AI Needs Architectural Governance.

𝗔𝗜 𝗡𝗲𝗲𝗱𝘀 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗮𝗹 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲. 𝗪𝗲 𝗔𝗹𝗿𝗲𝗮𝗱𝘆 𝗞𝗻𝗼𝘄 𝗛𝗼𝘄 𝘁𝗼 𝗗𝗼 𝗜𝘁.

Over the last year, discussions with fellow CTOs, architects, and engineering leaders have repeatedly surfaced the same question:

❓ What does good governance look like in AI-first engineering?

I'm fully convinced that AI changes how we should build software, but it doesn't change the fundamentals of engineering.

Architectural Decision Records (ADRs) have been one of my favourite tools for 𝗱𝗲𝗰𝗲𝗻𝘁𝗿𝗮𝗹𝗶𝘇𝗲𝗱 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻-𝗺𝗮𝗸𝗶𝗻𝗴. They keep the experts closest to a domain informed and able to influence outcomes, without creating delivery bottlenecks.

ADRs are deliberately lightweight. Anyone can propose a change, but they must record the decision, alternatives considered, trade-offs, and rationale. Stakeholders can then review, challenge, refine, or endorse it.

The value is that ADRs capture 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗮𝗹 𝗶𝗻𝘁𝗲𝗻𝘁. They record not just what was decided, but why, and make you think. That becomes invaluable as systems evolve, teams change, and assumptions are challenged.

As architects, we've spent years moving away from top-down decision-making because it doesn't scale. Ironically, AI-first development can push us straight back into that behaviour. We become the person issuing commands: do X, no not like that, try again, still wrong.

I started using ADRs as guardrails for AI agents. Initially, they simply provided context: this is how we make decisions in this system, but then the agents started challenging me:

💬 "You asked me to do X, but that doesn't align with ADR-17?"

Context is winning, answer was "No. Good catch."

The problem wasn't the AI. The problem was me bypassing a decision-making process and AI willing to do what I say, which is an easy trap to fall into. As the human in the loop, I had unconsciously slipped back into top-down mode issuing orders not goals.

Today, my AI-first workflows generate ADRs by default. Agents propose options, document trade-offs, and raise questions when architectural direction is unclear. 𝗦𝗼𝗺𝗲𝗯𝗼𝗱𝘆 𝘀𝘁𝗶𝗹𝗹 𝗵𝗮𝘀 𝘁𝗼 𝗯𝗲 𝗮𝗰𝗰𝗼𝘂𝗻𝘁𝗮𝗯𝗹𝗲.

This is where I think many organisations are struggling. When AI produces poor outcomes, its easy to blame the technology. In practice, AI failures are usually 𝗴𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 𝗳𝗮𝗶𝗹𝘂𝗿𝗲𝘀.

⚖️ AI is capable of making decisions continuously and at a scale humans simply cannot match. That makes 𝗱𝗲𝗰𝗲𝗻𝘁𝗿𝗮𝗹𝗶𝘇𝗲𝗱 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻-𝗺𝗮𝗸𝗶𝗻𝗴, clear architectural direction, and explicit accountability more important than ever.

Architects already have all the patterns needed to solve this problem, we've spent years doing ADRs, architectural guardrails, clear ownership, and lightweight governance.

The challenge for AI-first engineering isn't inventing entirely new governance approaches. It's recognizing that what we've already developed are exactly what AI systems need.