As AI Drafts More Work, Human Value Shifts to Specification and Evaluation

AI agents are increasingly capable of producing first drafts of code, analysis, and documents, reducing the cost of execution across many professions. According to Google Chrome DevTools veteran Addy Osmani, speaking on The Pragmatic Engineer podcast, the skills gaining value are those on either side of generation: precisely defining intent before work begins and critically judging output afterward. Evaluating AI-generated work is harder than it appears, since reviewers lack the contextual understanding they would have built by doing the task themselves. A data scientist example illustrates the risk: an AI-produced cohort analysis silently excluded 8% of users with null signup dates, a flaw only catchable through deep knowledge of the data's history. Professionals are advised to audit how much of their weekly work involves throughput an agent could absorb, and to deliberately strengthen the judgment and domain expertise that remains.