Every software architect has faced it: the legacy system, a labyrinth of undocumented choices, where the original builders have long since vanished, taking with them the crucial "why." It's a problem that breeds technical debt and slows innovation. But what if AI could help bridge that knowledge gap, not by making the decisions, but by meticulously documenting them? A new approach, detailed in a recent article, suggests that AI-powered architectural decisions can transform how teams capture and understand the rationale behind critical design choices. According to the announcement, this isn't about AI replacing human architects, but augmenting their ability to formalize the often-overlooked "why."
The article, penned by Salesforce Lead Developer Advocate Dave Norris, champions a structured method called the Architectural Decision Record (ADR). An ADR serves as a formal document that captures the context, alternatives, and final rationale behind a significant design choice. The problem? Crafting a well-thought-out ADR is time-consuming. This is where AI steps in, not as a decision-maker, but as a tireless assistant.
The Architect in the Loop
At the heart of this new methodology is a "human-in-the-loop" (HITL) approach. It's a disciplined practice that recognizes the current limitations of Large Language Models (LLMs). While an AI can rapidly research options and format information, it lacks the project-specific context, stakeholder empathy, and nuanced experience that define a skilled architect. The architect remains firmly in control, using the AI to accelerate the tedious parts of the process while providing the critical oversight and judgment.
To achieve this, Norris outlines a five-step "prompt chaining" strategy. Instead of one massive query, the architect guides the AI through a series of focused prompts, with checkpoints for human review and refinement at each stage.
A Five-Step Blueprint for AI-Assisted Decisions
Using the common architectural dilemma of a single-org vs. multi-org Salesforce setup as an example, the process unfolds as follows:
