You've probably seen the stat: 95 percent of enterprise generative-AI pilots deliver no measurable impact on the P&L. It gets passed around as proof that AI is overhyped. I read it almost the opposite way.
If 95 percent of pilots flop and the other 5 percent are using the same models, the model isn't the variable. Something about how the work gets rolled out is. And the place that breaks down most predictably is training.
What the 95% really means
The same research that produced that number found something more useful buried underneath it. Mid-market teams were getting from pilot to production in around 90 days, while large enterprises took nine months or more. And projects done with an outside partner succeeded at roughly twice the rate of internal-only builds.
95%
of enterprise GenAI pilots show no measurable P&L impact
MIT, State of AI in Business 2025
2x
success rate for projects done with an external partner vs. internal-only
MIT, 2025
60%
of employees say hands-on learning would most increase their AI use
IBM, 2025
Read those together and a picture forms. The winners move fast, stay small, and get outside help. The thing they're not doing is gathering a hundred people in a room to watch a feature tour.
Why the generic version fails
I've sat through the generic AI training. You probably have too. Someone demos a handful of features, the room nods along, everyone agrees it's impressive, and then Monday comes and nobody does anything differently. The session wasn't wrong. It was just disconnected from the actual work.
A controller doesn't leave a generic session knowing how AI helps with month-end. A dispatcher doesn't leave knowing how it helps with scheduling. They leave with a vague sense that AI is powerful and no idea where to put their hands. That gap is where adoption dies.
The four moves that work
Train by department
Finance, field ops, and sales use AI for completely different things. Run separate sessions so every example on screen is relevant to the people in the room.
Survey first
Send a short survey before the session asking what people actually do all day and where they lose time. Now you're teaching to their real bottlenecks, not a generic agenda.
Bring real data
Put the team's own data in front of them. A finance team should leave having run an actual month-end task, not having watched someone else's demo.
Offer a 1:1 follow-up
Group sessions get people started; a short optional one-on-one gets the stragglers over the line on their specific workflow. That's where the real conversions happen.
Set a baseline first
One unglamorous thing that derails more sessions than people admit: basic computer fluency. We've had people join an AI workshop who couldn't share their screen. There's no shame in that, but it means the first ten minutes evaporate before you've taught anything.
Set a simple baseline before anyone joins — can they share a screen, find a file, copy and paste between windows. Five minutes of prep saves the whole session from stalling out of the gate.
Small teams, outside help
The research lines up with what we see in the field. Smaller groups convert faster because the session can stay specific. And outside help converts faster because someone who has run this twenty times can skip the dead ends and point straight at the workflow that'll pay off first.
None of this is exotic. It's just the difference between teaching people about AI and teaching people to use AI on the work in front of them. If your last training didn't change anyone's Monday, change the format before you blame the tool.
Key Takeaways
- The 95% pilot failure rate is mostly a rollout problem, not a technology problem.
- Generic, feature-tour training rarely changes behavior — it's disconnected from the real work.
- Train by department, survey people first, bring their real data, and offer a 1:1 follow-up.
- Set a basic computer-fluency baseline before the session so it doesn't stall in the first ten minutes.
- Smaller teams and outside help both correlate with faster, more successful adoption.