Does AI help big-picture design differently than routine coding, or is the real dividing line whether you hold context the AI lacks?

This explores whether AI helps more, less, or just differently with big-picture design work (framing the problem, choosing an architecture, deciding what to build) than with routine coding. The collection has no head-to-head study that splits the two. What it has is a set of findings that, read together, point to a counterintuitive answer: the split isn't really between design and routine work. It's between work where you hold context the AI doesn't, and work where you don't.

Start with the coding studies, because they disagree. In a randomized trial at Google, AI coding features cut time on a complex task by about 21%, though the confidence interval was wide Do AI coding features actually speed up engineer productivity?. Another randomized trial found that experienced open-source developers working in their own mature codebases were 19% *slower* with early-2025 tools, even though they expected a 24% speedup Do AI coding tools actually speed up experienced developers?. One factor the authors name is the developers' deep familiarity with their code. When you already carry the design of a system in your head, the AI's suggestions are mostly something to check, not something that saves you time. Anthropic's internal survey shows the same tension from the inside. Engineers report big output gains, yet most say they can fully hand off only 0–20% of their work. They also worry that giving away the routine coding wears down the hands-on skill they need to catch the AI's mistakes Does AI assistance erode the skills needed to oversee it?. So routine work is the easiest to delegate, but delegating it may weaken your judgment on the design work above it.

The strongest evidence about higher-level work comes from outside software. In a field experiment with 776 Procter & Gamble professionals working on real product problems, individuals using AI produced solutions as strong as two-person teams without it. Their ideas also became more balanced across business and technical perspectives Can generative AI replace the benefits of having a human teammate?. That suggests AI's value in design may come less from speed and more from standing in for a collaborator: it brings in the viewpoints of people who aren't in the room. Meanwhile, systems like AIDE2 and the Darwin Gödel Machine let AI do design iteration itself, rewriting agent architectures and testing them on benchmarks. They reach parity with human-built designs, including on tasks the AI wasn't tuned for Does automated evolution match human-built agent performance? Can AI systems improve themselves through trial and error?. That only works where the design can be scored automatically, which most human design work can't.

There are two less obvious points here. First, working with AI is itself becoming a design problem: because an AI's context keeps shifting, the core skill moves from writing interfaces to engineering what the model sees How does AI context differ from conventional software context?. Second, how much design control engineers keep is often decided by company policy (approved tools, data rules) before personal preference comes into it Does personal preference shape how engineers use AI tools?. There is also a philosophical warning. AI can produce the outward form of a design without the reasoning that normally goes into one Does AI separate intellectual form from the thinking behind it?. A plausible architecture document is not proof that anyone did the design thinking.