Writing in Fortune, Julio M. Ottino and Brian Uzzi, both professors at Northwestern University, examine the risks of pushing AI into executive decision-making. Ottino is a professor at Northwestern’s McCormick School, and Uzzi teaches leadership and organizational change at the Kellogg School of Management. They open with Mark Zuckerberg’s use of a Meta CEO agent to get answers that once required staff intermediaries, and they use the example to separate two kinds of friction inside organizations.
The authors distinguish coordination friction, meaning the delays of scheduling, approvals, and communication, from cognitive friction, meaning the resistance that comes from differing perspectives and assumptions. AI removes the first kind efficiently, they write, and they see that as a genuine benefit. Their concern is that it may also erase the second, which they treat as useful rather than wasteful, because disagreement is often what stops a decision from being wrong.
Their central warning is that an AI system can become what they call a sophisticated mirror, reinforcing a leader’s existing thinking instead of testing it. A tool tuned to be helpful and agreeable, in their view, tends to confirm what a chief executive already believes. To counter that tendency, Ottino and Uzzi argue companies should deploy AI as a sparring partner that generates counterarguments, rather than as a peacemaker that smooths contradictions away.
They cite Meta’s Project OT, an effort to build an AI-native company with sharply reduced headcount, which they say abandoned its most aggressive reduction targets. Internal findings, in their telling, suggested that gains from AI-assisted coding did not translate into comparable improvements in products that users actually see, a gap they read as evidence that removing human friction has limits and that output is not the same as impact.
The authors close with an analogy to science, where progress depends on methodological variation across many laboratories rather than a single optimized approach. Fully integrated end-to-end systems, they argue, optimize existing assumptions and grow fragile at exactly the moment an organization needs to ask whether it is solving the right problem. Their conclusion is that a measure of friction, kept on purpose, protects that questioning, and that leaders should design their AI tools to preserve it rather than engineer it away.
