One of the most instructive experiments Codurance ran was a projected 60-day rewrite of its own internal time-booking system, originally built in Clojure. The question was straightforward: how quickly and how safely could AI help migrate it?
The answer was three days. A significant result, made possible only by the discipline applied around it.
The central insight was that you cannot simply hand a codebase to an AI and walk away. It required the same treatment as onboarding a new team member: a clear workflow, well defined standards, and careful review of output. The work was broken into well-scoped vertical slices, with features extracted, requirements defined, and each piece built, reviewed, and validated before progressing. By completion, the system has a CI/CD pipeline, integration tests, and an architecture suited to continued AI-assisted development.
Technology choice mattered too. TypeScript’s type system gave the AI a mechanism to detect and resolve inconsistencies across the stack. A principle that holds broadly: the cleaner and more expressive the stack, the more value AI can extract from it.
We expand on this in our previous Case Study: AI Project Delivers 10x Speed Without Compromising Quality.
Prefer listening over reading? Press play and enjoy
The 10x Engineering Reality: Beyond the AI Hype to Proven ROI
12:42
There is a growing gap between the AI productivity promises circulating in the industry and what engineering teams are actually experiencing. The reality is more nuanced than the headlines suggest.
AI is not a shortcut. AI is an amplifier. And like any amplifier, the quality of what goes in determines the quality of what comes out. Strong engineering practices get stronger. Weak ones get weaker.
This article draws on Codurance’s own research, internal experiments, and client work to provide a grounded view of what it actually takes to extract real value from AI-assisted engineering.
What this article explores:
- How Codurance approaches AI modernisation through structured R&D
- What the rewrite of a real internal system taught us about working with AI
- The particular challenges of legacy codebases
- Practical advice on workflows, guardrails, and accountability
- Key takeaways for technical leaders and engineering teams
How do you build an effective AI engineering strategy across a team?
Codurance began its AI-assisted engineering journey when it became clear that large language models (LLMs) performed significantly better when equipped with the right tooling. In response, an R&D team was assembled with a specific focus: how do you build a structured approach to modernising using AI?
Rather than searching for a single solution, the team developed a toolbox of patterns and practices matched to different modernisation challenges. This included techniques for understanding legacy codebases, BDD-style specifications, and strategies for accelerating rewrites safely. From there, the focus shifted to scaling that knowledge: communities of practice, an AI Champions group, and dedicated training exercises to build capability across the organisation.
Can AI really rewrite a codebase faster?
What a 3-Day Migration Actually Looked Like
How do you use AI safely on a legacy codebase without tests?
Working with a legacy codebase (particularly one without tests, consistent patterns, and craftsmanship principles) presents a fundamentally different challenge. AI learns from what it sees. Inconsistent codebases produce inconsistent output.
Codurance’s experience with a legacy React 16 application illustrates this. The codebase blended genuine React patterns with a significant amount of vanilla JavaScript written as though it were React, creating an unreliable foundation for AI assistance. Before acceleration was possible, the team had to establish clear guardrails: defining what good looks like, identifying reference patterns, and being explicit with the AI about what to follow and what to disregard.
The approach centred on characterisation testing: building a test harness that captured the system’s existing behaviour before anything was changed. Without it, there is no reliable feedback loop, and no way of knowing whether a change has introduced a regression. With that foundation in place, the team could draw clear boundaries: legacy code stays untouched; new work follows defined standards. The acceleration that followed was still substantial, but it demanded more deliberation and senior engineering judgement than a greenfield build.
Why does TDD matter more than ever in AI-assisted development?
The central conclusion is this: AI does not replace the need for strong engineering practices. It makes those practices more important than ever.
Teams with solid foundations like test-driven development (TDD), automated pipelines, clear architectural patterns and fast feedback loops can achieve significant acceleration with AI. Teams without these foundations will find that AI accelerates their problems just as readily as their progress.
This has particular implications for TDD. The AI era has largely settled what was once a philosophical debate: TDD with AI assistance produces demonstrably better systems. Tests become the specification that guides the AI’s work, provide a mechanism for verifying output, and constrain it from unproductive directions. This isn't just our experience. Google’s 2025 DORA Report found that while AI boosts individual productivity, it also increases software delivery instability when not paired with robust automated testing. In other words, the teams getting the most from AI are the ones who already had strong testing discipline.
Any bottleneck in the development process will become more visible under AI-assisted development. Slow requirements, manual testing gates, and heavy approval processes are not new problems, but AI makes their cost more apparent.
Our co-founder Sandro Mancuso writes about the crucial requirement for strong engineering practices surrounding AI in his blog: Software Craftsmanship in the AI Era.
Who is responsible for code written by AI?
When AI writes the code, who owns it? The answer is unchanged: the engineer who commits it.
Teams that treat AI output as authoritative, assuming that if the model produced it, it must be correct, encounter a predictable outcome: untestable code, accumulating technical debt, and systems that even the AI struggles to navigate. The discipline required is real. Engineers must review AI output, understand it, and be willing to push back or discard work that does not meet their standards. You have to review what AI produces. You have to understand it. You have to be willing to push back, redirect, and sometimes discard work that does not meet your standards.
Architecture matters here too. Well-structured systems give AI a better surface to work with. An agent that can run the application, inspect a browser, check a type error, and observe the results of its changes has something close to a productive developer’s feedback loop. An AI working in a fragmented system with unclear boundaries is largely operating blind.
How do you improve AI coding workflows so they get better overtime?
AI-assisted engineering improves with deliberate iteration. A pattern Codurance has found effective:
-
Ask the AI to produce a plan before it begins executing. This surfaces problems early.
-
After task completion, ask the AI what it found difficult. The answers are often worth acting on.
-
Where friction is encountered, encode the solution as a build script or helper tools, so the problem does not reoccur.
-
Inject those solutions into deterministic systems: scripts, tooling, configuration. That way the context is always available without relying on model memory.
The counterintuitive lesson: using less AI in the right places makes AI more effective overall. Deterministic scripts, clear guardrails, and well-structured tooling reduce the burden on the model and increase the reliability of its output.
8 key takeaways for AI engineering
-
AI is an amplifier. Strong engineering practices get stronger. Weak ones get weaker.
-
You cannot leave AI to run unsupervised. Treat it like a capable but junior team member: give it structure, review its work, and hold it to your standards.
-
Legacy codebases require more care, not less. Build characterisation tests first. Establish guardrails before you accelerate.
-
TDD is no longer a philosophical debate. It produces better AI-assisted systems, and the evidence is clear.
-
Architecture is a lever. The simpler and more expressive your codebase, the more value AI can deliver.
-
Iteration compounds. Build feedback loops, encode solutions, and your AI workflows will improve over time.
-
Engineering accountability is non-negotiable. The engineer owns the code, regardless of who or what wrote it.
-
Keep it simple. Multi-agent complexity and elaborate orchestration often create more problems than they solve. Experiment in your context, with your tools, and apply what works for you.
How Codurance can help
Codurance works with engineering teams and technical leaders to build the foundations that make AI-assisted development genuinely productive rather than just fast. That means strong engineering practices, clear architectural direction, effective testing strategies, and the coaching to make it stick.
Whether you are starting to explore AI tooling, working through a legacy modernisation challenge, or looking to build lasting capability across your organisation, we can help you get there without cutting corners on the things that matter.
Ready to build the engineering foundations that unlock AI’s full potential? Talk to our team.
Frequently Asked Questions
Is AI actually delivering 10x productivity for engineering teams?
For teams with strong engineering practices, significant acceleration is achievable. The 10x figure is possible in specific, well-structured scenarios, but it depends heavily on the quality of your codebase, your testing discipline, and how actively engineers guide the AI. For teams with poor practices, AI typically accelerates problems rather than productivity.
Can AI be used effectively with legacy codebases?
Yes, but it requires more groundwork. You need to establish characterisation tests, set clear guardrails, and be explicit about the patterns AI should follow in new areas of the codebase. The acceleration is still substantial, but the preparation work is not optional.
Does TDD still make sense in an AI-assisted workflow?
More than ever. Tests provide the specification and feedback loop that AI needs to produce good output. Without them, AI-generated code is harder to validate and more likely to drift in quality. Even long-time TDD sceptics are finding that the combination of AI and Test-Driven Development produces results that are difficult to argue with.
Who is responsible for code that AI writes?
The engineer who commits it. AI is a tool, not a co-author with accountability. Engineers need to understand, review, and own every line of code that goes into production, regardless of how it was generated.
What does good AI-assisted engineering actually look like day to day?
It looks like disciplined, human-led development with a capable assistant. Engineers define the scope, set the standards, review output, and iterate deliberately. The AI handles execution at pace. The human provides context, judgement, and accountability. Neither works well without the other.
About the Author
Sam Griffiths is a Principal Client Solutions Consultant with over 15 years’ experience helping organisations navigate cloud transformation, platform engineering and AI adoption. Operating at the intersection of technology strategy and commercial decision-making, he works with clients to identify opportunities, shape solutions and deliver technology investments that create lasting business value.
Watch our recent webinar recording where Sam, alongside our other Codurance experts, discuss the topics present within this blog.
Citations
This article is based on Codurance’s internal experiments and client work, as discussed in the webinar “The 10x Engineering Reality: Beyond the Hype to Proven ROI”.
Other links:
AI Project Delivers 10x Speed Without Compromising Quality
Software Craftsmanship in the AI Era
2025 DORA Report
.png?%20AI%20Due%20Diligence%20%20(1).png&width=528&height=528&name=Codurance%20-%20Technical%20%26%20AI%20Due%20Diligence%20%20(1).png?%20AI%20Due%20Diligence%20%20(1).png)
-1.jpeg?width=528&height=528&name=IMG_2121%20(1)-1.jpeg)
.png?width=528&height=528&name=Blog%20header%20(2).png)