The Craft Behind the Output

• 3 min read •
AI Strategy

I came across this piece in HBR by David Duncan and Tyler Anderson, and it nails something I've been watching play out in my own work for months.

Their core argument is that as AI makes generating polished work easier, the skill that actually differentiates people becomes judgment. Knowing what to trust, what to question, and what to refine. Most organizations invest in prompting workshops and tool certifications, but almost nobody invests in the harder problem of developing professional judgment alongside the tools.

The insight that stuck with me is that traditional expertise rewards people who've internalized judgment so deeply they can't explain it. AI-era expertise rewards the opposite. You now need to articulate what good looks like, because articulation is the interface between human judgment and machine capability. AI has enormous knowledge and zero context. It knows everything published. It knows nothing about your specific client's politics, the recent shift in your market, or the anxieties of a particular stakeholder.

The four-step framework

Duncan and Anderson propose a process that turns every AI-assisted task into a development opportunity.

  • Form your own view first. Before opening any AI tool, scope the task and sketch a preliminary position. Without your own hypothesis, you have no basis for evaluating what AI produces. The test is whether you could critique a finished version specifically. Not "this seems fine" but "this framing doesn't fit because this client needs X."

  • Play devil's advocate with the output. Most people only use AI in one mode. They ask it to generate. But generation is where AI is strongest and your judgment is least visible. Push further by asking probing questions. What are the weakest assumptions here? How do these versions differ in their tradeoffs? How would this specific person react? What data sources are weakest? Each question forces reasoning into the open that would otherwise stay hidden.

  • Diagnose the gap between your view and AI's output. Three questions matter here. What did AI add that you missed? What did AI get wrong or miss entirely? And the most dangerous category, what looks right but isn't? Plausible, well-structured, but wrong in subtle ways that require domain knowledge to catch. These are the errors that survive casual review and cause the costliest failures.

  • Deliver with a reasoning trail. The deliverable isn't just the output. It's a brief explanation of how you got there. What AI produced, what you changed, and why. For managers, requiring this single thing makes judgment visible and coachable. For individuals, it builds a calibrated sense of exactly where to trust AI and where to scrutinize in your specific domain.

Why this matters

We're at a weird inflection point. The people who adopted AI earliest are the ones most at risk of stalling their own professional development. The tool is so good at making them look competent that nobody notices they're not actually growing. That's a problem for individuals, but it's a much bigger problem for organizations that need deep expertise to navigate novel situations.

The full article goes deeper on each step with a worked example of a competitive analysis. Worth the read if you're thinking about how to develop your team in this landscape.