DORA 2025: AI in Software Development and Teams

About the publication

3.12.2025 This post is an adapted translation of the Google Cloud and DORA report “2025 State of AI-Assisted Software Development.” The original is available via the link. Reading time: 10 min. This chapter was prepared with contributions from: Kevin M. Storer, Ph.D. - user experience researcher, Google Cloud Derek DeBellis - quantitative user experience researcher, Google Cloud

AI use and adoption

As we continue to study AI adoption trends in the software development industry, it is becoming clear that AI use in development has grown significantly. In this fast-changing field, it is important to offer data-driven recommendations that help the professional community navigate the changes ahead. The "State of AI-assisted Software Development 2025" report is our deepest and most detailed study of AI use in development to date.

Key insights: adoption and usage experience

Overall, the study shows that software developers are widely adopting AI and relying on it actively for a broad range of tasks.

This leads to clear benefits, both in individual productivity and in code quality. Adoption: 90% of this year's survey respondents say they use AI in their work.

That is 14.1% more than last year.

Such a high level of adoption shows that AI use in software development has effectively become the industry's new normal. "I think over the past year many people have realized that generative AI has finally reached the point where it really works for a huge number of tasks. And now that this has become almost taken for granted, every team has cases where generative AI can be useful.

There are many reasons to adopt such solutions: launch new features, reduce costs, and shorten development time.

Nobody wants to be left behind. And I am sure that over time AI will become more deeply integrated into everything we do. And if you are not using generative AI in at least some form, yes - it will be hard to keep pace.”

Our respondents reported varying levels of experience with AI tools.

The median is 16 months, and the mean is 16.22 months.

For comparison: ChatGPT was released in November 2022, roughly 31 months before our survey began.

As a result, the sample included both early adopters and those who joined later.

The spike in AI adoption is especially noticeable from late 2023 to mid-2024.

According to the survey, respondents spent an average of two hours of their last workday working with AI, or about one quarter of a standard eight-hour day.

Reflective AI use

Although AI use among respondents is nearly universal, reflective use - when a person automatically turns to AI as soon as

Tasks

Among AI users, only 7% say they always use AI when they need to solve a problem or complete a task.
At the same time, 39% turn to AI only occasionally.
Nevertheless, 60% of AI users say they use it about half the time or more when they face a task.
This suggests that AI has already become a regular part of the development workflow.
Over the past 18 months, Sabre has closely tracked the adoption of generative AI assistants through usage analytics and satisfaction surveys.
Adoption has risen to 74% among developers with different experience levels and tenure, with especially notable growth in AI use for core development tasks.
Higher usage is directly linked to greater satisfaction: 86% of users report increased productivity.
The gradual increase in satisfaction and time savings shows that the benefits of generative AI tools grow as users become familiar with them and work with them more confidently.
Analytics also revealed slow adoption of newer capabilities such as agent mode - only 25% of users use it. In response, Sabre is strengthening training programs and expanding knowledge sharing among developers to increase engagement and ensure teams use AI capabilities confidently. -
Jacek Ostrowski, VP of Platform Engineering, Sabre
Only 5% of AI users say they are not dependent on it at all. At the same time, 65% say they rely on AI to a moderate extent (37%), significantly (20%), or to a very great extent (8%).

Tasks: where developers use AI

As in the 2024 DORA report, the leading AI use case among survey participants remains writing new code - 71% of respondents who do development use AI to help create code.

Most professionals whose roles include these tasks also use AI for: literature review - 68%, changing existing code - 66%, proofreading and text review - 66%, creating or editing images - 66%. Less common AI use cases among those whose work includes these tasks are: requirements analysis - 49%, internal communications - 48%, calendar management - 25%

UC Berkeley study: what we examined

As AI assistants become a standard tool in developers' work, a UC Berkeley research team studied how student developers use AI-powered IDEs in practice.

Using eye tracking and interviews, researchers observed how Python developers with one to five years of experience completed two short tasks: working with an unfamiliar library and interpreting a convoluted function. The findings from this study will help developers at any level choose ways of interacting with AI assistants that better fit their working style.

We'll curate materials for your task

clients@kt.team Telegram @kt_team_blog

We'll reply within 30 minutes and send relevant cases, diagrams, or analyses tailored to your context.

Customization as a solution

To reduce frustration and better support focused work, developers and teams can customize the behavior of AI tools.
Most modern IDEs let users: turn built-in suggestions on or off, work in request-only mode, and configure the style and format of generated suggestions
Repository-level configuration files and related documentation help AI assistants follow existing rules and processes.
Experimenting with these settings makes it possible to adapt AI behavior to the cognitive demands of different tasks, reducing distraction and increasing the usefulness of AI assistants.

When AI gets in the way

Although AI coding assistants are designed to save time and reduce workload, a UC Berkeley study found that in some tasks AI can create additional friction. Student developers readily used AI for mechanical tasks such as writing boilerplate code or installing packages. But when deep understanding of logic was required, for example when analyzing a complex code fragment, the same developers almost completely ignored the suggestions.

Eye tracking showed that for interpretive tasks, attention to the AI chat window was under 1%, while for mechanical tasks it was almost 19%. In many cases, students deliberately chose to do the work manually to stay in control and better understand what was happening, even when AI suggested correct and time-saving solutions.

What teams should consider

To use coding assistants as effectively as possible, developers and teams should pay attention to customization.
The study found that AI can hinder tasks that require interpretation by adding cognitive load, especially when a developer is trying to understand unfamiliar code.
AI support should match the type of task and the developer's preferences.
Fine-tuning tools can turn AI from a source of frustration into an assistant that boosts productivity and makes the development process more convenient and satisfying.

AI interaction formats and modes

Most often, developers interact with AI through conversational chatbots.
Second is AI built directly into the IDE.
Respondents say they interact less often with AI embedded in automated pipelines and other development tools.
However, the reason may be that such AI tools are simply less visible to the user, even though they are actively working behind the scenes.
Chat - step-by-step text interaction in a question-and-answer format
Predictive text - automatic code suggestions such as autocomplete
Collaborative mode - using AI to make broader changes to the codebase
Agent mode is when AI works almost autonomously, making changes without constant oversight
Because chatbots and AI built into IDEs are the tools respondents use most often, text chat and predictive text are the most common interaction modes.
Least common is agent mode: 61% of participants said they never interact with AI in this format.
This pattern likely reflects different levels of technological maturity: agent mode appeared relatively recently, while chat and autocomplete have long been familiar and more reliable features.

AI’s impact on productivity and code quality

More than 80% of this year’s survey participants believe that using AI has increased their personal productivity.
Although more than 40% report only a slight increase, fewer than 10% of respondents believe AI has reduced their effectiveness in any way.
Most respondents (59%) say AI has had a positive impact on the quality of their code.
At the same time, 31% consider the improvement minor, and another 30% see neither a positive nor a negative impact.
Only 10% of survey participants report any decline in code quality due to AI use. “Sometimes AI writes code better than I do - at least for certain tasks. I think the reason is that it is very well trained. I have said this before: code is binary, it either works or it does not. And the code AI generates is usually good enough for my tasks.
It often follows standards I might have forgotten or that I am too lazy to review and refactor.
So I think AI makes everything more coherent and polished and polished"

Trust in AI outputs

As in the 2024 DORA report, this year’s data show a complex and multifaceted landscape of trust in AI-generated outputs.
Most survey participants (70%) express some degree of confidence in the quality of AI-generated output.
This includes nearly a quarter of respondents (24%) who say they trust AI "a lot" or "to a great extent."
At the same time, 30% of respondents take a more cautious stance: 23% trust AI "a little," and 7% do not trust the quality of generated code at all.
These data highlight an important point: heavy AI use and clear benefits can coexist with a moderate, deliberate level of trust.
Complete trust is not a requirement for AI outputs to be useful.
During interviews, developers compared this to how they treat Stack Overflow answers: they use them heavily, but never accept them blindly - they always verify them.
Trust in AI for development remains an important topic for research and discussion.
We previously identified five strategies that help increase developers’ trust in AI.
However, this year's data show that developers have already learned to account for AI's limitations and adapt their workflows to minimize risks.

Final thoughts

Taken together, these results show that AI use in software development has become nearly ubiquitous. AI is used across a wide range of tasks, is embedded in respondents’ workflows, and is often the first tool they turn to when a problem arises.

While survey participants still express concern about the reliability of AI-generated code, they also note a significant positive impact from AI on personal productivity and code quality. So, despite some shortcomings, AI use has very quickly become standard practice for most organizations involved in software development. Last year, we found that competitive pressure is one of the key drivers of AI adoption in software development.

Many interview participants described this as a fear of missing out or a concern about falling behind peers and competing companies.

However, the question remains: is social pressure enough reason to adopt a new technology?

Although the data show many positive effects from using AI, we also observed significant drawbacks.

That is why we caution against treating AI’s ubiquity as a signal that every organization should rush to adopt AI regardless of its own needs. Instead, we see these results as strong evidence that everyone working in development - from individual contributors to team and company leaders - should carefully decide whether they need AI, where it will truly be useful, and how it should be used.

How conservative or, by contrast, flexible AI adoption should be depends on the context of the specific organization.

However, the widespread adoption of AI shows that its impact on development processes can no longer be ignored.

We understand that decisions about how much and in what way to integrate AI into development are complex and should be data-driven.

That is why, in the next chapter on DORA’s new AI Capabilities Model, we look at how cultural and technical capabilities within an organization influence AI adoption outcomes.

Our goal is to give companies that have decided to integrate AI into their processes practical guidance so they can do it effectively and realize tangible value.

To see how these findings fit into the broader picture of change, and which system-level mechanisms amplify or weaken AI's effect, move on to the fourth part of the study.

It shows why AI creates value only when the work system itself changes, and what organizations need to redesign to achieve sustainable results.

Contacts

Let's Discuss Your Project

Leave your current contact details and describe your task. We will come back with clarifying questions and a proposal for the next step.

DORA 2025, Part 3. How AI Is Transforming Software Development: Adoption, Use Cases, and Impact on Team Effectiveness