Blum describes AI adoption as an uneven three-act process. Initial experiments feel dramatically faster, larger tasks expose bugs and erode trust, and mature adoption begins only when teams learn to supply context, verification and guardrails. Organizations must support people at all three stages while preventing experienced engineers from becoming overloaded quality bottlenecks.
Verification provides the highest-leverage foundation. Figma shifts checks toward compilers, linters, unit tests and deterministic workflows, then uses agent review where rules can be encoded and keeps human review for product intent and functionality. Test-driven instructions work better than asking an agent to retrofit tests after implementation because the code is optimized against an independent target.
Blum distinguishes planning from prompting. Engineers can spend substantial time clarifying the purpose, decisions, independently verifiable phases and acceptance criteria, then let agents implement a sequence of small reviewable changes. This preserves human agency and craftsmanship while allowing large amounts of coding work to run quickly and in parallel.
He recommends putting skeptics in charge of the improvement roadmap, clearly labeling human-written and AI-generated communication, and meeting employees in the tools they already use. Adoption is as much a cultural shift as an engineering one, so teams should make attention costs visible and let reliable experiences build confidence incrementally.
Watch on YouTube
