AI-powered software migration: how one team transformed a six-week challenge into a four-week success story

The technical debt remained, leaving the company vulnerable with an unsupported PHP version that posed significant security risks.

Fast forward to 2024 and the story took a dramatically different turn. Armed with AI-assisted development tools, specifically Cursor powered by Claude 3.7, their Principal Engineer, Matt, accomplished what the entire team couldn’t complete the year before. He successfully migrated the entire codebase from PHP 7.3 to 8.3 in just four weeks, working essentially solo with AI as his primary collaborator.

The foundation for success

The Collecting Group’s success wasn’t just about having the right AI tools, though that was certainly crucial. Several foundational elements made this transformation possible. The company operates collectingcars.com and watchcollecting.com as online auction platforms, built on a complex, but well-structured hybrid monolith architecture combining PHP, Next.js, JavaScript and Node.js services.

Their existing codebase had good decoupling between components and comprehensive testing including end-to-end testing and unit tests throughout their build pipeline. This layered testing approach proved essential during the migration, providing confidence that changes would not break functionality.

Perhaps most importantly, they assigned their most experienced developer to the task. Matt, as Principal Engineer and one of the original builders of the codebase, brought deep institutional knowledge that proved invaluable when working alongside AI tools. This combination of human expertise and Artificial Intelligence amplification created a powerful force multiplier.

The journey from Copilot to Cursor

The team’s AI adoption didn’t happen overnight. They began their journey with GitHub Copilot about 18 months before tackling the PHP migration. The initial experience with autocomplete suggestions felt good to the development team, providing an immediate 20 per cent productivity boost that converted even the most sceptical team members.

As Olly recalls, the team’s excitement about Copilot naturally led them to explore other AI-powered development tools. Two engineers independently experimented with Cursor in their personal time and requested access to Cursor Pro. After running a two-month pilot programme, the results showed that Cursor offered capabilities beyond any that Copilot could provide in their context.

The key difference at the time was Cursor’s ability to understand entire codebases and maintain context across complex refactoring tasks. At the time, Cursor could engage in meaningful discussions about code architecture, maintain coding standards and handle larger-scale systematic changes that characterised the PHP migration project.

The PHP migration challenge

The transition from PHP 7.3 to 8.3 represented more than a simple version upgrade. This migration required updating virtually every dependency and library, touching nearly every code file and ensuring compatibility across the broader technology stack. The team couldn’t approach this incrementally since dependencies were interconnected and the entire system needed to move together.

During the migration process, Matt leveraged AI to handle approximately 80 percent of the coding work, while he focused on architectural decisions and quality assurance. The AI assistant could systematically work through dependency updates, import statement changes and compatibility adjustments, while Matt applied his expertise to determine which tests could safely be modified and where additional testing might be needed.

The structured nature of the PHP upgrade process played to AI’s strengths. Rather than asking the AI to create new features from loosely-specified requirements, Matt could provide clear context about the migration path and leverage the AI’s ability to apply consistent transformations across the codebase.

Risk management and security considerations

Adopting AI-assisted development tools required careful consideration of security and intellectual property concerns. Olly approached this systematically, starting with GitHub Copilot under their existing Microsoft enterprise license, which provided assurance that their code wouldn’t be shared externally.

For Cursor adoption, the team evaluated the terms and conditions carefully and ultimately moved to an enterprise license to ensure proper data protection. They also established access to Google’s Gemini Pro 2.5 through a separate contract that provided additional ring-fencing of their intellectual property.

The risk assessment considered that much of their core value proposition lay not just in the code itself, but in their operational expertise, hosting strategies and business processes. This broader perspective helped them feel more comfortable leveraging AI tools while maintaining appropriate security measures.

Building team capabilities and culture

One of the most significant insights from The Collecting Group’s experience is that AI-assisted development requires skill development similar to remote working. Teams can’t simply decide to use these tools effectively overnight; it takes practice, experimentation and developing familiarity with how to work alongside Artificial Intelligence.

The social proof aspect of their adoption proved crucial. When Matt demonstrated success with both the prototype product development and the PHP migration, his credibility as Principal Engineer helped other team members embrace the technology. Rather than mandating adoption top-down, the organic excitement and demonstrated results created genuine buy-in across the development team.

Individual developers have developed personal preferences for different AI models depending on the task at hand. Some favour Claude for complex reasoning tasks, while others prefer Gemini for specific types of analysis. This personalisation reflects the nuanced skill development that comes with extended AI tool usage.

Broader organisational impact

The success in software development catalysed AI adoption across other business functions. Marketing teams began using generative AI for photo editing and content creation, while the leadership team adopted AI tools for document writing and analysis. Olly set an explicit goal of getting the entire senior leadership team experienced with ChatGPT and other AI tools, recognising that familiarity with these technologies would become essential for strategic decision-making.

The productivity improvements extended beyond just code generation. For Olly, who describes himself as dyslexic, AI tools transformed email writing from an hour-long struggle into a five-minute task. These efficiency gains compound across an organisation, particularly when teams develop expertise in using different AI models for different purposes.

Looking forward

While The Collecting Group has achieved remarkable success with AI-assisted development, Olly remains realistic about current limitations. Fully autonomous code generation for complex features isn’t quite ready for production use and the team continues to focus on augmentation rather than replacement of human developers.

The rapid evolution of AI capabilities means staying current with new models and tools. The team actively experiments with Claude 4, continues using Gemini 2.5 Pro and evaluates emerging tools like Claude Code. This ongoing exploration ensures they can leverage improvements as they become available.

The competitive implications are significant. Organisations that don’t begin developing AI-assisted development capabilities risk falling behind teams that are building these skills now. As Olly notes, referencing Ayrton Senna’s wet weather driving advantages, AI represents a condition change that can enable dramatic performance differences between teams.

Key takeaways for development teams

The Collecting Group’s experience offers several important lessons for other development organisations considering AI-assisted migration projects. Most critically, the combination of experienced developers and AI tools produces significantly better results than either element alone. Garbage in, garbage out remains a fundamental principle, making senior developer involvement essential for complex migrations.

Comprehensive layered testing provides the confidence necessary to make large-scale changes. Without robust automated testing, teams cannot safely leverage AI for major refactoring work. Well-structured codebases with clear separation of concerns also enable AI tools to work more effectively than with tightly-coupled legacy systems.

Starting with pilot programmes and allowing organic adoption based on demonstrated results proves more effective than top-down mandates. Teams need time to develop skills and comfort with AI-assisted workflows and social proof from respected team members accelerates broader adoption.

The future of software development increasingly involves human-AI collaboration rather than human-only or AI-only approaches. Teams that begin building these collaborative capabilities now will be better positioned to tackle technical debt, implement large-scale migrations and deliver software more efficiently. For organisations facing similar migration challenges, the tools and techniques are available today to transform what once seemed impossible into achievable results within much shorter timeframes.