Modernizing a legacy system rarely fails because AI cannot generate code. More often, it fails because organizations no longer have a current, trustworthy specification of what their systems actually do.
Large enterprise landscapes evolve over decades. New systems are added through acquisitions. Others are rewritten, partially replaced, or integrated with third-party platforms. Documentation falls behind. Over time, organizations are left with hundreds of repositories but no single, trustworthy specification describing how the portfolio actually works.
Brownfield to AI-ready
To solve this problem, we built a structured recovery pipeline that reconstructs a complete picture of a legacy software landscape before modernization begins.
- It starts with Inventory & Triage, where every repository is profiled by technology stack, size, and purpose. Dead or duplicated repositories are filtered out, separating the active software landscape from accumulated noise.
- The workflow then moves into Recovery (AS-IS). Fan-out agents recover architecture, requirements, data models, APIs, and non-functional requirements directly from the existing systems. A separate skeptic agent validates the highest-impact findings against the source code before they become part of the recovered knowledge.
- The information is then transformed during Specification & Backlog into architecture documentation, feature specifications, non-functional requirements, rewrite specifications, and implementation-ready feature descriptions.
- Finally, Navigate & Decide provides an interactive view of the recovered system—from portfolio level down to individual features—allowing engineering teams to explore dependencies, understand system boundaries, and prioritize modernization based on verified information rather than assumptions.
Verification is built into every stage. Every recovered claim is file-cited and validated before it becomes part of the final specification. The pipeline ends with a human decision gate before any modernization work begins.
One framework, multiple outputs
Recovering a trustworthy specification is not the final goal. It becomes the foundation for everything that follows.
The same knowledge supports modernization planning, architecture documentation, security analysis, migration backlogs, and software testing.
We recently applied this pipeline to an enterprise software landscape consisting of 33 systems, hundreds of repositories, and nearly 12 million executable lines of code accumulated over years of growth and acquisitions. In approximately three days, we reconstructed a current, verified specification of the entire software landscape. That included all 33 systems, nearly 2,000 feature specifications, identified security concerns, and a sequenced modernization roadmap showing where engineering effort and risk should be prioritized.
Another example is test generation. Because tests are generated from reconstructed specifications rather than directly from legacy code, they reflect the intended behavior of the system instead of its accidental implementation. In one recent engagement for a large energy company with a legacy system of roughly 100,000 executable lines of code, this approach increased unit test coverage from 2% to 50% in 10 hours and generated the system’s first 125 integration tests in another eight. The same reconstructed specification scales toward whatever coverage target the context demands — from risk-based coverage for most systems to the near-complete coverage regulated industries require for certification.
From framework to practice
This framework, once originated in certification-grade testing, evolved into a reusable system of skills and agents, and is now part of Sigma Software’s AI SDLC practice within AI Compass.
Whether you’re modernizing a decades-old application portfolio, improving software quality, or preparing your engineering organization for AI, the first step isn’t generating more code.
It’s rebuilding a trustworthy understanding of the systems you already have.
Learn more about AI Compass and how Sigma Software can support your AI transformation: https://sigmasoftware.ai/
