Google launched a new AI model this week at a quarter dollar per million tokens. Let me say that again: a quarter dollar. Gemini 3.1 Flash-Lite costs so little that most companies wouldn’t even notice it on their bank statement.
Two years ago, the same thing cost a hundred times more. The price curve for AI models looks like the energy market in reverse – it falls and falls and falls. And yet, there’s a paradox.
Danish companies expect to increase their AI investments by 94 percent next year, according to Lenovo’s CIO Playbook 2025. At the same time, almost half lack clear policies for how they should even handle AI. Nearly one in three CIOs haven’t defined KPIs for their AI projects.
So they’re spending more money without knowing what they’re spending it on. That’s not a technology problem. It’s a leadership problem.
The hardware is cheap. The strategy is expensive.
It’s easy to look at token prices and conclude that AI has become cheap. It has – the part that involves running a model. But that’s a bit like saying fuel is cheap and concluding it’s cheap to run a taxi company. Fuel is a fraction of the bill.
The real cost of AI sits somewhere else entirely. It sits in figuring out what’s actually worth automating. In having data that’s clean enough for a model to use. In building integrations to the systems you already have. And in getting people to change the way they work.
McKinsey reported in 2025 that 88 percent of companies fail with their AI initiatives. Not because the models don’t work. But because they start in the wrong place, lack the right data, or never make it from pilot to production.
Three places the money goes
The first is data work. AI is only as good as the data it can access, and most companies have data sitting in silos, in formats no one has touched in years, or not digitized at all. Getting data ready for AI typically costs more than the AI solution itself.
The second is integration. An AI model living in its own little bubble is a demo, not a solution. It needs to talk to your ERP, your CRM, your customer service system. That requires data integration – and it’s rarely plug-and-play.
The third is change management. Computerworld recently cited an example where an organization spent months developing an HR chatbot that almost no one used afterwards. Not because the technology was bad, but because no one had ensured that employees understood why they should use it.
What it actually costs
If you’re a mid-sized company and want AI that does something real, the math typically looks like this: the pure AI costs – tokens, API calls, hosting – are the smallest line item. The biggest cost is figuring out what to build, and building it right.
That’s not a reason to hold back. It’s a reason to start with strategy instead of technology. Don’t ask “where can we use AI?” Ask “what’s our most expensive problem, and can AI solve it?”
The difference is that the first question gives you a pilot that never becomes anything. The second gives you a business case with a number at the bottom.
