State of the Art of DORA Metrics & AI Integration • Nathen Harvey & Charles Humble

The Evolution of DORA and its Metrics

The DevOps Research and Assessment (DORA) program has been analyzing software delivery performance for over a decade, collecting data from thousands of practitioners globally. The research identifies the key metrics, capabilities, and conditions that enable high-performing teams to deliver software with both speed and stability.

While many are familiar with the original four DORA metrics, the model has evolved. There are now five key metrics grouped into two categories:

• Throughput:

  • Lead Time for Changes: The time from code commit to running in production.
  • Deployment Frequency: How often an organization successfully releases to production.
  • Failed Deployment Recovery Time: How long it takes to restore service after a deployment failure.

• Stability:

  • Change Failure Rate: The percentage of deployments causing a failure in production.
  • Deployment Rework Rate: The percentage of deployments that require unplanned work or remediation.

A crucial finding from years of DORA research is that throughput and stability are not trade-offs; high-performing teams excel at both, moving quickly and reliably. These metrics serve as lagging indicators of a team’s underlying capabilities (like continuous integration or documentation quality) and leading indicators of organizational outcomes and team well-being.

The AI Paradox: A Dip in Performance

A headline finding from the 2024 DORA report was that increased AI adoption correlated with a decrease in both software delivery throughput and stability—the very metrics teams have spent years optimizing. The 2025 report shows a slight reversal: throughput is beginning to improve with AI adoption, but stability remains a challenge.

There are a few hypotheses for this initial performance dip:

1. New Tool Disruption: As with any new technology, there’s a learning curve. Teams experience a temporary drop in performance as they adapt their workflows and processes.
2. Optimizing the Wrong Bottleneck: The initial industry focus for AI was on code generation. However, writing code is rarely the primary bottleneck in the software delivery lifecycle. By using AI to generate more code faster, teams are simply overwhelming downstream constraints like code review, testing, and deployment processes. This creates more work-in-progress and amplifies existing friction, potentially harming overall system performance.

“It is rarely the case that writing code is the bottleneck in that process. And yet we’ve we put a lot of emphasis on AI to help us generate or author more code. Your bottlenecks are usually downstream from that.”

AI as an Amplifier

The DORA research suggests that AI acts as an amplifier of a team’s existing systems and processes.

• In organizations with smooth workflows and strong feedback mechanisms, AI accelerates the flow of software.
• In organizations with friction, silos, and bottlenecks, AI highlights and exacerbates those pain points.

The code review process is a prime example. DORA research from 2023 showed that teams with faster code reviews had 50% better software delivery performance. If a team’s code review process is already a bottleneck, using AI to generate more code will only make the problem worse, leading to longer review queues and increased frustration. Conversely, an organization might find more value in applying AI to improve the code review process itself rather than just the code authoring stage.

The 7 DORA AI Capabilities for Success

Before scaling AI adoption, organizations need to have certain foundational capabilities in place. The DORA team identified seven capabilities that, when combined with AI adoption, amplify positive outcomes like reduced friction and better product performance.

1. A Clear and Communicated Stance on AI: A well-defined policy on how, when, and where AI can be used.
2. A Good Data Ecosystem: High-quality, well-connected data is essential, following the “garbage in, garbage out” principle.
3. AI Access to Internal Data: The AI tools must be able to securely access relevant internal data and systems.
4. Strong Version Control Practices: With AI generating more code, frequent check-ins and robust rollback capabilities are critical.
5. Working in Small Batches: Breaking down large changes into smaller, manageable pieces is even more important to leverage AI effectively.
6. A User-Centric Focus: AI should be used to serve the needs of users, not just for the sake of using AI.
7. A Good Internal Platform: A solid internal platform acts as a force multiplier for AI adoption.

The Human and Socio-Cognitive Impact of AI

The research also delved into the human aspects of integrating AI into software development.

Documentation Quality

Surprisingly, the data shows that as AI usage increases, documentation quality improves. This isn’t just about quantity. One hypothesis is that AI helps overcome the “blank page” problem, making it easier for engineers to start writing. Another is that Large Language Models (LLMs) are effective at summarizing existing, and perhaps disparate, documentation, thereby increasing the utility of the information a team already has.

The Trust Paradox

A “trust paradox” exists. In 2025, 30% of respondents had little or no trust in AI’s output, while only 4% trusted it a great deal. Trust in this context is not necessarily about accuracy, but about meeting expectations. If you expect mediocre output and receive it, your trust in the tool’s predictability increases. This highlights the absolute necessity of maintaining strong, fast feedback mechanisms to validate any AI-generated output. You cannot blindly trust the output; it must be tested and verified.

Skill Devaluation vs. Expertise

There is a legitimate concern that AI devalues the skills of experts like technical writers, UX specialists, or SREs. While AI can democratize skills by allowing a novice to perform a task they couldn’t before (e.g., a product manager building a prototype), this does not replace the need for genuine expertise. The work of a novice, even an AI-assisted one, is not the same as the work of an expert. It’s crucial for organizations to continue to value and lean on the deep expertise of specialists.

The Future: The J-Curve and Systemic Thinking

Predicting the future is difficult, but past technology shifts offer clues. We are likely in the dip of a J-curve of transformation, where a new technology initially causes a drop in productivity before it bottoms out and leads to significant improvements. The key is to understand what helps teams get onto that upward trajectory, which is the focus of the DORA AI capabilities model.

The fundamental physics of software delivery—writing, validating, approving, deploying, and monitoring code—haven’t changed yet. However, teams must think systemically. For leaders looking to integrate AI now, a valuable first step is to perform value stream mapping to understand the current system, identify friction points, and pinpoint where AI could provide the most leverage.

• Tags:
• Developer Productivity
• Feature
• AI Adoption
• Software Delivery
• Engineering Culture
• DORA