AI has changed how design gets made. It has not changed what makes design good.

That distinction matters. Every week, teams ship AI-generated interfaces that look plausible and fail quietly: wrong hierarchy, generic patterns, no relationship to the actual user or the actual business. At Things, we've spent years shipping digital products (more than fifty of them) and the last few integrating AI into nearly every stage of our process. We've won a Red Dot for our design work; our founders have won Cannes Lions and Effie awards before that. So we care about craft, and we use AI heavily anyway. Those two facts are not in tension. This guide explains how they fit together.

What can AI actually do in UX and UI design?

AI is genuinely useful for synthesis, exploration, and acceleration. It can summarize research, generate wireframe variations, draft copy, produce visual explorations, and turn designs into working code. What it cannot do is understand your users, make judgment calls, or take responsibility for the outcome.

Think of AI as a tireless junior collaborator with an enormous memory and no taste of its own. It excels at:

  • Synthesis at scale. Feeding an LLM twenty interview transcripts and asking for recurring themes takes minutes. Doing it by hand takes days.
  • Divergent exploration. Generating fifteen layout directions for a dashboard, or forty naming options for a feature, is exactly the kind of wide-net work AI does cheaply.
  • First drafts. Wireframe skeletons, microcopy, empty states, error messages, component documentation: AI produces credible starting points fast.
  • Translation between mediums. Design-to-code tools now produce usable front-end scaffolding from Figma files or screenshots.

What it consistently fails at:

  • Knowing what matters. AI averages across everything it has seen. Your product wins by being specific.
  • Original interaction thinking. Ask AI for a settings screen and you get every settings screen ever made, blended.
  • Accountability. AI will confidently generate an inaccessible checkout flow and never lose sleep. Someone has to.

Can AI replace UX designers?

No, but it is replacing a certain kind of design work. Production tasks that were once billed by the hour (resizing, spec-writing, first-pass wireframes) are being automated. What remains, and grows in value, is judgment: research, strategy, taste, and the ability to decide.

This is the honest answer buyers deserve. If a vendor's pitch is "AI makes design cheap," they're selling you averaged output. The teams that thrive with AI are the ones who were already good without it, because AI amplifies whatever process it's plugged into. Amplified rigor produces speed. Amplified guesswork produces polished mistakes, faster.

The designer's job is shifting from making artifacts to directing outcomes. That's a promotion, not a replacement, but only for designers who can actually direct.

What does an AI-augmented UX/UI workflow look like in practice?

A practical workflow uses AI at every stage but keeps humans in charge of framing, judgment, and quality. Below is how it maps to our own process at Things (Understand, Shape, Build, Evolve), with the tool categories that actually earn their place.

1. Research and discovery (Understand)

Research is where AI helps most and misleads most. Use it to process evidence, never to invent it.

  • LLMs for synthesis. Transcribe user interviews, then use an LLM to cluster themes, extract verbatim quotes, and flag contradictions. Always trace insights back to real quotes, because an LLM will happily hallucinate a "user need" if you let it.
  • Heuristic evaluation support. LLMs with vision can run a first pass against Nielsen's heuristics or WCAG criteria on screenshots. Treat the output as a checklist to verify, not a verdict.
  • Desk research acceleration. Competitive teardowns, market scans, and pattern audits compress from weeks to days.

What stays human: writing the research questions, moderating interviews, and deciding which findings change the product. Synthetic "AI users" are useful for pressure-testing a protocol before a study, not for replacing the study. We still run real interviews and real usability tests, because the surprising answer, the hesitation, the workaround a user invented: that's where products get better, and no model contains it.

2. Ideation and information architecture (Shape)

  • LLMs as sparring partners. Feed the model your research summary and brief, then ask it to argue against your proposed structure, generate alternative user flows, or role-play a skeptical user walking through the journey.
  • Rapid IA drafts. Generate three competing sitemaps or navigation models, then critique them against the research. The value isn't in any single output; it's in how fast you can see the option space.

What stays human: the frame. AI answers the question you ask. Asking the right question (what job is this product actually hired to do?) remains the whole game.

3. Wireframing and prototyping

  • Figma AI and native design-tool assistants now generate low-fidelity layouts, rename layers, and populate realistic content. Good for scaffolding, not for decisions.
  • Prompt-to-prototype tools turn a text description into a clickable flow. We use these for testing a hypothesis in hours instead of days, then throw the prototype away. That's the correct relationship: AI prototypes are for learning, not shipping.

What stays human: hierarchy and flow. AI arranges components; designers decide what a user should see first, what should be hard to do, and what shouldn't exist at all.

4. Visual design

  • Image generation (Midjourney-class tools, and increasingly in-canvas generation) is excellent for moodboards, art direction exploration, and placeholder imagery. It is poor at consistency, which is precisely what a real interface demands.
  • Design systems are the multiplier. AI outputs become dramatically more usable when constrained by a well-built system of tokens, components, and rules. This is why we invest so heavily in design systems for clients: a strong system turns AI from a chaos generator into a disciplined production assistant.

What stays human: taste. Craft. The two pixels of optical adjustment that no model notices and every good designer does. Interfaces that win awards (and, more importantly, user trust) are built on thousands of small deliberate decisions.

5. Handoff and engineering (Build)

  • Code generation is the most mature AI capability in the stack. Design-to-code tools and coding assistants (Claude, Copilot-class tools) turn specs into components quickly, especially when the design system already defines tokens and behavior.
  • AI-written handoff documentation. Component specs, states, edge cases, and accessibility notes can be drafted by AI from the design file and reviewed by the designer.

Because Things is a design and engineering studio, this stage is where our model pays off: the people who designed the system review the code that implements it. AI-generated code without design review drifts; spacing collapses, states go missing, accessibility evaporates. Design. Code. Mastery. The middle word is not optional.

6. Post-launch (Evolve)

Use AI to summarize analytics anomalies, cluster support tickets and app-store reviews into themes, and draft A/B test hypotheses. Shipping is the beginning of learning, not the end of the project.

What are the quality risks of AI-generated design?

The main risks are genericness, hidden usability failures, accessibility gaps, and inconsistency. AI output regresses to the mean of its training data, which means it produces the most average possible interface, and average interfaces don't differentiate products.

Be specific about the failure modes:

  • Convergence. When every team prompts the same models, products converge on the same look. Distinctiveness becomes a competitive moat again.
  • Plausible-but-wrong flows. AI-generated screens look finished, which short-circuits scrutiny. A polished wrong answer is more dangerous than a rough one.
  • Accessibility debt. Generated designs and code routinely miss contrast ratios, focus states, semantic structure, and screen-reader behavior. These failures are invisible in a screenshot and very visible in a lawsuit.
  • System erosion. Ad-hoc AI output that bypasses the design system creates inconsistency debt that compounds with every screen.
  • Skipped research. The subtlest risk: because AI makes making cheap, teams skip understanding. You ship faster toward the wrong thing.

The mitigation for all of these is the same: human review with real standards, grounded in real research, enforced through a real design system.

How does Things use AI in its design process?

We use AI as augmentation, not replacement: at every stage, under human direction, with studio-grade review before anything reaches a client or a user. We're a studio, not an agency: small senior teams, deep involvement, no hand-offs to a production floor. AI fits that model well, because it removes grunt work without removing judgment.

Concretely: AI accelerates our research synthesis, widens our exploration, drafts our documentation, and speeds our engineering. Humans run the interviews, set the strategy, make the calls, and polish the craft. Our clients get the speed of AI with the standards that won us a Red Dot. That's the whole pitch, and it's an honest one.

FAQ

What are the best AI tools for UI design in 2026? Think in categories, not brand names: LLMs (Claude, GPT-class models) for research synthesis and copy; native design-tool AI (Figma AI) for layout scaffolding and file hygiene; image generation for art-direction exploration; prompt-to-prototype tools for fast concept testing; and AI coding assistants for design-to-code. Tools churn quarterly; a sound workflow outlasts any of them.

Can AI do UX research? AI can process research (transcribing, clustering, summarizing) dramatically faster than humans. It cannot conduct it. Real interviews and usability tests with real users remain irreplaceable, because AI can only remix what's already known.

Will AI-generated design hurt my brand? Unedited, probably: AI output trends toward the generic, and generic is the opposite of brand. Directed by strong designers inside a defined design system, AI speeds up distinctive work rather than diluting it.

Is AI-assisted design cheaper? It's faster more than cheaper. Production time drops; the share of budget going to strategy, research, and craft rises. The overall result is more value per week, not commodity pricing, and teams promising the latter are usually skipping the parts that make design work.

How do I evaluate an agency or studio that claims to use AI? Ask three questions: Where exactly does AI sit in your process? What does a human review before delivery? Can you show work shipped before AI existed? A team that was excellent pre-AI is being amplified. A team that wasn't is being disguised.

Ready to design with AI, properly?

If you're a founder, design lead, or marketing director trying to figure out where AI fits in your product's design and build, that's a conversation we have every week. We'll tell you honestly where AI will help, where it won't, and what it would take to ship something excellent.

Talk to Things: hello@things.ist · things.ist

Design. Code. Mastery.