Designing with AI in the loop

When generative AI became genuinely useful for design work, most teams treated it as a toy — a faster way to make the same things. I treated it as a question about how a design function should be structured.

We were a small team carrying a multi-product ecosystem. The constraint wasn’t talent; it was throughput. The interesting move wasn’t “use AI to design faster.” It was: where in the practice does AI remove a structural bottleneck, and where does it quietly degrade quality if you let it?

The bottlenecks, named honestly

I mapped our actual flow — not the idealized one — and found three places where time leaked:

1. Research synthesis. We gathered enough signal; turning it into something the team could act on took days.
2. Concept breadth. Under deadline pressure, we converged early. We were exploring two or three directions when we should have been exploring eight.
3. Handoff. The translation layer between a finished design and what engineering could build was slow and lossy.

What I changed

I rebuilt each stage with AI in the loop — and, just as importantly, decided where it stays out.

Research. AI accelerates the path from raw input to a structured, debatable synthesis. It does not decide what matters — that judgment stays with the designer. The output is a draft to argue with, not an answer to accept.

Ideation. This is where AI earns its place. The cost of generating a divergent option dropped close to zero, so we stopped rationing exploration. The skill shifted from producing concepts to critiquing them well — which is the more senior skill anyway.

Handoff. With an engineering partnership and a token-driven system underneath, AI compresses the mechanical translation work so the conversation can stay on intent and edge cases.

The goal was never to design faster. It was to spend the saved time on judgment — the part that doesn’t scale by adding tools.

Leading the team through it

A practice change is a people change. I trained the team end-to-end — not on prompts, but on a working model: where AI is a force multiplier, where it’s a trap, and how to tell the difference under deadline. We ran this through weekly critique, hands-on 1:1 coaching, and design sprints where the new flow was the default, not the experiment.

The honest part: some of this was the team’s, not mine alone. My role was to set the direction, build the scaffolding, and protect the quality bar while the practice changed underneath us.

What it produced

Directionally — and these are personal and team workflow gains, not audited or commercial figures — research moves about 40% faster, handoff about 50% faster, and we explore roughly three times the concepts per cycle. Early-stage prototyping went from days to hours.

The number that matters to me isn’t any of those. It’s that a small team started operating with the range of a much larger one — without the quality erosion that “move faster” usually smuggles in.