Early on, I did what every technology leader does. I demonstrated. Look what it can do. The room would lean in, someone would say “wow”, and I’d leave feeling the cause had been advanced.
It hadn’t. I eventually noticed that the wow almost never converted into changed behaviour. People were impressed the way they’re impressed by a magician. Entertained, briefly amazed, and completely unchanged, because none of it connected to the work sitting in their inbox.
Worse, demos quietly send the wrong messages. A polished demo says this is advanced stuff, performed by experts, on curated examples. It positions AI as spectacle rather than a tool. Some people left my demos more convinced than ever that this wasn’t for them.
So I stopped. No more impressive demos. What replaced them was much less glamorous: sitting with people and their actual work.
Bring the report you’re dreading. Bring the spreadsheet that eats your Friday. Bring the policy you have to summarise for your team. Then we’d work on it together, with AI, live. Sometimes the result was brilliant. Sometimes it was mediocre and we’d talk about why, and what that teaches you about where these tools can be trusted, where they still need to improve.
The mediocre sessions turned out to matter most. Watching AI fail on your own work, and learning to catch the failure, is precisely the judgement we need people to build. In critical infrastructure, a workforce that can spot a confident wrong answer is worth far more than a workforce that has seen a perfect demo.
The wow moments still happen, but they’ve changed character. It’s not “wow, look what it can do”. It’s “wow, I just got my afternoon back”. The first kind spreads applause. The second kind spreads adoption, because people tell each other.
If your AI program is still built around showcases and a degree of smoke and mirrors, ask what behaviour has actually changed. Then try swapping one demo for an hour with someone’s real work.