Most AI systems don't fail. They work. They hit metrics. They ship. They get used. And then the impact stalls.

The problem is not the model. It's the system.

AI optimises whatever signal you give it. Most companies get that signal wrong. The result: activity without impact, and value that never compounds.

This keynote shows what needs to change: how to select the right problem, define the right signal, and design systems that improve in the right direction over time.

We assume that when AI gets something wrong, a human will catch it. But humans review systems at scale, under time pressure, in environments where the system is usually right. When that happens, people stop questioning it.

Human-in-the-loop is not a safety mechanism. It’s a design decision. And most organisations are designing it under conditions where it cannot work.

This keynote shows why oversight fails in practice; and what that reveals about how AI systems actually behave.

Using AI is no longer a differentiation factor. Your moat is the way a system behaves in real customer interactions: the decisions it makes, the edge cases it handles, what it learns from every exchange.

Yet, most customer-facing AI is built to pass evaluations. Very little is built for how real customers use it.

The gap between those two things is where competitive advantage either compounds or collapses.

This talk gives you a lens for identifying where your current AI initiatives are optimising the wrong thing. And what building for real interaction actually requires.