AI coding tools: what developers actually use them for, and where they still fall short
Most developers now use AI tools, and half of professionals use them every day. Fewer trust the output than a year ago. That gap is what IT policy has to manage.
Stack Overflow's 2025 Developer Survey, answered by more than 49,000 developers in 177 countries and published in July 2025, found that 84% use or plan to use AI tools, up from 76% the year before. About half of professional developers use them every day.
Trust moved the other way. 46% said they don't trust the accuracy of AI tool output, up from 31% in 2024. The most common complaint, cited by 66%, was AI answers that are "almost right, but not quite." 45% said debugging AI-generated code takes more time.
For IT leaders, the useful reading is not "AI makes developers faster" or "AI writes bad code." Developers have already sorted the work into tasks where AI helps and tasks where it doesn't, and policy should follow the same lines.
Which tools they use
| Tool | Share of respondents using it |
|---|---|
| ChatGPT | 81.7% |
| GitHub Copilot | 67.9% |
| Google Gemini | 47.4% |
| Claude Code | 40.8% |
| Microsoft Copilot | 31.3% |
Most developers use more than one. That matters for procurement: if you approve a single tool, expect people to keep using others unless you give them a reason not to.
What they use them for
The survey asked which tasks developers mostly use AI for today. The top answers were searching for answers (54.1%), generating content or synthetic data (35.8%), learning new concepts (33.1%) and documenting code (30.8%). Writing code itself came lower, at 16.9%.
The tasks developers said they don't plan to hand to AI are telling: deployment and monitoring (75.8%), project planning (69.2%) and committing and reviewing code (58.7%). Those are the steps where a mistake reaches production.
AI agents
Agents, tools that take a series of actions on their own rather than answering one prompt, are earlier in adoption. 31% of respondents use them now, 17% plan to and 38% have no plans. Of those who have used agents at work, 69% said they had seen a productivity gain.
What this means for policy
The pattern in the data supports a simple rule: let AI draft, but keep a person accountable for anything that ships. In practice that means:
- Approve one or two tools with enterprise terms that keep your code out of model training.
- Keep code review and deployment approval human, which matches what developers already prefer.
- Run security scanning on every change, whoever or whatever wrote it.
- Give agents their own credentials with narrow permissions, so you can see and limit what they do.
What IT teams should do
- Find out which AI coding tools your developers already use, including personal accounts.
- Approve at least one tool on enterprise terms, so there is a sanctioned option.
- Keep a human approval step before code merges and before production deployments.
- Set separate, limited credentials for any AI agent that can run commands or open pull requests.