Keeping a quality bar in an AI design workflow
AI tools made one part of design almost free: producing more. More layouts, more copy variants, more icons, more screens. That’s genuinely useful, and it creates a new problem. When adding things costs nothing, the scarce skill is deciding what not to keep.
A quality bar used to survive on a few senior people’s taste and the friction of making things by hand. Both are weaker now. If you don’t write the bar down and put it in specific places in the workflow, it dissolves into a series of reasonable-sounding compromises. This is how I keep it.
Where AI drafts and where a person decides
My design method runs in five verbs: Sense, Frame, Explore, Decide, Ship. AI helps at every stage, but it helps differently.
- Sense. AI clusters research and support signal quickly. A person decides which themes are real and which are just loud.
- Frame. AI drafts outcomes and assumptions. A person picks the bet and writes the rule for what evidence would change it.
- Explore. AI generates options in volume. A person kills most of them.
- Decide. This one stays human. Agents can lay out trade-offs; they shouldn’t break ties on things customers will feel.
- Ship. Automated checks catch regressions. A person says "not yet" when it works but isn’t good.
The pattern is consistent: machines widen the funnel, people narrow it. The quality bar lives at the narrowing points. Product thinking covers the decision rules, and AI action stacks covers the pipeline. This piece is about the narrowing itself.
A rubric an agent can score against
"Make it better" gives an agent nothing to work with. It also gives a junior designer nothing, which is why good design leads have always translated taste into criteria.
A useful rubric is short, specific and scored. Mine usually covers:
- One job. Can you say in a sentence what this screen or page is for? Does everything on it serve that?
- Hierarchy. Is there one obvious first thing to look at, and one obvious next action?
- Restraint. Is every color, weight and element earning its place? Could you remove one thing without loss?
- Voice. Does the copy sound like us, and does it make only claims we can back up?
- States. Are empty, loading, error and edge cases designed, or only the happy path?
- System fit. Does it use existing parts and tokens, or did it invent something?
Agents can score drafts against criteria like these before a person sees them, which filters out the obvious misses. The rubric doesn’t replace judgment; it makes judgment faster by clearing the noise first. Rules about the system itself, like which tokens are allowed, are a separate layer I wrote about in design systems with AI.
See the first draft before anyone polishes it
The most expensive mistake in an AI workflow is polishing the wrong thing. Generative tools make polish cheap, so a wrong direction can look finished within an hour, and finished-looking work is hard to kill.
So I look at version one early, while it’s still rough. Structure, message and priority get argued about before anyone refines a pixel. If the direction is wrong, we’ve lost an hour, not a week. If it’s right, polishing is the easy part.
This is a habit worth making explicit on a team: first-draft reviews are short, frequent and expected, not a sign that someone is behind.
Subtraction is the main edit
Most of my edits on AI-generated work are deletions. Extra sections, decorative elements, hedge words, a third call to action, a second accent color. Generative tools default to more because more looks like effort.
I’ve written before about letting go of your first idea so the team can find a better one. The AI-era version is letting go of the second, third and fourth idea the tools hand you. A strong edit leaves fewer, sharper things.
A practical rule: after every round, try to remove one thing. If nothing can go, the work is probably ready.
Build quality in early, then ratchet
Years ago at Plume I learned that the foundations you lay early, systems, processes and insisting on quality in the first release, barely move later. I wrote about that in What I’d Build Earlier. Agent-written work raises the stakes, because the volume of output is higher and the cost of a weak default multiplies.
The other half is the ratchet: using work already in motion to fix papercuts, rather than waiting for a quality sprint that never comes. That idea is in Papercuts, Tech Debt and Stalled Features. With agents, the ratchet gets easier, since small fixes are cheap to make, as long as someone decides which papercuts matter.
Running the review itself
A rubric is only as good as the review it lives in. A few habits keep design review useful when much of the work under review came from a tool:
- Review the brief before the output. If the brief was vague, the output will be too, and critiquing pixels wastes everyone’s time. Half of the fixes I make in review are really fixes to the instructions.
- Ask what was cut. Generative tools offer many options. The interesting question isn’t "why this one?" but "what did you throw away, and why?" Good designers can answer quickly.
- Separate rule violations from judgment calls. A wrong token or a missing error state is a rule problem and can be sent back without discussion. Whether the hierarchy is right is a judgment call and deserves the conversation.
- Write down every decision that will come up again. If the same taste call happens twice, it belongs in the rubric or the system rules, not in someone’s memory.
- End with a verdict. Ship, revise, or stop. Reviews that end with "some thoughts" produce another round of the same work.
The goal is a review that gets faster over time, because the recurring calls have moved into written criteria and the conversation is saved for the decisions that are genuinely new.
Signs your bar is slipping
Watch for these on any team using AI heavily:
- Reviews mostly approve. If almost nothing gets sent back, the bar has moved.
- Pages and screens get longer over time, never shorter.
- New colors, sizes or components appear without anyone deciding to add them.
- Copy starts sounding generic, and nobody can say whose voice it is.
- Empty and error states are missing from reviews.
- People say "it’s fine" more than "it’s good."
Any one of these is normal. Three at once means the narrowing points aren’t doing their job.
The short version
AI widens what’s possible. A quality bar decides what’s acceptable. Write the rubric, review early, delete often, and keep the final call with a person who can say, plainly, "not yet."
Related When design is the bottleneck: fixing flow between design, product and engineeringWhat "AI builder" means on a Head of Design's résuméAI action stacks