Quick answer: Johnson & Johnson gave its people three years to experiment freely with generative AI — "let a thousand flowers bloom." Nearly 900 projects bloomed, but only 10–15% produced about 80% of the value, so CIO Jim Swanson switched most of them off. The focused survivors cut lead-optimization time in drug development in half and shrank a regulatory report from roughly 700 hours to about 15 minutes. MIT's 2025 report "The GenAI Divide" put the broader pattern starkly: about 95% of corporate AI pilots showed no measurable profit impact. So "switch off the rest" isn't pessimism — it's hygiene.
The shallow lesson: "run more experiments, do more AI"
Motion mistaken for progress. Almost nobody claps for the person with the courage to turn off the initiative that doesn't work.
The real lesson: the AI skill of 2026 is subtraction
Not adding more — finding what actually works, reinforcing it, and having the discipline to kill the rest. The value isn't in the 900 flowers; it's in the nerve to prune 850 of them.
The sobering part
Those ~900 pilots weren't wasted money — they were tuition. But tuition only pays off if you act on what the experiments taught you. The failure mode isn't experimenting; it's refusing to switch anything off once the data is in.
Full disclosure
Azati lived this. We spent about 30 months building an AI-first "Revenue OS" for our own company — and were moving the wrong way for roughly 24 of them before it clicked. That's ~80% of the time in the wrong direction, almost the same ratio as J&J. So we're sympathetic to the math, and biased toward killing what doesn't work. See our GenAI & product work.
FAQ
Adapted from a LinkedIn post by Andrew Babkin, Managing Director & CTO at Azati – read the original. Sources: Reuters Momentum AI / Wall Street Journal reporting, April 2026; MIT, "The GenAI Divide," 2025.
