How AI became standard in software development over three years, what effects it delivered, and why some metrics remain stable.
3.12.2025 This post is an adapted translation of the “2025 State of AI-Assisted Software Development” report by Google Cloud and DORA. The original is available via the link. Reading time: 20 min.
Derek DeBellis - quantitative UX researcher, Google Cloud
In just three years, AI use in development has undergone a radical shift.
If in 2022 a developer using AI seemed unusual, today 90% of technical professionals use AI at work.
The “AI Adoption and Use” chapter shows that AI has become nearly universal, and this trend is confirmed well beyond DORA’s data.
According to the Stack Overflow Developer Survey 2025, 84% of developers already use or plan to use AI in the development process, up from 76% the year before.
Everyday use has also become normal: 47% of respondents use AI daily.
This trend is also confirmed by Atlassian State of DevEx 2025: 99% of developers save time thanks to AI tools.
The growth in individual use is also reflected in corporate strategies.
According to the LinkedIn Corporate Communications report (2025), 88% of business leaders consider accelerating AI adoption a priority.
A McKinsey study shows that 78% of organizations regularly use AI in at least one business function.
Priorities also show up in budgets: AI Index 2025 (Stanford HAI) reports that in 2024 global corporate investment in AI reached $252.3 billion, 26% higher than a year earlier. But the labor market may reflect the new reality most clearly: the number of U.S. job postings mentioning generative AI skills grew by 323% in 2023 alone.
With interest in AI growing rapidly, it is important to ask: do the benefits keep pace with adoption?
Widespread AI adoption does not automatically mean that value becomes widespread as well.
It is important to keep in mind that adoption is often chaotic: it is frequently driven by hype and fear of missing out (FOMO), rather than a well-thought-out strategy.
In addition, adoption can be constrained and slowed by internal systems and organizational structure.
This was also noted in the DORA Report 2024, which showed that AI delivers promising results but at the same time increases software delivery instability and reduces development process throughput.⁷
The same study found a 1.5% decrease in software delivery speed and a 7.2% increase in software delivery instability for every 25% increase in AI use.
are not unique; similar results appear in other studies as well. For example, a recent Model Evaluation & Threat Research (METR) study showed a sharp gap between developers' perceptions and reality: developers whose work slowed by 19% because of AI tools were convinced those tools had made them 20% more productive.
This underscores how subjective assessments of AI effectiveness can be.
Other independent studies show similar results: they point to a possible impact of AI on developers' thinking, focus, and overall well-being.
In other words, the industry is still only beginning to understand how AI actually affects professionals and workflows.
Nevertheless, when developers were asked to assess AI's impact across each aspect of the SPACE model, they reported mostly positive effects and only a small number of negative ones. Most participants in our AI adoption and use chapter also noted that AI improved their code quality and increased their individual productivity.
These contradictory signals show that more data-driven research is needed to truly understand AI's impact on product development, especially given the scale of investment and the speed of adoption.
We believe the developer community and employers should form realistic expectations. The first step is to gain a clear, objective understanding of how AI actually affects the development process.
Only then can expectations be managed responsibly and informed decisions about AI adoption be made.
To understand how key development outcomes depend on the level of AI use, we need to measure AI adoption accurately. This year, we developed a metric based on a few simple principles:
The metric should not favor any single role or function. For example, a developer should not automatically score higher just because they use AI to write code.
The metric should reflect an overall orientation toward using AI, regardless of role or task type, only real, meaningful AI use.
As every year, we let the data lead the way, even when we have strong hypotheses.
The process includes exploratory factor analysis¹⁴, a method that depends less on prior expectations and makes it possible to identify real structures in the data.
The approach to defining and measuring AI adoption should align with recognized methods for studying AI and be consistent with our qualitative observations.
The result was a factor made up of three closely related variables:
How much did you rely on AI at work over the past three months?
How much did you trust the quality of AI-generated results in your workflow over the past three months?
How often in the past three months did you automatically turn to AI when a work task or problem came up? Elements that form the AI adoption factor
The analysis showed that responses to these three survey questions follow very similar patterns.
This suggests there is a single hidden characteristic that causes these variables to change in sync.
We believe this factor reflects three interconnected conceptual dimensions: behavioral, how often a person uses AI; operational dependence, how deeply AI is embedded in workflows; and attitudes and trust, the person's view of AI quality and how much they trust it.
This combination fully aligns with both the research literature and our qualitative observations.
Most likely, this relationship is driven by a feedback loop.
Trust is a prerequisite for use, but use itself becomes the mechanism that builds trust.
This creates a strong cycle: when users begin to trust the system enough to use it, the use itself increases dependence on it and raises trust even further, forming an adoption spiral.
This cyclical nature is what makes combining the variables into a single factor logical, especially since the survey captures only a snapshot in time rather than the full dynamic process.
AI adoption is a psychological process involving beliefs, intentions, and actions.
With a valid AI adoption metric, we can determine whether key outcomes change depending on how actively a person uses AI.
The basic logic of the analysis is as follows:
If you compare two people with the same characteristics, conditions, and processes, the person with a higher level of AI adoption will on average score {number} higher or lower on the {result} measure. Of course, almost no person, team, or organization is truly “average,” but analyzing average effects helps reveal common patterns.
These patterns provide context for a deeper analysis in the DORA AI Capabilities Model chapter, where we examine the conditions under which AI adoption delivers the greatest, or least, benefit.
Each year, we choose and shape a set of outcomes that reflect the real goals and interests of the professionals reading this report.
In other words, we assess practices based on what truly matters to technology professionals.
These are the outcomes we examined this year:
A high-level indicator of overall organizational success, including factors such as profitability, market share, and customer satisfaction.
Team effectiveness A metric reflecting the team's perceived performance and its ability to interact and collaborate effectively.
Product effectiveness A metric that assesses the success and quality of the product or service the team is working on.
It considers characteristics such as the product's ability to help users complete important tasks, data security, and performance metrics such as latency.
Software delivery throughput A metric that reflects the speed and efficiency of the software delivery process.
This aspect is described in more detail in the second part of the study.
Software delivery instability A measure of the quality and reliability of the software delivery process.
This aspect is described in more detail in the second part of the study.
Reflects a developer's personal assessment of the quality of the code underlying the main application or service they work on.
Individual effectiveness Measures how productive people feel in their work and how highly they rate their own achievements.
Work value Measures the share of time a person estimates they spend on tasks they consider useful, meaningful, and worth the effort.
Friction / barriers Measures how strongly different obstacles interfere with work.
The lower the level of friction, the better.
Measures the level of work-related fatigue, exhaustion, and cynicism.
Lower values are considered a positive result.
The chart below shows how AI adoption relates to key work outcomes, from code quality and efficiency to team performance and delivery stability.
We estimate that, all else being equal, with the same conditions, environment, and processes, a person with a higher level of AI adoption will show: higher individual effectiveness; higher delivery instability; higher overall organizational performance; a larger share of time spent on truly valuable work; higher code quality; better product outcomes; higher delivery speed and efficiency;
higher teamwork scores; comparable burnout levels; comparable friction levels.
The remaining sections of the chapter will help explain why this is the case.
We build hypotheses based on research, interviews, expert opinions, and the experience of the DORA community.
Since last year, several metrics have continued to show a stable positive association with AI adoption level. These are: higher individual effectiveness; higher code quality; stronger teamwork; higher organizational performance.
These effects can now be considered baseline: they have repeated for a second year in a row and align with what many practitioners observe.
The reasons can vary greatly depending on the company, team, and context. AI helps developers by automating routine actions, quickly suggesting solution options, generating contextual prompts, summarizing large volumes of information, and handling higher-level tasks from design to analysis.
These targeted improvements add up and reinforce one another, giving teams and organizations a tangible gain.
We'll reply within 30 minutes and send relevant cases, diagrams, or analyses tailored to your context.
Some metrics show stable but not especially encouraging patterns for the second year in a row: no connection with job frustration levels; no connection with burnout levels; and greater AI use is linked to increased delivery instability.
These effects are tied not to AI itself, but to the system in which it operates.
Today, AI mainly helps at the keyboard: it speeds up coding, improves quality, and increases individual effectiveness.
But frustration, burnout, and delivery failures depend on processes, communication, team rituals, and decision-making methods, all of which go beyond a developer's individual work.
These metrics are properties of the sociotechnical system. AI by itself does not change culture and processes. So despite higher productivity, frustration does not fall, burnout does not disappear, and delivery instability increases until the system itself is adapted to work with AI.
Although a tool that removes routine tasks may seem to automatically reduce friction at work, the data show that friction is much more complex than simply automating repetitive actions.
A number of studies indicate that a significant share of friction comes from processes that lie outside an individual developer's control.
A 2019 Microsoft study identified the main sources of friction in the workday: • unstable and slow infrastructure; • outdated documentation; • excessive administrative burden; • lack of time; • repetitive tasks.
One conclusion of the study is that leaders should first and foremost improve processes and tools.
Even if AI reduces friction at the individual-work level (for example, by speeding up code writing), inefficient processes can fully cancel out that effect, especially if the system is not ready for more changes and new ways of working. For example, increasing the volume of changes without updated rules, roles, and “golden paths” raises verification and coordination costs.
At the same time, friction does not disappear entirely, it shifts.
Previously, it showed up as manual routine; now it appears as the need to make decisions, verify results, refine prompts, and work through code that looks almost right. As a result, the overall level of friction may not decrease, even if AI helps with more valuable work and automates some repetitive tasks. "We realized that AI works best when it amplifies the skills of strong engineers.
By automating routine and repetitive tasks, AI frees developers for strategic work and real innovation." Exabeam
It may seem that a tool that increases productivity should reduce burnout.
However, the results show that burnout is barely affected by technology.
It is influenced most strongly by the work culture around the person.
We see this in our own data across different years as well: burnout is closely tied to leadership quality, priority stability, and cultural maturity.
A 2017 meta-analysis identifies stable factors that lead to burnout: weak support from colleagues and managers, a sense of unfairness, low rewards, and job insecurity.
Even if AI could reduce burnout, the effect can easily be absorbed by cultural pressure.
There are also additional signals from our qualitative research.
They align with the literature on labor intensification. The idea is this: the productivity gains from AI have, in some companies, led to higher expectations for employees.
In other words, AI increases individual effectiveness, but the balance between demands and available resources stays the same, so burnout does not go away. "Software development definitely changed because of AI, and I really felt it this year. Not last year. I think that with the arrival of MCP servers (Model Context Protocol) and the ability to write code together with the model, a lot changed: feature release, deadlines, and the amount of work that can be completed in a given period.
Stakeholders now expect more to be delivered in the product, faster.
Deadlines and projects are under tighter time pressure, and that definitely changes how I work. And that is a bit worrying: leadership and product management were given a pretty hard deadline for shipping the product."
After 11 years of DORA research, one thing is clear: an organization’s technical practices and processes are directly linked to software delivery performance.
If these basic capabilities are missing, they completely erase the benefit of AI adoption. For example, a team may use AI extensively but still face high instability if it lacks a reliable delivery pipeline and depends on other teams.
Our data shows that AI does not fix instability. Today, its adoption is even associated with higher instability. Perhaps classic technical practices have become even more important and require stricter discipline.
But it may turn out that this is no longer enough.
Everything points to a more radical conclusion: existing technical practices and metrics must evolve, be supplemented, or be replaced to fit the AI era.
Some argue that increased instability is an acceptable price for the development speed AI makes possible.
The logic is that if delivery speeds up enough for mistakes to be fixed almost immediately, the negative effect of instability can be softened.
However, the data do not support this hypothesis.
To test it, we examined whether a high level of AI adoption reduces the harm caused by instability in outcomes that have historically suffered from it.
We found no evidence of such an effect. On the contrary, instability still leads to worse key outcomes, such as product quality and burnout levels.
These consequences can completely cancel out any gains in delivery speed.
We observed several notable shifts compared with the 2024 results: AI's relationship with the share of valuable work time moved from negative to positive; AI's effect on delivery speed became positive, after previously being negative; AI's effect on product quality moved from neutral to positive.
These changes show that people, teams, and tools have adapted. Developers had an extra year to learn AI, organizations to redesign processes, and vendors to improve models and the user experience.
In many tests, the quality of AI solutions has improved.
Fine-tuning models on your own data no longer requires deep ML expertise. Platforms have made the process accessible.
Cloud providers offered secure ways to connect internal data, such as customer databases, documentation, and repositories, to the fine-tuning process without sending it to external services.
Organizations have started using AI more intentionally. AI is getting new roles and workflows: automated code review, test generation, bug finding, refactoring, and documentation.
Such feedback loops reduce risk and help tools work more reliably.
Developers better understand where AI adds value.
Routine and repetitive tasks are handed to AI, while people focus on analysis, design, and problem solving.
This explains why the share of valuable work time increased and delivery performance sped up.
When AI takes over technical rough work, such as templates, helper code, and routine transformations, developers have more time for releases, which increases throughput and improves product performance.
At the same time, organizational systems are becoming more AI-friendly: teams are adjusting processes, building new protocols, and creating conditions in which the benefits of AI extend not only to individual employees, but also to the product and the company as a whole.
We also examine these limitations and growth areas in the AI Capabilities Model and AI Mirror chapters.
Why did some metrics improve while others did not, such as delivery instability?
There is not enough data for a precise explanation. It is likely related to where teams focus their efforts, which systemic constraints are most visible, and how complex individual problems are. Each metric has its own learning curve. “If AI does for me in 30 minutes what would otherwise take two hours or more, that is good. I get free time, and I can do something else, take on another task, and get more done.
It accelerates professional growth: I learn faster and move forward faster."
AI has a positive effect on most key metrics, but there are important exceptions: it does not reduce burnout or friction, and it remains associated with increased software delivery instability.
Comparing 2025 results with last year makes it clear that models, tools, and workflows are adapting alongside people and organizations.
Practitioners have found ways to use AI to free up time for higher-value work, and we are beginning to see how AI adoption helps improve delivery speed and product quality.
The persistence of effects such as rising instability, no progress in reducing friction, and burnout points to something else: it is not only about the tools, but also about how the work system itself is designed.
If an organization does not change its processes and practices, AI capabilities run into existing constraints. In our view, AI’s value is unlocked less by the models themselves and more by changes to the work system into which those models are embedded.
This conclusion is explored further in the next parts.
DORA focuses on the people who build software. The environment in which developers work directly shapes how they experience their work and how they feel about their profession. AI is changing that environment: companies are rethinking priorities, leaders are looking for new opportunities for innovation, and AI is becoming more deeply embedded in workflows.
This is gradually changing both the tasks themselves and the nature of developers' work.
History shows that technological shifts can change our sense of what is normal.
Before Uber, the idea of getting into a stranger’s private car for money seemed impossible.
Developers are at the forefront of a similar AI-related shift, and our task is to capture these changes as they happen. In the study, we focused on six socio-cognitive aspects that reveal how developers see themselves and their work:
(interest in challenging tasks)
Psychological ownership of the product Redistribution of skill importance The effect of AI adoption level on key metrics
Pride is a basic emotion that appears when a person attributes an achievement to their own efforts. Like a runner who feels satisfied after a marathon, knowing how much training went into it. We feel uplifted when we master a difficult task or make progress. Why is it important to measure pride in the context of AI? There are two opposing assumptions.
First: if a person relies too heavily on AI and automates a large part of their work, the space for achievement through their own effort may shrink. Second: AI frees up time, allowing people to work on more meaningful tasks, the kind they can genuinely be proud of. The data support the second. Higher levels of AI use are associated with a stronger sense of genuine pride.
The mechanism is also clear: the higher the AI adoption, the more time people have for tasks they consider valuable. And the larger the share of such work, the higher the level of professional pride. This shows an important effect: when developers hand routine duties over to AI, they gain the most valuable thing, time, and can invest in projects and ideas that truly matter to them.
Meaningful work is a person's desire to spend time on what feels important and worthwhile. Research shows that when people find meaning in their work, their well-being and satisfaction increase. We wanted to understand whether AI adoption affects how meaningful developers consider their work to be. Two scenarios were possible: AI takes on tedious routine tasks and increases the share of truly valuable work, in which case the sense of meaning grows.
When AI interferes with core elements of the profession, the sense of significance declines. The results did not show any effect. Developers did not see their work as either more or less meaningful. We are probably trying to capture such shifts too early, as the transformation is only beginning.
The arrival of AI gave people a simple way to reduce cognitive load. Research shows that student developers sometimes use AI to literally switch their brains off. And while many enjoy working through hard problems, there is a risk that easy access to ready-made answers will reduce the satisfaction of mental effort: even complex problems can now be solved instantly. But the study data showed that AI adoption did not change developers' need for mental effort.
Existential connection is the feeling that there is a bridge between your inner world and another person's experience. Philosopher William James noted that it is precisely the gap between people's inner experiences that makes it so hard for them to understand one another. This feeling reflects our ability to form deep human connections and feel less alone in our views. We studied this factor because the growing use of AI may change the nature of workplace interactions.
AI gives fast, personalized answers, which reduces the need to ask colleagues for help. This speeds up work, but it can reduce live conversations and joint problem-solving. We assumed that the more developers rely on AI, the weaker their ties with colleagues might become. However, the data showed no link between AI adoption level and a sense of existential connection. There are two possible explanations.
Either we are too early in the adoption cycle to see the effect, or AI, even while replacing some communication, is also creating new points of contact by helping people share knowledge, speed up discussions, and broaden the shared base of experience.
Psychological ownership is the feeling that something belongs to you, even if it does not legally.
This feeling can arise both toward tangible objects and intangible things, such as one's own ideas.
Many developers feel a sense of ownership over the code they write, so we asked whether using AI might weaken that feeling.
Our results show, with 78% probability, that AI adoption does not reduce a sense of personal contribution to the outcome.
Simply put, developers do not see code as any less their own, even if it was written with AI.
This supports the interpretation that modern AI tools are seen more as advanced assistants than as autonomous coauthors.
Developers psychologically place AI in the category of tools, like a compiler or a linter.
They use it, but they do not share authorship with it.
Nevertheless, there is a small but noticeable probability (21%) that AI can reduce the sense of ownership.
When a developer writes code entirely by hand, authorship is clear.
But when AI enters the process, the boundaries of authorship become blurred for some: the sense of personal investment and control decreases, and those are what create a feeling of ownership over the task or outcome.
As developers increasingly work alongside AI, discussion is intensifying about how this may change the value of different skills.
We decided to test whether AI adoption affects which skills developers consider more or less important for their work.
We hypothesized that, in the context of AI, skills related to working with people and with AI itself would become more important, while skills related to writing code would decline in importance.
We asked participants to rank eight skills by importance:
Problem-solving skills Prompt engineering
Memorizing programming language syntax
Understanding your team's codebase
As expected, AI adoption has increased the importance of prompt engineering.
Unexpectedly, it also increased the importance of memorizing syntax.
This is an interesting observation: syntax would seem to be the first thing to take a back seat in the age of AI.
Most surprisingly, AI did not affect perceptions of the importance of the other skills.
There are many possible explanations: developers may be in an adaptation phase with new tools, or they may believe their professional expertise will remain indispensable.
Taken together, the results show that AI adoption has not yet had a noticeable effect on how developers perceive their work.
We will continue to monitor possible changes closely. In the meantime, organizations should give developers more room to do the work they consider valuable.
It is useful to create conditions where they can learn to use AI effectively, eliminate routine tasks, and free up time for truly important work.
To avoid reducing the sense of ownership over outcomes, it is important to treat AI as a tool that works for the developer.
Even as technologies become more autonomous, a human must remain in control.
To understand how organizations can unlock the real value of AI, and which technical and cultural mechanisms strengthen or weaken its impact, move on to the fifth part of the study. It introduces DORA's first AI capabilities model.
It shows that AI's effect depends not on the tools themselves, but on the environment in which they operate, and explains which elements of the development system need to be redesigned for AI to deliver sustained results.
Contacts
Leave your current contact details and describe your task. We will come back with clarifying questions and a proposal for the next step.
