AI in Design: What Really Works and What Is Still Overpromised

Steven Heine, founder of the design tool Paper, spent months conducting anonymous interviews with designers at Atlassian, Shopify, Notion, and similar companies, finding that practice differs from public discourse. AI proves strongest in auditing and documentation: using Cursor and Figma MCP, Kirk from UI Collective built a system that checks design components against token tables, achieving roughly 50x acceleration compared to manual checks. However, attempts to generate dashboards through Figma Make produce visually similar results that don't use real variables when placed back into Figma. Another pattern is designers getting their own forks of production repositories to prototype against real code rather than imagined mockups, though Heine notes no one could confidently say this is faster. The AI in Design Report 2026, based on a survey of over 900 designers in 60+ countries and case studies from teams at Anthropic, Framer, Linear, Notion, Shopify, Sierra, and Stripe, shows peer learning rose from 24% to 70% while trust in top-down recommendations fell from 32% to 16% year over year. Alexey Kirdyaev writes on Habr that AI exposes whether trust exists in a team; a PM building a working dashboard prototype in Claude Code tests the team's willingness to reassign responsibilities. The most concrete example comes from X5Tech, where ControlNet, LoRA, and style transfer turned generation from a lottery into predictable variations, with 40-45% of visual content for Pyaterochka and Chizhik brands now created with AI, parts of teams abandoning stock photos, and design speed nearly doubling. Overall, AI is embedded where tasks are narrow and verifiable, but fails where creating new structures is needed; the boundary between smooth and chaotic adoption is determined by organizational trust.