This document is a scorecard for your team on the seven capabilities that DORA's data says make AI pay off. Each signal comes with one metric you can take from systems you already have, one question to ask, and a green, amber, red threshold. Score the team this week, then read what usually fails together and what the score cannot tell you.
1. What DORA found in 2025, in one sentence
"AI is an amplifier." That is the central claim of the 2025 report by DORA, Google Cloud's DevOps Research and Assessment program, based on survey responses from nearly 5,000 technology professionals and more than 100 hours of qualitative data: AI magnifies an organization's existing strengths and its existing dysfunctions.
The adoption numbers behind it: 90% of technology professionals report using AI at work, more than 80% perceive that AI increased their productivity, 65% rely on it at least a moderate amount, 60% use it about half the time or more when they encounter a problem, and 59% observe a positive effect on code quality. DORA's advice to technology leaders has three parts: have a systems view, invest in foundational systems (internal platforms, data ecosystems, core engineering disciplines), and focus on effective use to guide, evaluate and validate AI generated work.
The companion model names seven capabilities that, in DORA's data, amplify the positive effect of AI adoption on outcomes such as team performance, code quality, individual effectiveness, product performance, friction, throughput and organizational performance. The seven are the signals below.
2. The seven signals
Signal 1. A clear and communicated AI stance. What DORA found: when the organization's stance is clear and well communicated, it amplifies AI's positive effect on individual effectiveness and organizational performance, and turns AI's neutral effect on friction into a decrease. Without it, DORA writes, developers act too conservatively for fear of crossing an invisible line, or too permissively with unvetted tools. What to measure: whether a written stance exists (which tools, for what work, reviewed how) and what share of the team can state it without looking it up. Question for this week: if a developer asks whether they can point an agent at the production database schema, is the answer written down?
Signal 2. A healthy data ecosystem. What DORA found: internal data that is high quality, accessible and unified is a foundational capability that significantly amplifies AI's positive influence on organizational performance. Applied to fragmented data, DORA says, AI simply scales the confusion. What to measure: how many sources a developer has to consult to answer a routine question about the system, and whether metrics, logs and documentation agree with each other. Question: when two dashboards disagree, which one does the team trust, and why?
Signal 3. AI accessible internal data. What DORA found: giving teams AI tools that can access internal data directly is a statistically significant multiplier for individual effectiveness and code quality. DORA describes this as context engineering, the system that gathers relevant documentation, policies and compliant code before the agent acts, and warns that an agent replicating deprecated patterns amplifies debt. What to measure: whether the repository contains context for the agent (architecture, conventions, how to run tests, what is deprecated) and when it was last updated. Question: does the agent learn the codebase from a file the team maintains, or from whatever it finds?
Signal 4. Strong version control practices. What DORA found: strong version control practices amplify AI's positive effect on individual effectiveness and team performance. Frequent use of rollbacks, specifically, amplifies the positive influence of AI on team performance. Version control is the safety net as AI accelerates the velocity of change. What to measure: rollback frequency and time to revert; share of generated changes that land through the same branch and review path as human changes. Question: how long does it take to undo a change an agent made yesterday?
Signal 5. Working in small batches. What DORA found: working in small batches amplifies AI's positive effect on product performance and helps turn AI's neutral effect on friction into a net positive. What to measure: median pull request size and its trend since the initiative began; share of pull requests above the team's size limit. Question: did pull requests get bigger when the agents arrived, and did anyone decide that?
Signal 6. A user centric focus. What DORA found: teams that focus on the user have 40% higher organizational performance than teams that do not (DORA publishes the figure without the baseline). When such teams adopt AI, their effectiveness and team performance grow; without that focus, DORA writes, AI can propel a team faster in the wrong direction and accelerate the production of low value software. What to measure: whether user metrics (adoption, retention, task completion, satisfaction) are visible next to delivery metrics, and whether engineers observe users directly. Question: of the features shipped faster this quarter, how many does anyone use?
Signal 7. A quality internal platform. What DORA found: when platform quality is high, the effect of AI adoption on organizational performance is strong and positive; when it is low, the effect is negligible. By 2025, 90% of organizations had adopted at least one internal platform, 76% had a dedicated platform team and 29% ran a multi platform environment. What to measure: whether a developer can build, test and deploy a generated change through an automated path without a ticket; platform satisfaction from the team's own survey. Question: is the platform the foundation the agents run on, or the place their output gets stuck?
3. The scorecard
| Signal | Metric | Where the data lives | Green | Amber |
|---|---|---|---|---|
| 1. AI stance | Written stance exists; share of team that can state it | Wiki, onboarding, a five minute poll | Exists, most can state it | Exists, few know it |
| 2. Data ecosystem | Sources consulted per routine question | Ask five developers | One or two | Three or more, or "depends" |
| 3. Agent context | Context file in repo; days since last update (a proxy for whether the context exists, not for its quality) | Repository | Present, updated this month | Present, stale |
| 4. Version control | Time to revert; rollbacks per month | CI/CD, deploy logs | Minutes, used regularly | Hours, rare |
| 5. Small batches | Median PR size and trend | Version control | Flat or down since AI | Up since AI |
| 6. User focus | User metrics on the team dashboard | Analytics, product | Visible, discussed | Not visible |
| 7. Platform | Automated build to deploy path for generated code | Platform, CI/CD | Yes, no ticket | Manual steps |
Red is the absence of the thing: no stance, no context file, no rollback path, no size limit, no user metric, no automated path. Score green as two, amber as one, red as zero. Seven is our threshold, not DORA's; the model describes capabilities and does not score them. A team below seven is missing at least half of the foundation DORA's data says AI amplifies.
4. Reading the scorecard from inside a client's team
On our own work inside client teams, the scorecard rarely fails on one signal. It fails on two together. The most common pair is signals 3 and 5: the agent works without a context file, so it produces changes that touch more than they should, and pull requests grow. The reviewer then absorbs both problems at once. The second most common pair is signals 1 and 4: nobody wrote down what the agent may do, so it does more than expected, and the team discovers that reverting is slow.
The signals that move fastest are 1 and 3. A written stance and a maintained context file are days of work. Signals 6 and 7 are quarters. The scorecard is not a ranking of importance; it is a way to see which quarters of work are ahead.
5. What the scorecard does not tell you
Two limits, stated plainly.
DORA's findings are survey based. Faros AI, on two years of telemetry from 22,000 developers (April 2026), reports that organizations with mature delivery practices and strong DORA metrics (deployment frequency, lead time for changes, change failure rate and time to restore service) show the same downstream deterioration under high AI adoption as everyone else, and says this contradicts DORA's conclusion that foundations protect. The two are not measuring the same thing: DORA asks people; Faros reads systems. A team should hold both. The seven signals tell you whether the foundation exists. Telemetry (review time, incidents per merged change, code churn, unreviewed merges) tells you whether it is holding.
The second limit is that the scorecard measures the organization, not the individual. It will not tell you which developers are faster with an agent. In our experience that question stops being answerable once adoption is real, because the developers stop working without one.
Sources
- DORA (Google Cloud), 2025 State of AI-assisted Software Development, report page and official infographic, September 2025. https://dora.dev/research/2025/dora-report/
- DORA, Capability catalog: Clear and communicated AI stance; Healthy data ecosystems; AI-accessible internal data; Version control; Working in small batches; User-centric focus; Platform engineering. https://dora.dev/capabilities/
- DORA, DORA AI Capabilities Model report, last updated November 25, 2025. https://dora.dev/ai/capabilities-model/
- Faros AI, Ten takeaways from the AI Engineering Report 2026: The Acceleration Whiplash, April 12, 2026. https://www.faros.ai/blog/ai-acceleration-whiplash-takeaways
Analysis and conclusions by Viaro Networks, based on the published reports listed above. DORA material is used under CC BY 4.0.