The slide has survived three CEOs.

A stick figure. A chip. An arrow to a bigger output. Human plus AI beats human alone. That part is true.

The slide skips the only comparison that would tell you whether the human is adding anything on top of what the AI already does by itself.

The CPO who had never run the control

I was on a coaching call last week with a CPO at a European property insurer. She shared her screen and walked me through the all-hands deck for a $14M copilot rollout. The augmentation slide was on page six.

I asked a dumb question. Has your team ever measured the copilot working alongside your underwriters against the copilot running the same claim on its own?

She paused. Then she said she had no idea. Why would they?

That line stayed with me all week. It is the question most collaboration decks politely skip.

The pooled version of my question

Vaccaro, Almaatouq and Malone at Harvard Business School and MIT Sloan posted the full meta-analysis on arXiv on 30 October 2026. 420 papers. 1,955 effect sizes on humans and AI working on real tasks.

Human plus AI against human alone: pooled Hedges' g of 0.53, confidence interval 0.45 to 0.61, p < 0.001. A medium effect. The slide is right on that axis.

Human plus AI against the better of either side alone: g = -0.35, confidence interval -0.52 to -0.18, p < 0.001. The subset is 171 experiments that bothered to also run AI by itself. A third of a standard deviation worse than the stronger partner working on its own.

Across the 171 experiments that ran the control the standard slide implies but never shows, the human in the loop drags down the better of either alone by g = -0.35.

The same team ran this analysis in 2024 and got g = -0.23. Two more years of data, mostly generative AI, made the picture worse. The decision-tasks moderator (p < 0.001) says the AI-over-human gap is largest in decision work. That is exactly the setting where most firms bolt a human in as a check on the stronger partner.

The CPO's team runs claim decisions through the copilot and then back through an adjuster. They have never measured the adjuster against the copilot. The pooled evidence suggests the adjuster is probably costing them about a third of a standard deviation of accuracy.

Three operational reads fall out of the moderators. Division of labour is the only process variable that moves both estimates (p = 0.005 on augmentation, p = 0.031 on the better-of-either test); specify who does which step cleanly. AI explanations are a measured performance lever (p = 0.002); design the explanation surface instead of signing off on it. Crowdworkers get less lift than non-crowdworkers (p < 0.001); the "AI levels the playing field" story runs the other way in the pooled evidence.

Run the control on your own rollout before you run the rollout again. If the AI alone beats the pair, you have been paying for drag.