Human Magic in the Loop

Ask a board whether humans oversee its AI systems and the answer is yes. There is a review step. Someone signs off. The governance slide is complete.

Then look at where that person sits. They are at the end, approving. By the time a recommendation reaches them the question has been framed, the options generated, the trade-offs weighed. Nothing is left to shape. And when agent systems produce thousands of consequential decisions an hour, no human can meaningfully review them at all.

A human who rubber-stamps is deferring, not heeding. The distinction comes from Tsai and Ku, who separate deferring to a machine from heeding it, which means comprehending the output, understanding its reasoning, and then deciding. Deferring replaces first-person judgment. Heeding preserves it. A workflow containing a person who defers has acquired a procedural check that adds neither comprehension nor judgment. Oversight can be fully staffed and still be empty.

The evidence points upstream. In the largest study of multi-agent failures so far, more than forty percent traced to how the systems were designed rather than to the models themselves. The machines are not mainly failing at execution. They are failing at what they were asked to do.

Human Magic in the Loop is the design principle that follows. Put people where the questions are still open, at framing, and at the moments when values conflict and someone has to decide what matters. Split the work by consequence rather than by speed, and give the machine what is reversible. Name who owns the decision before the agents go live.

The practical question a board can ask is narrower than it sounds. Where in our systems does a human still change an outcome, and does that person understand what they are approving? Two organizations can run the same tools, report the same productivity, and answer that question in opposite ways.

The argument runs through Human Magic: Leading with Wisdom in an Age of Algorithms (Routledge, 2026), which traces how the capabilities that make professionals valuable erode when algorithms substitute for them and grow when they work alongside them. Johan and Bart Victor named the design principle in a 2026 paper, arguing that where agent operation cannot carry practical wisdom on its own, that wisdom lives in the humans the system is engineered to require.

This is a design principle rather than a recipe. What it looks like in a given organization depends on the work, the stakes, and the failure modes worth guarding against. That is the conversation Johan has with leadership teams.