A manager at a company I work with pulled up a dashboard to show me how well the AI rollout was going. Prompts run this month. Tokens used. Seats active. The numbers were up and to the right, and he was proud of them.
I asked what it had actually changed. Which decision got made faster. Which report stopped taking a full day. Which euro showed up that wasn't there before.
The room went quiet.
That silence is the whole problem with how most companies are doing AI right now.
Activity is not the outcome
We are measuring the wrong thing. Tokens burned, prompts run, "AI adoption." None of that is a result. It's motion. A company can burn a fortune in tokens and be exactly as effective as it was a year ago, just louder about it.
I produced electronic music for years, and there's a beginner mistake everyone makes: you turn everything up. More compression, more layers, more loudness. It feels like progress because the meters are moving. Then someone with real ears plays you a mix that's quieter and clearer and hits ten times harder, and you realize you were measuring the wrong thing the whole time.
AI in a business is the same. The token counter climbing is the meters moving. It is not the mix.
The token-burn trap
Here's how the trap works. You buy the seats. Everyone starts prompting. Usage climbs, so you feel good. But most of that usage is people asking AI to write emails they'd have written anyway, or generating slop that a human then has to clean up. You've added a step and called it transformation.
The version pointed at the customer is worse. Every SaaS wants an AI feature on the homepage now: a chatbot, a "generate" button, an assistant nobody asked for. Front-facing AI as decoration. It burns tokens on every interaction and rarely moves a number that matters. You can feel it as a user. You know when you're talking to a wall.
I don't think that's where the value is. I think it's almost the opposite.
Put the AI in the plumbing, not the shop window
The work that actually pays is unglamorous and invisible. It lives in the infrastructure.
When I set up AI for managers at Politiken, we didn't build anything a reader would ever see. We sat down and did two boring things. We named the five recurring tasks each leader actually does every week, the ones that eat their Monday. And we named the red zones, the things AI must never touch (source protection, personnel data, a contract with a counterparty).
Then we made AI good at the five, and fenced off the rest. That's it. No demo, no chatbot, no launch. Just a leader getting an hour of their week back and trusting the thing, because they know exactly where its hands aren't allowed.
That's the forward-deployed way of thinking. The model is a commodity now, everyone has the same one, so the whole advantage is in the deployment: you go into the real workflow, find where the money and the time actually leak, and you build against that specific leak.
What to measure instead
If tokens are the wrong number, here are the ones I'd put on the dashboard.
Revenue you can trace to it. Not "AI-influenced," a real line: a recovered customer, a proposal that closed faster, a campaign that shipped in a day instead of a week.
Time back on a named task. Pick the task before you start. "The weekly board summary went from three hours to twenty minutes." That's a number a CFO believes.
Decisions made faster, with the same or better judgment. Speed only counts if the quality holds.
And one nobody tracks: what you chose not to automate. Governance is a metric. "Claude drafts, never sends" is a result, because it's the reason people trust the system enough to actually use it.
None of these have anything to do with how many tokens you spent getting there.
I burned tokens too
I don't have this fully solved, so I'll be honest. I've done the thing I'm describing. I've spun up workflows that felt clever and moved nothing. I killed three of them last month because they failed exactly this test. The instinct to measure activity is strong, because activity is easy to see and revenue is hard to attribute.
The fix I keep coming back to is the gardener's fix. When a plant gets sick, the instinct is to add: more water, more feed, more attention. Usually wrong. Usually you have to cut, and let the energy go back to what's healthy. Most AI rollouts need less, aimed better.
The part you can't get from a blog post
Everything above is the what, and I'm giving it away on purpose. The frame, the governance move, the five-tasks-and-red-zones exercise, take all of it.
The hard part was never the information. The hard part is doing it inside your actual business, with your actual leaks, your actual red zones, your actual people who are quietly afraid the AI is going to embarrass them. That part doesn't come from reading. It comes from sitting in the room and building it against your specific mess.
That's the only thing I'd ever charge for. The reading is free. The implementation is the job.
I'm still learning the edges of this. But I'm fairly sure the companies that win the next few years won't be the ones who burned the most tokens. They'll be the ones who can tell you, to the euro, what the burning was for.