The value of AI-assisted development is not simply faster coding. It is about where AI can remove routine work, give developers more room to experiment and improve the development process without creating problems later.
Helmes introduced Claude Code across a software delivery team, offering a practical view of where AI genuinely saves time, where human judgement remains essential, and what needs to change in development practices to make AI-assisted workflows work at team level.
Executive summary
We introduced Claude Code across a 14-person delivery team to see where AI coding assistants could make a meaningful difference in day-to-day software development. Developers and QA engineers were encouraged to experiment with the tool and find the ways of working that made sense for their roles.
The results varied across the team, but several clear patterns emerged:
- QA engineers with no prior coding background began building automated tests
- Developers could prototype ideas faster and work more easily with unfamiliar technologies
- Repetitive testing work was reduced
- Ideas and prototypes could be validated much faster
The experience also helped us understand where AI genuinely improves development workflows, where its limitations become apparent, and where engineering judgement remains essential.
The thinking behind AI-assisted development
By mid-2025, AI coding tools had evolved far beyond isolated code suggestions.
Newer AI assistants had become capable of understanding larger codebases, searching across project files, analysing relationships between components, and assisting with implementation planning. For us, this created an opportunity to explore how AI-assisted workflows could streamline time-intensive engineering tasks and make it easier to experiment with new ideas and technologies.
At the same time, we recognised an important limitation. AI coding tools are designed to produce working output quickly, but they do not naturally prioritise simplicity, maintainability or long-term readability.
As one team lead described:
« If you let these coding agents run unchecked, they can produce layers of code that become increasingly difficult for humans to maintain.«
The question was therefore not simply how much AI could do, but how to integrate it into real engineering workflows in a way that remained useful for developers.
Where AI fits into real software development
The rollout was intentionally lightweight and practical. Rather than enforcing a single workflow, developers and QA engineers were encouraged to experiment with the tools in ways that suited their day-to-day work.
Different team members developed their own approaches, but one common workflow emerged among senior developers:
- Discuss the problem with the AI assistant
- Clarify requirements and possible approaches
- Create or review tests before implementation
- Implement the solution
- Review the final code manually
The process resembled working with a junior developer:
« You explain the problem, it comes up with a solution, maybe the solution is lacking, then you nail down the plan to implement.«
The team also introduced safeguards around the use of AI-generated code, including peer reviews, test-driven practices, reusable component patterns, static analysis tools and mandatory approvals before changes could be merged.
How AI-assisted workflows evolved across the team
Claude Code found its way into several different engineering workflows across the team.
- AI-assisted planning and debugging
For many developers, the biggest value was not fully automated coding, but using the tool as a way to think through technical problems more clearly.
Developers used Claude Code to discuss issues, validate assumptions and explore possible implementation approaches. This proved especially useful when debugging more complex problems, where describing the issue step by step often helped identify the root cause faster.
One developer compared the experience to “rubber duck debugging”, where simply talking through a problem helps clarify the solution.
In many cases, developers still implemented the final solution themselves, maintaining a strong understanding of the codebase and control over architectural decisions.
- Faster prototyping in unfamiliar technologies
The team also used AI coding agents to experiment with technologies where internal expertise was limited.
One project involved a small C++ codebase, despite the team not having dedicated C++ specialists. Traditionally, this would likely have required additional expertise or significant time spent getting developers up to speed.
Instead, developers used Claude Code to understand unfamiliar code, prototype solutions and build proof-of-concept implementations before deciding whether further investment made sense.
This lowered the barrier to experimentation and made it easier to test ideas involving unfamiliar technologies.
- QA automation without traditional coding experience
One of the clearest operational improvements appeared in QA workflows.
Before introducing automation, regression testing took place every two weeks and required roughly three to four person-days of repetitive manual work.
To reduce this workload, the team introduced an automated testing framework supported by Claude Code. Technical leads created the initial framework and example tests, then worked directly with QA engineers through screen-sharing sessions to demonstrate how AI coding agents could help generate and extend test cases.
Notably, the QA engineers involved did not come from traditional software development backgrounds. By combining reusable testing patterns with AI-assisted code generation, they were able to contribute directly to test automation without needing to write large amounts of code manually.
The automation suite eventually grew to around 40–50 tests.
The limits of AI-generated code
The rollout also highlighted a clear limitation of AI-assisted development.
AI coding tools are highly effective at generating code quickly, but where human developers instinctively simplify systems when complexity grows too high, AI tools do not. Without guidance, they may continue adding layers of code as long as the output technically works.
This can make systems harder for people to understand and maintain, allowing technical debt to accumulate quickly.
We therefore treated AI assistance as exactly that: assistance. Significant changes still required human review, architectural oversight and testing. AI could accelerate parts of the work, but responsibility for the resulting software remained with the engineering team.
Results in practice: what changed
Regression testing
Test automation
QA workflows
Prototyping
Day-to-day development
Making AI useful in software development
The experience showed that the value of AI coding tools depends less on how much code they can generate and more on where they fit into the development process.
The clearest gains came from reducing repetitive work, supporting technical problem-solving and making experimentation faster. At the same time, faster code generation does not remove the need for engineering judgement. Architecture, maintainability and code quality still require experienced people making deliberate decisions.
For us, the real value of AI-assisted development lies in that balance: using AI where it improves the work, while keeping engineers firmly in control of what gets built.
If you are looking at how AI coding tools could fit into your development work, get in touch. We are happy to share what we have learned and compare approaches.
Get in touch
