Tool, teammate or threat? Experimenting with AI in digital product design
AI used to be theory. Now it sits in your stack, speeding up work and forcing harder questions about what quality actually means.
The real shift isn't whether AI can build products. It's whether it can help you build products that matter, repeatedly, without losing the taste and judgment that set great teams apart.
The risk of sameness
AI is great at patterns and averages. That's also the trap. Left alone, it trends toward lookalike interfaces and me-too ideas that feel flat.
Exceptional products don't come from probability. They come from taste, perspective and intent. Human-in-the-loop isn't a buzzword-it's the guardrail that keeps purpose and originality intact.
Tools, not teammates
Treat AI like a tool and you'll get leverage. Treat it like a teammate and you'll get blind spots-context misses, weak decisions and work that's "fine" but forgettable.
Real productivity is alignment, not speed. Agentic tools help coordinate steps, but the quality jump happens when people shape constraints, set standards and step in at the right moments.
Case study: a minimum testable product for meal subscriptions
We built a minimum testable product for Soph's Plant Kitchen to assess appetite, retention and the commercial case for curated meal plans. Underneath, the goal was to pressure-test AI across ideation, design and delivery.
Generative UI tools like UIzard and v0 gave us fast starts and slick demos-about 80% there. The last 20% required human eyes: visual harmony, creative finesse, accessibility, and the micro-decisions that make an interface feel intentional.
What the AI stack did well-and where it didn't
- ChatPRD to draft backlogs in Linear: useful scaffolding, needed pruning.
- Custom GPTs for recipe formatting: fast, required spot checks for consistency.
- NotebookLM for research synthesis and user stories: strong summaries when tightly prompted.
- Cursor, Figma and GitHub Copilot to bridge design and code: quick components, light on polish and accessibility.
Setup had a cost. New workflows, new habits and adoption time created an initial dip before the lift.
Most AI outputs still needed refinement or replacement. Speed increased the QA burden. We shifted from making to moderating-less "generate" and more "critique," where taste, context and constraints do the real work.
The outcome wasn't "built by AI." It was built by people using AI with intent.
A practical playbook for product teams
- Use AI for scope and scaffolding: hypothesis generation, backlog drafts, content transforms, component stubs.
- Keep humans on the hard parts: product bets, interaction flows, information architecture, accessibility, editorial voice.
- Design the constraints: tokens, grids, tone, forbidden patterns, acceptance criteria. Feed them to your tools.
- Write playbooks, not prompts: multi-step instructions with examples, edge cases and "don't do this" rules.
- Define the 80/20 line: what "good enough to hand over" means for design, copy and code.
- Close the loop: linting, accessibility checks, visual diffing, content QA, performance budgets.
- Label everything: what's real, what's mock, what's generated. Prevent demo confusion.
- Track rework: time saved vs. time spent reviewing and fixing. Kill steps that don't pay back.
Managing expectations with stakeholders
AI can make something look finished before it works. That's the trap. Call out what's conceptual versus functional in every demo.
Anchor progress to a working slice, not screenshots. Use a definition of done that includes usability, accessibility and system coherence-not just "it runs."
Metrics that matter
- Time to first interactive prototype (with real data if possible).
- Rework ratio: reviewer minutes per AI-generated artifact.
- Defects caught by QA vs. users.
- Accessibility audit pass rate (see WCAG).
- Task success and time-on-task from usability tests.
- Activation, conversion, retention on live experiments.
Where to deepen practice
If you're formalizing AI-in-the-loop, Google's People + AI Guidebook is a solid reference for team patterns and ethical guardrails: People + AI Guidebook.
For hands-on upskilling by role, explore curated AI learning paths: AI courses by job.
Bottom line
AI expands output. People create meaning. The gap between the two is your advantage-if you build systems that make judgment the center of the process.
Treat AI like a tool, keep your standards high and ship work that doesn't blend into the feed.
Your membership also unlocks:
