Software developers are the first large profession to work alongside AI every day, and how their teams adapt is a preview of what is coming for everyone else. AI coding assistants, distributed teams and open-source collaboration have changed how software gets built. They have also changed what leaders need to manage: not just output, but review quality, security, skills and culture.
The lesson so far is that the tools matter less than the engineering culture around them. Teams with strong review habits, good tests and clear ownership get real gains from AI. Teams without them get more code, faster, with more hidden defects. For business and IT leaders, the question is not whether developers should use AI, but how to set up the practices that make it pay off.
What the evidence says about AI-assisted development
Adoption is close to universal. Stack Overflow’s 2025 Developer Survey found that about 84% of respondents were using or planning to use AI tools in their development process, while more developers said they distrusted the accuracy of AI output than trusted it. Tools such as GitHub Copilot, ChatGPT, Claude, Tabnine and AI-enabled editors are now part of everyday work.
The productivity evidence is more mixed than the marketing. In a controlled experiment published by GitHub researchers, developers using Copilot completed a well-defined programming task about 55% faster than a control group. But in July 2025, the research group METR published a randomized study of experienced open-source developers working on their own large, mature codebases. With AI tools allowed, they took 19% longer to complete tasks, even though they expected to be faster and afterward believed they had been.
Both results can be true. AI helps most on bounded, well-specified tasks, unfamiliar languages and boilerplate. It helps less, and can slow people down, on complex work in large codebases where the developer already knows the system well and must check every suggestion. The practical conclusion: measure outcomes in your own environment instead of assuming a fixed percentage gain.
From code writer to AI orchestrator
As assistants take on more drafting, the developer’s job shifts toward specifying problems precisely, reviewing generated code, designing systems, and guiding agents that can make multi-file changes and run tests. That is a real change in skills. Reading code critically, understanding architecture and writing good tests become more valuable, not less, because they are how a developer catches a confident but wrong suggestion.
It also creates a training problem. Junior developers traditionally learned by writing the routine code that AI now produces. Teams that want a pipeline of future senior engineers need to create deliberate learning paths, such as requiring juniors to explain AI-generated changes in review, pairing them on debugging, and rotating them through design discussions.
Guardrails for AI-generated code
AI assistants can reproduce insecure patterns, invent package names that do not exist, and leak sensitive code or credentials into prompts. None of these risks requires banning the tools. They require the same discipline good teams already apply to human code, applied consistently. A practical policy covers:
- Approved tools and accounts. Provide business-tier AI tools with contractual limits on data retention and training, and prohibit personal accounts for company code.
- Human review for every change. AI-generated code goes through the same pull request review as any other code, and the author must be able to explain it.
- Automated checks in the pipeline. Run static analysis, secret scanning, dependency scanning and tests in CI/CD so problems are caught regardless of who, or what, wrote the code.
- Dependency verification. Confirm that any package an assistant suggests actually exists, is maintained, and comes from the expected publisher before adding it.
- Limits on agent permissions. Coding agents that can run commands should operate in sandboxes with no access to production credentials.
- Outcome metrics. Track lead time, change failure rate, defect escape rate and time to restore service, not lines of code or AI acceptance rates.
Hybrid teams and developer culture
The pandemic made distributed development normal, and many engineering teams have stayed hybrid or remote. Cloud-based development environments, asynchronous code review and CI/CD automation let a developer in Berlin, Austin or Nairobi contribute to the same codebase on the same day. That widens the talent pool and brings more perspectives into product decisions.
Distributed work puts more weight on written communication: clear tickets, design documents, decision records and thorough pull request descriptions. Those same artifacts also make AI assistants more useful, because a well-documented codebase and clearly written requirements give the tools better context. Culture and tooling reinforce each other.
Security has to travel with the developer. Managed devices, single sign-on to source control, protected branches and signed commits matter more when the team is spread across home networks and time zones. For the individual career angle, see You’re Not Competing With AI, You’re Competing With People Using It.
Open source: shared intelligence, shared risk
Open-source communities remain where much innovation happens first, from AI frameworks to security libraries. Supporting open source helps attract talent, speeds up development, and gives companies a voice in the tools they depend on.
It is also a supply chain. In March 2024, a backdoor was discovered in the widely used xz Utils compression library (CVE-2024-3094), planted by a contributor who had spent years building trust with the project. It was caught before it reached most stable Linux releases, largely by chance. The lesson for businesses is to know which open-source components they ship, keep a software bill of materials, monitor for vulnerabilities, and contribute back to the projects they depend on most, whether through code, funding or maintainer time.
Frequently asked questions
Do AI coding assistants really make developers faster?
Sometimes. Controlled studies show large gains on well-defined tasks, while a 2025 study of experienced developers on familiar, complex codebases found a slowdown. Measure delivery outcomes in your own teams before planning around a specific gain.
Is AI-generated code a security risk?
It can be, in the same ways human code can, plus a few new ones such as invented dependencies and prompt leakage. Code review, automated scanning in CI/CD and approved business-tier tools address most of the risk.
How should we measure developer productivity with AI?
Use delivery and quality outcomes such as lead time for changes, change failure rate, defect escape rate and time to restore service. Lines of code and suggestion acceptance rates reward volume rather than value.
Getting AI-assisted development right
Delana Technologies helps engineering teams adopt AI coding tools and agents with the right policies, pipeline controls and metrics, through our AI consulting and agentic AI solutions. For the operations side, see The Rise of the DevOps Engineer. Call 239.414.5126 or contact us.
Sources: Stack Overflow 2025 Developer Survey; Peng et al., “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot” (2023); METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity” (July 2025); NIST National Vulnerability Database entry for CVE-2024-3094.