Put your best analyst next to a capable AI and give them the same job. You would expect the pair to beat either one working solo. They don’t. More often than not, the team does worse than the stronger of the two would have done alone.
That is not a rumor. It is a systematic review, published this year in Sustainability.
In fifty-eight percent of the cases examined, human and machine together did worse than the better of them would have done apart. The tool made the team worse.
The authors don’t stop at the diagnosis. They set out twelve design principles to fix the pairing, and almost none of them are new — most were learned with earlier machines and quietly forgotten the moment the next one arrived.
Make the system show what it doesn’t know. Have people write down their own judgment before they see the model’s. Keep a running, visible list of where the AI has already been wrong. Protect regular stretches of work without it, so the skill doesn’t wither. Read that list as a leader and the instinct is immediate: hand it to the implementation team and the fifty-eight percent takes care of itself.
Here is what that instinct misses.
Every one of those principles asks something of a person that only trust can buy.
Writing down your own view before the model answers takes effort, and people spend effort where they believe it counts. In an organisation where nobody reads what you flag, you stop flagging. Keeping the error log honest means saying out loud that the shiny new tool leadership just bought is wrong a third of the time — and people only say that where saying it is safe. Protected time away from the tool lands as development in one culture and as surveillance in another. The difference is never the schedule. It is whether the people using it trust the organisation that set it.
You can wire a house exactly to code. Every circuit correct, every switch where the drawing says it should be. Leave the main breaker off and nothing in the building lights up.
The twelve principles are the wiring. Trust is the current. Get the wiring right and leave the power off, and you have a beautifully specified system that does nothing at all.
This is the part the design conversation keeps stepping around. Complementarity gets treated as an engineering problem — calibrate the trust in the model, force the engagement, protect the human skill. All sound. All downstream of a prior question no workflow can answer: do these people trust the organisation that dropped the tool into their day?
We have watched this exact sequence for three decades, long before the tool was AI. Across 1,072 change leaders in 80 countries, one variable moved resistance more than any other, and it was not communication, training, or sponsorship. It was trust in the people leading the change. Every other lever gets multiplied by that one. Where trust is missing the multiplier is zero, and the finest change plan ever written returns nothing.
The machine changed. The finding did not.
So the brief for any leader putting AI into an organisation is not the one the vendor hands you. It is older and harder. Before the tool goes in, the people who will use it have to believe their judgment still matters, that flagging a problem is safe, that the organisation asking them to work a new way has earned the right to ask. That belief is not a setting. You cannot configure it during the rollout.
You build it first, in the ordinary work, before the tool ever arrives — or you do not, and you join the fifty-eight percent.
Serve the people first and the twelve principles do exactly what the research says they will. Skip that, deploy on schedule, and you have wired the house and left the power off.
The tool was never the variable.