Where AI Fits in My Design Process (and Where It Doesn't)

Two years ago, AI in design mostly meant novelty — generated images for moodboards, auto-written placeholder copy, demos that looked impressive on social media and fell apart in real projects. Today it's quietly woven into a large part of my week.

I use AI tools for research synthesis, rapid prototyping, copy exploration, and a surprising amount of thinking out loud. They've made me faster in some places and better in a few. But they've also introduced a new kind of risk: the temptation to skip the parts of the job that actually matter.

This is an honest account of where AI fits in my process, where it doesn't, and the rules I've settled on to keep the tools serving the work instead of the other way round.

Research synthesis: a strong first pass#

The most valuable use for me is synthesis. After a round of interviews, I might have eight transcripts, a pile of support tickets, and a spreadsheet of survey responses. Reading all of it is essential. Organizing it is tedious.

AI is genuinely good at the organizing. I'll feed in transcripts and ask for recurring themes, contradictions between participants, and direct quotes that support each theme. In a few minutes I have a draft affinity map that would have taken me half a day.

But it's a first pass, never the final word. The tools are good at surfacing what's said frequently. They're much weaker at noticing what's said once but matters enormously — the offhand comment from a nurse about a workaround that reveals a deep flaw in the workflow. They also smooth over tension, presenting a tidy summary where the real insight is that two user groups want opposite things.

So my rule is simple: I read every transcript myself. The AI synthesis is a scaffold I check against my own notes, not a replacement for them.

Rapid prototyping: from idea to clickable in an hour#

Tools like Lovable and Replit have changed the early stage of the second diamond for me. When I want to test an interaction idea — a novel filtering pattern, a multi-step form, a drag-and-drop flow — I can often get a working, clickable prototype in an hour instead of a day.

This matters because real interactions reveal problems that static mockups hide. A filter that looks clean in Figma might feel sluggish and confusing when you actually use it with realistic data. Getting to "real" faster means I find those problems faster.

The caveat is that these prototypes are disposable. They're for learning, not shipping. The code they generate is rarely something an engineering team should build on, and I'm careful to frame them that way. A prototype that's mistaken for a near-finished product creates false expectations about timelines.

Copy exploration: more options, faster#

Interface copy is one of the most underrated parts of design, and AI is a helpful sparring partner for it. When I'm stuck on an error message or an empty state, I'll ask for fifteen variations with different tones — direct, warm, technical, playful.

Most of them are mediocre. A few are interesting. Occasionally one is better than anything I'd have written. More often, seeing the range helps me articulate what I actually want, and I write the final version myself.

Where I'm careful: anything involving money, health, legal implications, or safety. AI-generated copy tends toward confident-sounding phrasing that can be subtly inaccurate. In a fintech or healthcare product, "your transfer is complete" versus "your transfer has been submitted" is the difference between a calm user and a support ticket — or worse.

Usability testing support#

Tools like Maze and UX Pilot have AI features that help set up tests, summarize responses, and flag drop-off points. These are useful for speed, especially for unmoderated tests where you're processing dozens of sessions.

But I still watch the recordings. Summaries tell you where people struggled; watching tells you why. The hesitation before a click, the cursor drifting back and forth between two options, the moment someone scrolls up to reread instructions — that's where the real insight lives, and it doesn't compress well into a bullet point.

Thinking partner#

This one surprised me. Some of my most useful AI sessions don't produce any artifact at all. I'll describe a design problem in detail — the constraints, the user, the trade-offs — and ask the model to argue against my preferred solution.

It's like having a colleague who's always available and has no ego about being contradicted. It won't always make good arguments, but it frequently surfaces a consideration I hadn't weighed properly: an accessibility concern, an edge case, a business constraint. Writing out the problem clearly enough for the tool to understand it is itself clarifying.

Where AI doesn't fit#

There are parts of the work where I've deliberately kept AI at arm's length.

Understanding people#

AI can summarize what users said. It can't sit across from a clinic manager and notice that she sighed before answering a question about scheduling. It can't build the trust that makes someone admit they've been using a spreadsheet behind the product's back for two years. Empathy built through direct contact is the foundation of everything else, and there's no shortcut for it.

Making the actual decisions#

AI can generate options. It can't own the decision of which one to ship, because that decision depends on context it doesn't have: the political dynamics of the team, the engineering capacity next quarter, the brand's long-term direction, the conversation you had with the founder last week. Taste and judgment come from accumulated context, and that context lives in people.

Final visual design#

Generated UI tends to converge on the same polished-but-generic look. It's fine for a quick prototype. It's not fine for a brand that needs to feel distinct. The details that make a product feel considered — a slightly unusual type pairing, a specific motion curve, a layout that breaks the grid on purpose — are still human work.

Anything I can't verify#

If I can't check whether the output is right, I don't use it. That rules out AI-generated statistics, citations, and accessibility claims unless I can confirm them independently.

My rules of thumb#

After a couple of years of daily use, here's where I've landed:

1. AI drafts, I decide. Every output is a starting point for my judgment, never a replacement for it.
2. Read the source material. Summaries are for navigation, not understanding.
3. Prototypes are disposable. Use them to learn, then throw them away.
4. Be stricter in high-stakes domains. Healthcare, finance, and anything legal get extra scrutiny.
5. Watch for sameness. If the work starts looking like everyone else's, that's a sign the tools are leading.
6. Protect the human parts. Interviews, critique, and decision-making stay human.

The work is the same#

The tools have changed a lot. The job hasn't. Design is still about understanding a problem deeply, making good decisions under constraints, and caring about the details that make something feel trustworthy and humane.

AI has given me back time — time I used to spend on organizing sticky notes and writing the twentieth variation of a button label. I try to spend that time on the parts of the work that only a person can do: talking to users, thinking hard, and caring about the result.