Architecture as Code: how LLMs accelerate system design

Architectural documentation in large companies is almost always a pain. Weeks of approvals, endless diagram edits in Draw.io, versioning through Confluence that looks more like an archaeological dig than an engineering process. The AI department team of the BCS Group decided it was time to end this and showed how the Architecture as Code approach combined with LLM assistants can radically change the rules of the game.

Alexey Pronsky, who heads architecture in BCS's AI division, described a problem familiar to every corporate architect. His team builds agent systems, AI assistants, OCR solutions, speech analytics, and classical ML models. Behind each such project lies an architectural decision — a document that must go through a multistage approval process with business, corporate architecture, the information security service, and owners of related systems. On average, the path from receiving business requirements to final approval takes two to three weeks. For an industry where iteration speed determines competitiveness, this is unacceptably slow.

The essence of the proposed approach is moving architectural documentation from visual editors and wiki systems into code. The BCS team chose Structurizr — a tool that allows describing software system architecture using a special DSL. Instead of manually dragging blocks on a Draw.io canvas, the architect describes components, relationships, and contexts in text. This provides all the benefits that developers have long gained from the Infrastructure as Code approach: versioning through Git, code review, automatic diagram generation, and critically, the ability to connect an LLM assistant.

This is where things get really interesting. Pronsky shows how Claude Code, acting as an LLM assistant, can take on a significant portion of the architect's routine work. When architecture is described in code rather than pictures, a language model can analyze existing structure, suggest changes, generate new components, and even help prepare documentation for approval. Essentially, the same thing that happened with code writing over the past year is happening — LLM assistants don't replace the specialist, but dramatically accelerate their work by handling routine tasks.

It's important to understand the context in which this practice emerged. Architecture as Code is not a new idea. Tools like Structurizr, PlantUML, and Mermaid have existed for a long time. However, before the advent of powerful language models, textual architecture descriptions remained a niche approach: the entry threshold was high, and the advantage over visual editors was unclear. LLMs changed this equation. A model that works freely with text and code transforms Architecture as Code from an elegant but labor-intensive practice into a genuinely efficient workflow. An architect states requirements, the assistant generates a DSL draft, the person reviews and corrects — and the cycle shrinks from weeks to days.

For the enterprise environment, this approach carries additional benefits. When architecture lives in a Git repository, every change is transparent and traceable. Code review of architectural decisions becomes as natural a process as software code review. The security service can automate part of the checks. And most importantly, dependence on a specific visualization tool decreases. Diagrams are generated automatically from the code and can be rendered in any compatible renderer.

The BCS experience is indicative also because it's not about a startup experimenting with new approaches, but about a large financial group with strict documentation and approval requirements. If Architecture as Code with LLM support works in such a regulated context, the approach is mature enough for broad corporate adoption. We will likely see a wave of similar implementations in the coming year — especially in companies that already actively use LLM assistants for development and want to extend this practice to related engineering disciplines.