57 structured experiments. Research +60% faster, Design +80% faster, compound savings across every sprint.
Most organizations experiment with AI reactively and inconsistently. I launched a structured program in Q3 across the UX team. By Q2 2026, AI was embedded in how the team researches, designs, writes, and delivers.
The traditional UX process has a speed problem. Research takes weeks. Synthesis takes days. Concept generation requires protected time that delivery schedules rarely respect. The result is a design function perpetually catching up to the roadmap rather than getting ahead of it.
In 2025, I decided to fix that—not by cutting corners on quality, but by systematically identifying where AI could absorb the time-consuming parts of UX work without degrading the thinking behind it.
I launched a deliberate experiment program. Every team member—across research, design, and writing—was asked to incorporate AI experimentation into their Individual Development Plans. The goal wasn't to mandate tools. It was to create accountability for learning and a shared language for evaluating what worked.
The team ran 21 experiments in Q3. Each one documented hypothesis, tools used, estimated time saved, AI versus human effort ratio, and whether the person would use the approach again. The failures were as instructive as the successes. Some AI-assisted workflows produced output that wasn't worth the time it took to correct. Those experiments established where AI belongs in the process and where human judgment remains non-negotiable.
What We Learned About Where AI Adds Value
A consistent pattern emerged across disciplines. AI accelerated work that required scale without judgment—first drafts, synthesis of large inputs, option generation, structural scaffolding. It consistently fell short where precision, product context, and UX craft were non-negotiable. The human effort shifted from doing to directing and evaluating.
- UX Research: Custom GPTs and Miro AI synthesized interviews faster, enabling PM stakeholders to provide sharper research inputs. Result: faster time to first research synthesis and cleaner outputs.
- Product Design: Teams tested ChatGPT, Figma Make, Lovable, and v0 across multiple use cases. Lovable produced the strongest visual quality and flow coherence for early exploration. Figma Make worked best for refining against existing design comps but struggled under complexity. v0 was a reliable middle ground for quick concept validation. None were plug-and-play. All required designer oversight, UX judgment, and alignment to Muck Rack's design system. The tools accelerated generation. Designers elevated the output.
- UX Writing: Custom GPTs reviewed copy against accessibility standards and tone guidelines, ensuring messaging remained clear, inclusive, and consistent. The tool surfaced design system inconsistencies that would have required manual review.
Q3 Results
- 21 experiments completed against a target of 14.
- Recurring AI assistants—the Competitor Insights GPT, Accessibility Checker, Component Decision Tool, and UX Research Brief Assistant—saved on average 47 minutes per use.
- One project went from briefing to concept testing with users in under five work hours. A fully revised scope of work was delivered in less than 48 hours after the briefing—a cycle that traditionally would have taken over two weeks.
- The program was acknowledged at a company-wide All Hands as a model for responsible AI adoption.
Q3 answered whether AI could accelerate UX work. Q4 and early 2026 answered a harder question: could we build AI into how the team actually delivers, not just how individuals experiment?
In mid-February, we got access to Claude with GitHub integration into production code. The workflow evolved immediately.
- Prototyping moved into code. Teams began using Claude to explore interaction ideas directly, pushing output back and forth with Figma for refinement and finalization. The prototyping loop was now a conversational design process. What would have taken eight hours through traditional methods was produced in two hours, with an additional five to six hours saved on interaction refinement. Design gaps that would have taken significantly longer to identify manually emerged faster through the conversational workflow.
- Design specifications became a delivery artifact. Designers now use Claude to generate markdown design specification files from their work. Those specs feed directly into PRD requirements and engineering tickets, compressing the handoff cycle and reducing the ambiguity that typically lives between design and engineering.
- Copy generation became systematic. We converted the Voice and Tone GPT into a Claude Skill, enabling consistent, on-brand copy generation across features. What had been a bottleneck became a first-pass quality gate any designer could run.
- Accessibility and design system compliance became continuous. AI-powered accessibility audits and design system token creation now run as part of the workflow, not as a separate QA step at the end. Code syncing with the design system reduces drift and catches inconsistencies earlier.
- This program was intentionally designed from day one. The efficiency numbers didn't come from tool adoption, but from that deliberate approach.
- Every experiment was evaluated, not just completed. The failures informed the guardrails. The wins got codified into the Vibe Design Prototyping Guide, which establishes principles and guardrails for AI-assisted prototyping across the team.
- The core argument cuts against how most teams approach this: AI prototyping tools are a means to faster validation, not a replacement for problem-first thinking. The risk of AI adoption in a design practice isn't that teams will use it badly. It's that they'll use it to accelerate in the wrong direction faster.
- The team's craft, judgment, and product intuition remained the quality bar. AI raised the floor on first drafts. Human review and refinement is what elevated those drafts into work that met that bar.
Most organizations approach AI adoption reactively—individuals experimenting independently, no shared standards, no measurement, no codification of what works. We ran a structured program, measured it, built results into permanent practice, and shared it across the organization. That's the difference between AI as a productivity hack and AI as an organizational capability. One fades when the novelty does. The other compounds.