DORA AI Capabilities Model (Seven AI Capabilities)
When you need this method
You have rolled out AI tools and see no return that matches the spend. What gets measured is licenses, usage rates, and acceptance, in other words the rollout itself. What does not get measured are the conditions under which usage converts into an outcome at all. Where internal data sits scattered, changes arrive in large batches, and the platform is brittle, added speed merely surfaces existing damage faster. The internal debate then circles around tool selection while the constraint sits one layer below.
Approach
- 1Survey the seven capabilities per team, using the published question blocks rather than your own wording, so answers stay comparable across teams and over time.
- 2Build a profile across all seven capabilities per team and name the two weakest explicitly, instead of collapsing them into a single score.
- 3Apply the amplification claim as a funding rule: while a foundational capability is weak, money goes into that capability first, not into more licenses or more use cases.
- 4Work the tactics the report sets out for each weak capability, meaning make internal documentation and code reachable for models, cut batch sizes, improve the feedback the internal platform gives, or put the AI policy in writing and make it known.
- 5Evidence the suspected constraint before fixing it: map the value stream across the delivery path, then narrow the result down to exactly one first step in a facilitated prioritization session with the whole team.
- 6Repeat the survey after one or two quarters and hold the movement in the profile against throughput, change instability, and product outcomes, rather than celebrating the absolute score.
Typical application
A typical B2B SaaS in the logistics field equips its roughly fifty developers with coding assistants. After two quarters usage is high and team sentiment is good, but the number of changes that have to be rolled back after release has gone up. Surveying the seven capabilities produces a clear profile: version control and user-centric focus are strong, yet internal data is not reachable for the models, and batches are large because approvals run in one bundle each week. The company halts the planned expansion to two further tools and works on two things instead: internal interface documentation and runbooks are made available to the assistants, and the release cadence moves from weekly to daily. A quarter later the rollback rate is back at its earlier level at higher throughput. The license count has not changed in that time.
Limits and counter-indications
The basis is a single-wave self-report survey, analyzed through associations and a cluster analysis. The report speaks of capabilities proven to amplify the benefits of AI; what is methodologically established are associations in a cross-section, not a causal chain. The scope is software development. For sales, marketing, or administration, transferring the model is an analogy and not a finding. The publisher sells platform and AI products, which is worth keeping in mind for the recommendations on platforms and data access. The questionnaire measures perception, and perception about AI is demonstrably unreliable: in a controlled METR study, experienced developers worked more slowly with AI tools while believing they had been faster. Four of the seven capabilities come from the older DORA core model, so teams already running those mainly gain the two data capabilities here. Finally, the model states no thresholds: no published value marks a capability as sufficient, so that judgment stays with leadership.
How to measure impact
Keep a profile across all seven capabilities per team, collected with the published items, and place two delivery outcomes beside it, namely throughput and change instability. What you evaluate is the movement between two waves relative to those two outcomes, not the absolute score of any single wave.
Related methods
Tools for this
Sources
- 1.DORA AI Capabilities Model (v. 2025.1) (opens in a new tab) · DORA / Google Cloud · 2025 · practitioner source · describes the methodCarries the full procedure: the seven capabilities with their wording and rationale, implementation routes and tactics per capability, plus guidance for placing your own team via seven team profiles derived from a cluster analysis, a value stream mapping exercise, and a facilitated ninety-minute prioritization session. Limit: the report calls the capabilities proven amplifiers, yet states itself that it is meant as a frame for forming your own hypotheses and running your own experiments; it rests on a cross-sectional survey without causal identification, and the model is explicitly slated for further validation and revision.
- 2.DORA AI Capabilities Model: Measurement questions (opens in a new tab) · DORA / Google Cloud · 2025 · practitioner source · provides benchmark figuresPublishes the survey items for all seven capabilities, among them eight questions on AI stance, four on the data ecosystem, five on the reachability of internal data, three on batch size, and eleven platform characteristics, mostly as seven-point agreement and frequency scales. This makes the method self-administrable without buying consulting. Limit: these are self-report items with no published thresholds, norms, or scoring rule.
- 3.State of AI-assisted Software Development 2025 (opens in a new tab) · DORA / Google Cloud · 2025 · practitioner source · evidence of effectivenessThe study the model emerged from. It draws on more than one hundred hours of qualitative material and responses from close to five thousand professionals worldwide, and from that states the amplification claim: AI magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones. Limit: the finding comes from a single-wave, non-random self-report survey, and the direction of the association is not identified.
- 4.Becker et al.: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (opens in a new tab) · METR · 2025-07-10 · academic and scholarly literature · limits the methodA randomized controlled trial with sixteen experienced developers on two hundred forty-six real issues in their own repositories. When allowed to use AI tools they took nineteen percent longer, yet judged themselves twenty percent faster. That limits what any instrument measuring AI impact through self-report can claim, this model included. Limit: a very small sample, mature open-source projects only, and participants with little prior experience of the tool used; the authors explicitly decline to generalize to software development at large.
Origin: DORA / Google Cloud