Richard Hill made an interesting observation that is worth bearing repeating for those who see AI as a magic wand that fixes everything.

AI is not a substitute for process improvement; it is a force multiplier. Well-structured organisations that invest in standardising their workflows before adopting AI will see efficiency gains. Those that attempt to use AI as a shortcut to bypass process refinement will likely find that it exacerbates their existing problems.

Additional snooping around led to the 2025 DORA Report reinforcing this with its research findings:

The primary role of AI in software development is that of an amplifier. It magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones.

Oddly, this pattern isn’t new. I saw organizations adopt event-driven architectures because they heard it would make them faster and more scalable, without understanding message flows, eventual consistency, or debugging complexity. The technology amplified existing dysfunction. Poor observability became distributed tracing nightmares, unclear data contracts became multi-service coordination hell. Same thing is happening with AI.

Hill’s healthcare example makes it scarily concrete. If a hospital implements AI for patient record-keeping but medical notes are recorded inconsistently with varying terminology, incomplete fields, ambiguous shorthand. In this situation, AI will struggle to extract meaningful insights. Worse, it may generate unreliable summaries, leading to clinical errors. The problem isn’t the AI itself but the inconsistent data feeding it, or in today’s world: the context you are prompting the AI with.

This connects directly to measuring what matters. The obsession with “20% faster coding” ignores whether you’re solving the right problems. A well-prompted LLM that generates code faster doesn’t help if the team ships features customers don’t need.

Hill points out AI can play a constructive role as an efficiency identifier. AI-powered process mining tools can track how work flows through an organization, revealing bottlenecks that aren’t apparent to human observers. Retailers can use AI to spot discrepancies between stock levels and purchasing patterns. Manufacturers can apply AI analytics to identify supply chain inefficiencies. But organizations must be willing to act on these findings. If you lack the discipline to standardize processes before AI adoption, you’re unlikely to leverage AI-generated insights effectively.

Drew Breunig’s work on DSPy for geospatial conflation shows what systematic optimization looks like. He went from 60.7% to 82% accuracy on address matching by having GPT-4 optimize prompts against actual success metrics:

DSPy’s role is to optimize the prompt used for that smaller model… producing a 700 token prompt that increased the score from 60.7% to 82%.

This is the gap between systems thinking and prompt theater. DSPy optimizes against verifiable metrics. Manual prompt engineering optimizes against vibes. One scales with proper infrastructure, one doesn’t.

I’ve noticed teams that successfully adopt AI tend to have strong existing processes. Comprehensive tests, clear documentation, good version control habits-what Simon Willison calls vibe engineering. His characterization is spot-on that “AI tools reward top-tier software engineering practices”. Automated testing lets coding agents fly. Planning in advance makes iteration productive. Good documentation lets models use APIs without reading all the code. Strong version control habits become even more important when an agent might have made the changes.

AI amplifies these practices too. If your process is “ship fast, debug in production,” AI just makes you ship broken things faster. If your process is “test thoroughly, review carefully, deploy incrementally,” AI accelerates the entire quality loop.

Hill’s conclusion is worth sitting with: AI works best when it has a solid framework to build upon. Without that, it’s merely an amplifier of whatever dysfunction already exists. The organizations seeing real AI productivity gains are likely the ones who already had strong systems thinking cultures-they just got a multiplier on existing capability.

For organizations without that foundation, the uncomfortable truth is you probably need to rebuild core processes first. That’s not the tax you pay for jumping on the hype, it’s the prerequisite for making the technology work. As is common with automation; AI doesn’t make bad processes good. It makes bad processes worse, faster, and at scale.