team performance.
- When internal platform quality is low, AI adoption’s effect on organizational performance is negligible; when it is high, the effect becomes strong and positive. 90% of organizations have adopted internal platforms and 76% have dedicated platform teams.
- About 21% of respondents are beginning to store AI prompts in version control.
- AI adoption alone has only a modest impact on organizational performance, and that effect is amplified when paired with strong value stream mapping practices.
- A cluster analysis of eight outcome factors yields seven team archetypes, from “foundational challenges” (10% of respondents) and “the legacy bottleneck” (11%) to “pragmatic performers” and “harmonious high-achievers” (20% each); the report notes the names and descriptions are an interpretation of the data.
Guidance
For a clear AI stance, the report recommends a cross-functional working group producing a risk-based policy that sorts uses into three buckets — prohibited, permitted with guardrails (for example human-in-the-loop review of all AI-generated code), and allowed — published as a living document. For AI-accessible internal data it contrasts prompt engineering with Context engineering, which it describes as the system that automatically gathers relevant company data, up-to-date documentation, and tools and rules for the model, and it names retrieval-augmented generation and a Model Context Protocol server as the two emerging patterns for supplying only the most relevant context. Under version control it recommends versioning everything, trunk-based development, small frequent commits, and storing AI prompts and agent configuration files such as GEMINI.md or CLAUDE.md. Under user-centric focus it points to Spec-driven development as an emerging paradigm that orients LLMs toward user needs.