One team, not two halves

Two weeks ago I argued that holding back on AI is often rational, not fearful — because autonomy devalues the expertise people spent a career building.

This chapter is about the answer to that. And the best evidence I know for it isn't a survey. It's a controlled experiment.

758 BCG consultants. Two groups. One with AI, one without.

Dell'Acqua, Mollick, Lakhani and colleagues measured what happened when consultants worked within the technology's capability frontier: +25% speed, +12% task completion, +40% quality.

And outside that frontier? They were 19 percentage points more likely to be wrong.

Same technology. Same people. Two entirely different outcomes. The variable was where the boundary sat — and whether anyone knew where it sat.

That's the finding I'd underline for anyone rolling out AI right now. The question is not "is the model good". It's "does the person using it know what this thing is good at, and what it only looks good at".

Ethan Mollick calls the working version of this Co-Intelligence. Human and machine are not tool and user. They're a collaboration with clearly different strengths — and the value shows up in the handover between them, not in either half alone.

Which is also the honest answer to last chapter's problem. In a real alliance, deep expertise isn't devalued. It becomes the thing that tells you where the frontier is. AI makes people better — not obsolete — but only where the work is built for two halves.

Two halves. One whole. That's when it works. 🧩

💬 Where have you seen human and machine truly work as an alliance — not as two separate systems?