Margaret Gould Stewart, 2014

“You have to design with data, but you have to follow your gut.”

Margaret Gould Stewart described this balance in her 2014 TED talk on designing for billions of people. Drawing on product work at YouTube and Facebook, she explained how large-scale design uses extensive evidence while still requiring judgment about human behavior, meaning, and the details worth preserving. Data reveals patterns in what people do, but it does not automatically explain why or determine what should matter. Gut judgment is not a license for taste to override users. It is informed synthesis shaped by craft, qualitative observation, values, and responsibility for consequences. The business value of the line appears when it changes a real decision. It directs attention toward assumptions that routine can hide and invites a team to examine the system behind an outcome instead of treating the visible result as a complete explanation.

Product teams can pair behavioral metrics with interviews, usability observation, accessibility testing, and explicit design principles. Before a broad rollout, they should identify the decision that data informs, the limits of the sample, and the human concern no metric fully captures. Application therefore requires an explicit mechanism: what should people do differently, what signal should move, and when will the team examine the result? Writing those elements down turns general agreement into a testable management practice.

At scale, small design choices can impose friction or harm on millions of people. Intuition may encode the experience of a narrow team, while data may optimize the easiest behavior to count. Diverse review, staged deployment, and accessible reversal are essential. Context matters because an idea that improves one situation can create harm in another. Leaders need boundaries, dissent, representative evidence, and a way to correct course. Accountability cannot be delegated to the quotation or the tool used to implement it.

AI systems add another layer between user behavior and design decisions. Teams need to inspect model error, feedback loops, and unequal outcomes, while preserving understandable controls and human support. Personalization should not make the product impossible to explain or govern. For one high-traffic feature, write what the quantitative signal suggests, what qualitative research adds, and which principle should govern the tradeoff. Test with users poorly represented in the average before turning a local improvement into a universal default. The review should preserve what was known when the choice was made, document what changed, and share the learning beyond one team. Repeated with discipline, that cycle converts a memorable sentence into better judgment rather than another short-lived management theme. It also makes the cost of assumptions visible, allowing future teams to reuse proven knowledge while recognizing the point at which a different customer, risk, or operating environment requires fresh testing.