282,000 prompts is not a result

Companies have finally built the one structure that genuine AI capability building needs: a network of respected people, embedded in the work, one per team, helping colleagues get better. It is the best distribution mechanism for capability an organisation will ever have. And almost everyone has pointed it at the wrong target.

The Wall Street Journal ran a piece this week on the "AI superfans" companies are mobilising to win over reluctant colleagues. Read it for the numbers, because the numbers tell you exactly what these networks are being used for.

Law firm Ropes & Gray celebrates its champion programme taking the legal AI tool Harvey from about 32 users and a few hundred prompts a month to 2,200 employees and more than 282,000 prompts a month. Partner Howard Glazer sums up the mission: "Humans don't like to change." Citigroup ran its Citi Stylus champion effort with more than 4,000 participants and watched adoption climb from single digits to about 80%. Mars Snacking has 500 champions; its CTO Ramesh Kollepara says "Champions are the ones who socialize the broader potential of AI."

Now notice what every one of these stories celebrates. Prompts per month. Adoption percentages. Seats active. Usage metrics, all of them. The champion's job, in every case, is framed the same way: convert the skeptic. Sell the tool.

There is a reason the selling is needed. For two years, executives told their people AI would replace them, then acted surprised when those people didn't rush to feed it. Even the CEOs are now softening the fear message. The champion networks are partly cleaning up an adoption problem that top-down pressure created in the first place.

A prompt count is an input

Here is the problem with 282,000 prompts a month: it measures nothing that matters. A prompt count is an input, not a result. An adoption percentage tells you people opened the tool. It does not tell you a single workflow got better, a single handoff got cleaner, or anyone's judgment got sharper. You can hit 80% adoption while every process in the building runs exactly as it did before, just with a chat window open next to it.

And that is what makes the current setup so wasteful. Look at what these companies have actually built. Trusted peers. Embedded in the daily work, not parachuted in from L&D. One per team, at Mars 500 of them. This is precisely the vehicle capability building needs, the thing no training programme and no licence rollout can buy. Spending it on adoption cheerleading is like hiring master teachers and grading them on attendance.

Worse: the skeptic the champion is sent to convert is often right. The tool doesn't fit the current workflow. That is a true statement about most tools and most workflows. The honest response is not a better pitch. It is a different question: what in this process should change, now that this tool exists? Which step disappears, which handoff moves, which check becomes someone else's job? That question redesigns work. "Have you tried prompting it differently?" does not.

Same people. Same network. Same trust. Different question. That is the entire distance between an adoption programme and a capability programme, and it costs nothing to cross.

One thing to do this week

Change the single thing your champions report back on.

• Retire the old question: how many people used it, how many prompts this month. Stop collecting it from champions entirely. If you still want it, pull it from the admin dashboard, where it belongs.
• Replace it with this: bring one workflow your team actually redesigned, or one task the team has stopped doing the old way, because of the tool. One per champion, per month. Described in a sentence, not a slide.

Some champions will come back empty-handed the first month. Good. That is real information about where capability isn't building, which is more than a prompt count ever told you. One changed handoff beats a thousand logged prompts.