DORA 2025, Part 8. How systemic AI adoption improves productivity, stability, and software quality

Last year, the DORA report showed that teams using AI reported lower throughput and greater delivery instability.

These unexpected results sparked discussion: how can a technology designed to speed up work lead to delays and failures? This year, the picture changed.

AI adoption is now associated with higher throughput, but instability still remains.

This mix of progress and friction pushed us beyond a simple AI use versus no AI use comparison and deeper into what actually drives AI's impact.

Our findings point to a factor that matters more than tools and skills.

AI's impact is determined most of all by the environment in which it operates.

One of the most important outcomes of this year's research was the DORA AI capability model.

AI's impact on metrics such as throughput, code quality, team effectiveness, and organizational performance was consistently strengthened by seven key capabilities:

A clear, publicly stated company position on AI

Availability of internal data for AI

These systemic conditions reflect how an organization structures work, supports teams, and adapts its environment to modern development practices.

They determine whether AI use turns into real results, and make those results scalable. The fact that all these capabilities belong to the team and organizational levels underscores the importance of changing how AI's role in development is approached.

We see that AI's impact on metrics depends on the system in which the work is performed. For example, a mature data ecosystem integrated with AI tools creates conditions where AI's effect can shift from individual productivity gains to organization-wide leaps.

Without such fundamental efforts to prepare the environment for working with AI, the effect may stall, remain uneven, or fail to appear at all.

To use this insight, the focus must shift: less attention to how individual people use AI, and more to how organizations design the systems around them.

To understand what it takes to scale AI's impact, from individual productivity gains to organization-wide benefits, it is important to think systemically.

Organizations are not a collection of separate people and tools, but networks of interdependent elements.

Work flows through teams, processes, rules, infrastructure, and shared norms.

Individual abilities matter, but the final result is determined by how all parts of the system interact with one another.

This principle underlies systems thinking, the approach that has shaped the development of high-performing organizations. U.

W. Edwards Deming, one of the founders of modern quality management, emphasized that most performance problems come not from people, but from systems.

As he said: "A bad system will beat a good person every time." In any system, improving an individual element does not guarantee a better outcome.

Moreover, local improvements can be blocked, weakened, or even backfire if the rest of the system is not ready to adapt.

The same principle underlies the theory of constraints. Every system has a limiting factor, the bottleneck that determines how much value it can produce.

Focusing on something else may create a sense of progress, but it will not improve the value stream in any real way. These ideas have direct implications for AI adoption.

Even if developers write code faster with AI, it still has to go through testing, review, integration, and deployment.

Overall delivery speed will barely change until surrounding processes are adapted to the new tools and the faster pace of work.

The system is not ready to absorb this gain, let alone amplify it.

During cloud migration, companies that simply moved infrastructure without changing architecture and development practices saw minimal benefit.

Real value came from those who reworked applications, teams, and processes for cloud-native approaches.

The same happened with Agile and DevOps: they worked only when they were accompanied by deep changes in roles, feedback, and team boundaries.

New technologies have a transformative effect only when the system around them changes.

That is why AI adoption should be treated as a transformation.

If an organization wants to move faster, experiment more, and change how developers spend their time, it needs to rethink the workflow itself.

Are the underlying systems - integration, testing, deployment, compliance - ready to operate at the new pace?

Are management structures able to make decisions fast enough?

Are teams motivated to delegate tasks to AI, verify results at scale, and share knowledge in new ways?

Without deliberate changes to processes, roles, rules, and expectations, AI will remain a point speedup inside the old system, a missed opportunity.

To scale AI's impact, an organization must invest in redesigning the system: identify constraints, simplify flow, and create conditions where local acceleration becomes a broader organizational momentum.

When developers start using AI tools and feel more productive, but the team does not see a matching increase in throughput or delivery speed, the system itself may be the limiting factor.

Amplification means identifying and removing these friction points so individual speed gains can flow further through the workstream. Code review and handoffs: identify where AI can accelerate and simplify existing stages. For example, an AI-generated first-pass code review helps surface issues faster and reduces time spent on routine comments.

Structuring inputs so that AI highlights risks or briefly summarizes changes makes review faster and more convenient for developers.

CI/CD pipelines: AI code generation is speeding up. Are your delivery processes keeping pace?

CI/CD pipelines may need to be redesigned to reduce waiting between steps and support more frequent releases.

AI-enhanced quality checks can be introduced without extra manual gates, preserving flow and reliability.

The AI capability model highlights the importance of investing in data infrastructure. First, AI becomes more useful for productivity and code quality when AI models are connected to internal data, such as repositories, tracking systems, documentation, and decision-making tools.

This context improves the quality of AI's work. Second, AI has a greater impact on organizational performance when the data ecosystem is healthy: data is accessible to AI, accurate, and complete.

Security and privacy protocols Now that AI is involved in development and operations, security practices must evolve.

This includes safe use of tools, policy updates, and monitoring that accounts for AI work, without creating bottlenecks.

Automating part of these processes helps teams maintain speed.

Change management and cultural alignment

As with any transformation, AI adoption requires vision, support, and clear communication - key signs of transformational leadership.

Leaders should define the long-term goals of AI adoption, whether faster delivery, innovation, or higher quality, and support the transition with training, shared practices, and realistic expectations.

Teams must be able to experiment, make mistakes, build skills, and share experience.

Encouraging practices such as checking AI output, delegation, prompt tuning, and agent orchestration shapes what success means in an AI-accelerated environment.

Beyond enhancing existing systems, AI opens the door to creating new workflows that are built around how AI works from the start and use its potential in new ways. In AI-oriented delivery pipelines, AI can continuously analyze code for bugs, security vulnerabilities, and violations of team standards.

It suggests tests and even generates them automatically.

With the right data and integrations, AI can predict deployment risks and performance regressions before they appear. In AI-oriented data systems, AI can maintain its own environment by organizing, labeling, cleaning, and analyzing data.

This provides deeper insights and speeds up data-driven decision-making. AI also reveals patterns in team work, opening up new opportunities to improve operational processes. AI-focused security expands the capabilities of security teams: it detects threats earlier, identifies anomalous behavior, and automates parts of the incident response process.

For teams that often work with limited resources, AI reduces workload and speeds up response time. AI-oriented collaboration models

Emerging practices such as agentic workflows and swarming are starting to change how people interact with AI.

Agentic workflows distribute tasks across autonomous AI agents, while swarming lets people and AI dynamically come together to solve complex tasks.

Although these approaches are still taking shape, they point to the emergence of new adaptive ways of collaborating.

Emerging approaches such as continuous AI show how AI-oriented workflows can be sustained over the long term.

Continuous AI treats the AI system as a living part of the development pipeline and a full participant in team processes.

Continuous AI tracks project-context events in real time and, while operating autonomously but in alignment with the team, supports collaboration and adjusts the work direction together with it.

The key requirement is that AI systems or agents must continuously receive up-to-date context and be regularly checked for accuracy, usefulness, and cost.

This allows AI to stay aligned with the organization's architecture, practices, and priorities as they change over time, and ensures that AI outputs remain relevant and high quality. Whether extending existing systems or building new ones, intent matters most.

Adopting AI tools on its own does not lead to transformation.

But if AI is combined with pragmatic and systemic changes, it becomes a catalyst that changes how software is built, delivered, and secured.

The sooner an organization treats AI adoption as a transformation rather than the rollout of isolated tools, the more control it will have over how that transformation unfolds.

Technology changes quickly, but the real difference between companies is defined by how they adapt.

If you start before old processes harden around new tools, the organization has a chance to shape the future of its system rather than inherit it by inertia.

The natural first step is to study how work flows today.

In practice, that means creating shared visibility into how ideas move from concept to delivery.

This process, known as value stream management, helps teams visualize every stage of delivery, from coding and code review to testing, deployment, and production release.

When a value stream map is built well, it shows where coordination costs arise, where delays and rework happen, and where the system absorbs or slows the acceleration that AI tools can provide.

This helps identify the factors indicated by

Theory of constraints identifies the points where improvements will have the greatest impact.

To begin mapping the flow of work, companies can form small cross-functional working groups of practitioners who take part in software delivery every day: engineers, product managers, data engineers, operations specialists, and security specialists.

These groups understand the system best from the inside and can identify coordination gaps, bottlenecks, and areas where AI can play a transformative role.

These efforts are especially effective when supported by leadership.

Such support highlights the strategic importance of the work, provides resources, and creates a clear path from research to action.

The mandate of such working groups is to develop strategic recommendations:

Which processes or roles can AI amplify?

Which capabilities should be developed to create long-term value? In some cases, an external facilitator or consultant can help navigate the process, provide benchmarks, and keep focus on systemic opportunities for broader improvements.

For change to stick, insights must emerge from within the system.

When practitioners lead the initiative and leadership commits to providing the necessary conditions, a foundation for real transformation emerges.

Applying systems thinking is critical.

As discussed earlier, organizations are complex systems, and improving one node (for example, speeding up code generation) will not have an effect if adjacent elements do not evolve in parallel.

The DORA AI capability model helps identify which organizational interventions will amplify AI's advantages. For example, a working group may find that although AI tools can generate valuable suggestions, their answers often miss important context: team agreements, architectural history, and past incidents.

For many organizations, this is not surprising - such information is usually stored in fragmented systems and informal channels. In response, the group may recommend structurally and securely surfacing internal documentation, decision records, and historical tickets to AI models.

They can also suggest workflows that automatically tag and retrieve this context during development or review, reducing search time and improving the quality of AI output.

Or a working group can explore how AI can be used to identify outdated documentation, summarize long project discussions, and find mismatches between what the system does and what the documentation says.

This approach helps turn fragmented knowledge into a structured and usable asset.

In addition to process changes, working groups may also identify new skill and role requirements.

As developers delegate more work to AI tools, verification, orchestration, and workflow design become increasingly important.

Organizations will need to define what these roles look like, how to support them, and how to build an incentive system.

This includes targeted training not only on the tools themselves, but also on how AI changes the nature of development.

None of these changes happens on its own or gets implemented all at once.

They require intent, involvement, and ongoing support.

However, they do not need to be perfect at the start.

When organizations start small, act deliberately, and spread responsibility across roles, they create momentum.

Step by step, capability by capability, the transformation takes hold. “Cultural transformation is just as important as tools.

Success requires not only introducing automation and metrics, but also aligning teams around shared goals and shared accountability." TBC BANK

AI can change how software is built, but by itself it does not change the organizational system.

Its effect is different: it quickly and accurately reflects how the system actually works. In well-aligned teams, AI amplifies flow. In fragmented ones, it exposes weak points. Where practices, flexible processes, and shared context are in place, AI's effect appears immediately.

Where success depends on fragile procedures and informal knowledge, the gaps become more visible.

AI is therefore both a mirror and an amplifier.

It highlights what works by accelerating processes already in motion, while also revealing what needs to change.

For organizations willing to look honestly, this reflection becomes a roadmap.

As with other technological shifts, those who see AI adoption as a chance to change the work model itself come out ahead.

It is precisely such organizations that get the greatest benefit, both from the tools and from the transformation AI makes possible.

Written with Matt Beane, Ph.D., Associate Professor

Digital Fellow, University of California, Santa Barbara

Stanford and MIT, CEO and co-founder of SkillBench The same principle applies in software development as in other professions: expertise flows from the top down.

Senior engineers pass architectural thinking to mid-level and junior engineers, while they bring fresh perspectives and new skills.

Pair programming is not just about finding bugs; it is a way to transfer tacit knowledge that is hard to formalize. In an ideal setup, the three-level model junior, mid-level, senior works as a single learning mechanism through joint problem-solving. AI is now interrupting this familiar cycle.

My years of research on intelligent automation show that in many fields, widespread AI use weakens the traditional apprenticeship model and reduces the number of situations where a newcomer can do the "real work" needed for growth.

Experts can do everything themselves, and they do.

A few junior engineers break through because of persistence, but most simply do not get enough participation in the process. Since 2023, I have been studying the use of AI in complex engineering tasks, and the same patterns are already visible in software development.

But there are important exceptions, and we need to understand more deeply how AI really changes skill development.

As in past technological revolutions - from the printing press to the internet - we are in an era of rapid change, and we do not know exactly how people will adapt to the new conditions.

Today, companies focus on productivity metrics: the speed of AI adoption, the amount of code generated, and the number of merged pull requests.

But we almost never measure what reflects skill development: for example, greater code expressiveness, solution diversity, and the complexity of implemented patterns.

The best organizations know how to improve productivity and develop skills at the same time. In part of my research, high performance appeared only where employee skill development was a mandatory part of the process.

Measuring and developing both directions at the same time is the path to sustainable results. AI is inevitably becoming part of the future of development, and that includes learning as well.

We can use AI to analyze developer-AI interactions and connect them to Git history, progress, and development outcomes.

This kind of analysis used to be too expensive, but the cost of data labeling APIs is falling rapidly. AI can give us a coordinate system for redesigning work so that skill development is preserved and strengthened, both individually and collectively.

Current AI usage habits deliver record speed gains while also blocking skill development for most developers.

To preserve our ability to innovate, we need to use AI not only to speed up tasks, but also to measure, support, and develop engineering craftsmanship.

AI's impact is determined not only by how teams work, but also by how organizations make decisions.

To change a system, you first need to learn how to see it.

The next chapter explains how metrics help do this: not as a set of numbers, but as a tool that directs focus, shapes team behavior, and enables intentional change management.

We are moving from asking, "How does the system work?" to asking, "How do we measure it in order to improve it?"