The changing face of migration work: discovery, automation, and human oversight

Legacy system migration is one of the most dreaded projects in software engineering. It's expensive, risky, and often compared to rebuilding a bridge while traffic continues to flow. While the industry debates AI's potential to transform migration work, some of us have moved past the conversation and into execution. As an engineer who's navigated multiple AI-assisted migrations, the reality is more nuanced and more practical than the headlines suggest. 
 

The AI Revolution in Development 

The shift from AI as experimental tooling to production-ready capability has fundamentally changed what's possible in software development. While generative AI is typically used to create new code, this same capability can be redirected toward understanding and migrating legacy systems. By applying AI to these problems, the legacy code itself becomes part of the context. This unlocks powerful capabilities to drive development via specifications and rebuild systems better than before. 
 

A New Development Paradigm 

My approach to development has drastically changed with a mix of AI-generated code and AI-assisted decision making, all driven from specification definitions and authorised codebase examples. This has improved the feedback loop of development, and I cannot deny the vibe-flow experience has made coding more joyful and educational. 

Yes, AI has teething pains. Hallucinations, context loss, and faulty assumptions remain issues. A human in the loop is vital to prevent misdirection while capitalising on the benefits. 
 

Practical Applications in Migration 

Migration of legacy systems can be a huge burden for any business. Old code over time becomes less functional, secure and understood. However, this is the perfect opportunity to modernise with generative code to capture workflows and logic. When directed by a small talented team, legacy systems can be rebuilt and improved to unlock developer productivity and ultimately a better user experience. 

I've benefitted from using this approach to scan an undocumented legacy code base, understanding and replacing capabilities that would otherwise take enormous time to reverse engineer. Legacy migration is largely about discovery, and with stretched resources, AI becomes a real enabler. It keeps complex migration work on track while freeing developers to focus on optimisation. 
 

Domain Expertise Remains Critical 

Generative AI does not by default produce the most domain-optimised output. This is especially true for code security which, is best managed by incorporating suggested code into existing security tooling. For other domain specialised tools and frameworks such as CMS and Digital Asset Management, AI agents can be introduced to key parts of the workflows. These can augment tasks which are often complex and precisely tuned to the business's needs. 

Esoteric languages and legacy software versions require unique considerations, as LLM training cut-offs limit information validation. Domain expertise combined with traditional unit testing is invaluable for steering interpretation. 

Where there are gaps in legacy systems and security standards (say one application does not have strict mode enabled for JavaScript whereas the new web application does), this could potentially mean a comprehensive rewrite and replacement of code where suitable libraries and functions need to be factored in. Only an experienced developer with working knowledge of the Document Object Model can navigate these challenges, and this in turn changes the direction of questions and decisions that AI would otherwise not see. 

Four Levels of AI Decision Making 

It's worth considering there are four levels of AI decision making: 

• Assisted Decision Making: AI supports human choices  

• Verified Decision Making: AI proposes, humans approve  

• Delegated Decision Making: Humans oversee AI autonomy 

• Autonomous Decision Making: Full AI independence 

Different decision level suit different use cases and scenarios. For environment with set business flows and data for repetitive tasks such as credit checks or loan approval, this could fall into level 4 where automation is not quite enough due to more dynamic real time data. For higher-level decisions such as deciding on complex architecture decisions with SaaS product trade-offs, only level 1 would be the right and diligent choice. 

Real-World Success: CMS Migration 

Migration tasks can be complex and can encompass the entirety of the architecture. This is especially true with Content Management Systems (CMS), where even advanced teams can experience issues ensuring the quality, performance and governance of content on a new platform. Use cases like this are where I'd recommend Nimbus, our own CMS platform migration framework that utilises bespoke optimisation capabilities with AI enrichment to streamline the upgrade journey. 

Using Nimbus to replace a legacy CMS, we were able to capture an entirely manual asset management process. Updates to content and structure were outage-prone due to large complex files that made asset versioning nearly impossible. Once all the visual content was mapped, enriched and pivoted to a fully automated workflow, we overcame a huge pain point. 

Just as software needs updating and coding standards, the user experience must be considered a priority. The visual look and feel of any competitive website or application are felt before the performance comes into play. Having high quality digital assets and CMS is vital for an end-to-end experience. At MSQ DX partner with some of the leading CMS providers and take an active part in defining roadmaps to help achieve award-winning results for our clients. 

Looking Ahead 

AI will not replace me but change me. I'll be able to achieve more, innovate more and explore new possibilities, but not without human insight and meaning. In the long-term, I'd expect to see new roles emerge specifically around AI-augmented migration work as the demand to modernise legacy systems continues to grow. 

Products like Nimbus wouldn't exist without AI technology. What was once purely manual, requiring significant developer time to painstakingly map and transfer content, can now be intelligently automated with AI enrichment. This shift is making complex migrations feasible for organisations that previously couldn't justify the cost or risk.