How AI is changing software development

For decades, the image of someone who builds software was that of a person typing code. That image is becoming incomplete. In just a few years, artificial intelligence tools went from autocomplete suggestions to agents that open change requests on their own.

This article looks at the process, not the individual task. The question is not “how much faster someone writes a function”, a topic we cover in AI in programming: how much work it automates. The question is what changes in the production cycle: planning, writing, review, testing, security, delivery and careers.

The data comes from surveys of professionals, company reports and academic studies, all dated. Opinion surveys measure perception, not outcomes. Vendor reports have a commercial interest. We point this out throughout the article.

From experiment to routine

Adoption is no longer the news. Some figures, always with the date:

• Stack Overflow, 2025. In the community’s annual survey, 84% of respondents were using or planning to use AI tools, up from 76% the previous year. Among professionals, 50.6% used them daily (Stack Overflow Developer Survey 2025).
• DORA, 2025. The DORA research program, run by Google, surveyed almost 5,000 technology professionals. AI adoption reached 90%, with a median of about two hours of use per day (Google, September 2025).
• Stack Overflow, 2026. The 2026 edition, with 30,903 responses from 169 countries, records that 80% of respondents use AI at least one hour a day (Stack Overflow Developer Survey 2026), and that 73% use coding assistants and agents daily (AI section, 2026).
• GitHub, 2025. In the Octoverse report, covering September 2024 to August 2025, GitHub states that nearly 80% of new developers on the platform use Copilot in their first week (Octoverse 2025).

The samples are different and self-selected: those who answer a Stack Overflow or DORA survey do not represent all programmers. Even so, the direction is the same in all of them.

From autocomplete to agents

The first wave was autocomplete: the tool suggests the next line and the person accepts it or not. In June 2024, Google reported that about 50% of its engineers’ code characters were already being completed with AI assistance, with an acceptance rate of 37% of suggestions (Google Research, June 2024). In the same post, the company said it expected the next wave of benefits in testing, code understanding and maintenance, with agents carrying out larger tasks.

The second wave is coding agents: systems that receive a goal, read the project, edit files, run commands and propose a change ready for review (see what AI agents are). GitHub reports that its coding agent created more than 1 million pull requests (requests to merge code) between May and September 2025. The company calls the number “an early glimpse”.

The pace of this transition shows up in the surveys. In 2025, only 14.1% of Stack Overflow respondents used agents daily at work, and 37.9% had no plans to use them. In 2026, among the tools cited as coding agents, Claude Code (66%) and GitHub Copilot (59%) led. We explain how one of them works in Claude Code: how it works.

In the editorial team’s assessment, the practical consequence is a change of role. Part of the work stops being writing and becomes specifying what you want, giving the agent context and judging what it delivered. GitHub itself describes developers who “orchestrate agents”.

The bottleneck moves

Software is not just writing code. After writing come review, testing, integration and delivery. If writing speeds up and the rest does not, the backlog appears further down the line.

DORA has been tracking this for two years. In the 2024 report, greater AI adoption was associated with improvements in documentation quality, code quality and review speed, but also with an estimated 1.5% drop in delivery throughput and 7.2% drop in stability (DORA 2024). The authors concluded that AI “does not appear to be a panacea”.

In 2025 the picture changed in part. DORA found a positive relationship between AI adoption and delivery throughput, but the relationship with stability remained negative. The authors’ explanation: AI speeds up development, and without robust control systems, such as strong automated testing, mature version control and fast feedback loops, the increase in the volume of changes leads to instability (DORA 2025).

Telemetry data from a vendor points the same way. Faros AI, which sells an engineering metrics platform, analyzed two years of data from about 22,000 developers across 4,000 teams, in a report published in the second quarter of 2026. Comparing teams with high and low AI adoption, it found more tasks completed, but also median review time five times longer, pull requests 51% larger and 28% more defects per pull request (Faros AI, 2026). These are correlations, not causes, and the page itself states that the detailed limitations are not public.

Octoverse records a similar symptom in GitHub’s workflow: merges still stall when approval depends on people or policies. In short, review has become a central part of the work, and reviewing code you did not write is harder.

Trust: heavy use, little trust

The surveys show a stable pattern: use grows faster than trust.

In the 2025 Stack Overflow survey, 45.7% distrusted the accuracy of AI tools and 32.7% trusted it, with only 3.1% highly trusting it. The main frustration, cited by 66%, was solutions that are “almost right, but not quite”. And 45.2% said debugging AI-generated code takes more time. When faced with a doubtful answer, 75.3% would still turn to another person.

In 2026, the survey shows 48% trusting AI when they can easily validate the answer, and only 16% extending that trust to most tasks that do not involve important decisions. In DORA 2025, 30% reported little or no trust in AI-generated code.

This has a direct effect on the process. If the output needs to be verified, the capacity to verify (testing, review, knowledge of the system) becomes what limits the team’s speed. We cover verification in how to evaluate AI responses.

Security: the code compiles, but is it secure?

Concern about security has accompanied assistants from the start.

• In 2021, researchers asked Copilot to complete code in 89 scenarios tied to known vulnerabilities. Of the 1,689 programs generated, about 40% were vulnerable (Pearce et al., 2021).
• A user study published at the CCS conference in 2023 found that participants with access to an assistant based on the codex-davinci-002 model wrote significantly less secure code and, at the same time, were more likely to believe their code was secure (Perry et al., 2023).
• In March 2026, Veracode, an application security company, reported evaluating more than 150 models on 80 coding tasks in Java, JavaScript, C# and Python, without specific security instructions. Only 55% of generations produced secure code. According to the company, the rate of syntactically correct code rose from about 50% to 95% since 2023, while the security rate remained practically flat (Veracode, March 2026). Veracode sells security analysis tools, which should be taken into account.

The studies use models from different periods, and the older ones do not describe current tools. The recurring point is another one: improvement in “working” does not guarantee improvement in “being secure”. In practice, this requires automated security analysis and human review to remain in the workflow, especially when agents change many files at once.

Roles and careers

The most sensitive change concerns who enters the profession.

A Stanford study, using ADP payroll data from the United States, found that employment of software developers aged 22 to 25 fell almost 20% between the late-2022 peak and September 2025, while employment of older age groups continued to grow. For the set of occupations most exposed to AI, the relative decline for workers aged 22 to 25 was 16%, after controlling for firm-level shocks (Brynjolfsson, Chandar and Chen, November 2025).

The authors are careful. In a February 2026 note, they state that interest rates do not appear to explain the decline concentrated among entry-level workers, but that AI is not the only factor and that causality cannot be firmly established by this analysis (Stanford Digital Economy Lab, February 2026).

There is a tension here. Experiments cited in our article on AI in programming show that less experienced developers tend to gain more from assistants. At the same time, entry-level positions appear to be shrinking. If the simple tasks that used to train beginners move to agents, teams need to decide how to train the people who will review those agents’ work in the future.

Among Stack Overflow respondents in 2026, 17.1% said they avoid AI out of concern about losing skills or being replaced.

What sets apart those who benefit

DORA 2025 sums up its main conclusion in one idea: AI works as an amplifier, magnifying what the organization already does well or badly. In December 2025, the program detailed seven capabilities associated with a more positive impact of AI (DORA, December 2025):

1. A clear and communicated stance on AI use.
2. Healthy data ecosystems.
3. AI-accessible internal data.
4. Strong version control practices.
5. Working in small batches.
6. User-centric focus.
7. Quality internal platforms.

Almost nothing on this list is about the tool itself. These are engineering and organizational practices that were already recommended before AI. To measure whether adoption is working on your team, see productivity with AI. For the current limits of agents, see autonomous agents: limits.

Key takeaways

• AI use in software development is near-universal in 2025 and 2026 surveys, but trust remains limited and conditional on verification.
• Work is shifting from writing to specifying, orchestrating agents and reviewing. Review, testing and delivery stability have become the new bottleneck.
• Security has been a documented risk since 2021: code that compiles is not necessarily secure, and people using AI may overestimate the security of their own code.
• Employment of entry-level developers has fallen in the United States since 2022, according to payroll data, but the authors do not claim AI is the sole cause.
• Teams with good engineering practices tend to make better use of AI. It amplifies the process that already exists, for better and for worse.