−19pp
Experts get fooled too.
On work past AI's edge, seasoned pros using it were 19 points less likely to be right, and didn't clock it.
BCG × Harvard · 758 consultants
E /The evidence
Not opinion. Published research, caveats and all.
The method is built on published work about how people and AI actually perform together. The headline findings are below. Want the full detail, the sources, and every caveat? It's all in the brief.
01 /The headlines
Six findings that shape what we do.
+43%
It lifts the floor first.
The biggest gains went to the lowest performers. AI raises the floor of work more than the ceiling.
BCG × Harvard · 758 consultants
+17pts
A good reference does the work.
Checking against an independent reference beats unguided judgement, and helps your weakest reviewer most.
RefEval, 2026
666
Lean on it and you dull.
Across 666 people, heavy AI use tracked with weaker critical thinking. Skill and method buffer it.
Gerlich, 2025
18models
How you prompt decides it.
Same model, different prompt, measurably different quality. The operator matters as much as the tool.
Serapio-García et al., 2025
2skills
Knowing the subject isn't knowing the tool.
Domain depth and AI-failure literacy are separate skills. We test and train both.
Jagged-frontier evidence, 2023
02 /Kept honest
We show the caveats, not just the wins.
Last reviewed: June 2026.
- Each finding in the brief carries its own limit. We'd rather under-claim than oversell.
- The page has a named owner, a review date, and a refresh cycle.
- Spot a finding we've overstated, or research that complicates one? Tell us. That's the correction loop the method runs on.
03 /Next step
Want this walked through for your board?
Take the deck into your next governance conversation, or book a short discovery call. We reply within one working day.
Prefer email? Use sales@aielabs.co.uk.
Download the deck (PDF) Book a discovery call
Prefer slides you can edit? Download the PowerPoint (PPTX).