Liongate Modernises Legacy ERP with AI in 65-Person-Days

The system was running stably and had largely been written off. Day-to-day operating costs attracted little attention, yet technical debt continued to grow. Ten business domains shared a single codebase, permissions and process control ran on custom-built components, and schema changes were made by hand.

Three options were on the table. Continuing as-is simply defers the problem. A complete rebuild with equivalent functionality ties up a budget that creates no additional business value. A lengthy analysis phase consumes months of capacity across the business and typically ends with slide decks rather than results.

There was a further complication: the system had no written specification. Its logic lived in the code and in the institutional knowledge of individual developers. Any modernisation effort launched without first reconstructing that knowledge would have been built on uncertain ground.

We chose a fourth path: measure first, understand second, build third. Each phase delivers a result that is usable independently of what follows. From the outset, the business, IT and procurement teams worked from the same set of numbers.

The starting point was an AI-assisted function point analysis to ISO 14143. The method is established and vendor-neutral. It measures functional scope across inputs, outputs, data stores and interfaces, providing an objective basis for effort and budget estimates before a single euro of development budget is committed.

In the second phase, we reconstructed the business logic from three independent sources: source code, existing documentation and automated tests running against the live system. This produced a complete domain model, structured use cases, and rules and algorithms in pseudocode. In the third phase, we defined the target architecture: strictly layered, one package per business domain, dependencies flowing downwards only. Four proprietary components were replaced by established standards.

In the implementation phase, a squad of more than ten AI agents worked in parallel across design, implementation, testing and quality assurance. Experienced developers managed the orchestration and made all significant architecture and quality decisions. Because the requirements from phase two were complete and unambiguous, the agents were able to work in parallel effectively.

Across all four phases, human effort totalled 65 person-days, compared to over 500 in a conventional approach. The work does not disappear — it shifts from writing to reviewing and deciding.

The ISO 14143 measurement required 3 person-days, against around 15 in a conventional project. Specifying the business logic took 7 person-days, compared to around 120. Architecture planning and decision-making accounted for 5 person-days instead of around 40. Implementation concluded at 50 person-days, where projects of this size conventionally budget around 330.

Beyond the reduction in effort, the project delivers two further outcomes that extend well past go-live. For the first time, the system’s business knowledge is fully documented: domain model, use cases and calculation rules are committed to writing and remain available to the organisation regardless of which individuals leave in future. And the target architecture is built on technologies with broad market availability. Long-term operations and maintenance remain plannable, and the freedom to choose who develops the system further is preserved.

The approach is repeatable. What we applied here can be transferred to any legacy system that has grown over years and whose documentation no longer reflects reality. The function point analysis creates transparency before budget decisions are made. The AI-assisted specification makes implicit knowledge explicit. And the layered target architecture ensures that the new system does not face the same problems in ten years that the old one faces today.