The real AI cost, beyond the sticker price

Ask what AI costs and most people quote the sticker price, and the sticker price is the easy, small part. As of mid-2026, a per-seat assistant runs roughly $30 a user a month for Microsoft 365 Copilot, around $60 and up for ChatGPT Enterprise, while Google folded Gemini into Workspace plans, and those numbers move every few months. One thing to flag: Copilot is an add-on that stacks on a base plan, so the true figure is often two to three times the number you first see quoted. The other model, the metered one that shows up when you build on AI, is token pricing: you pay per million tokens, split into input for the prompt and output for the response, and output usually runs three to five times the input rate. Both of those are real AI costs, and both are the small ones. The expensive part of AI cost never appears on the order form.

That's the part worth understanding, because it's where budgets actually go, and it's the part you can do something about.

Of all those, the idle seat is the largest and the most invisible. You pay the full per-seat price every month whether or not anyone uses it, and a lot of seats sit unused. McKinsey found leaders estimate only about 4% of employees use generative AI for a meaningful share of their day, while the real figure is roughly three times that, and only about 1% of companies call themselves AI mature. BCG's data shows regular frontline use stuck around half, with only 36% of employees feeling adequately trained. Read those together and the picture is a lot of paid licenses that never become habits, which pushes the effective cost per active user well above the list price. The idle seat isn't a tooling problem; it's an adoption problem, and it's the most expensive line item nobody puts on a purchase order.

There's a real counterargument: AI is getting radically cheaper, so cost barely matters, just buy the cheapest model and wait for prices to fall. It's well supported. Stanford's AI Index found the cost to query a capable model fell from around $20 to roughly $0.07 per million tokens in about eighteen months, close to a 280-fold drop, with roughly tenfold annual declines for equivalent performance. That collapse is genuine, but notice which number it is: the token and inference cost, the line item that's already small and falling on its own. The costs that actually blow AI budgets aren't the model; they're people and adoption, the idle seats, the rework, the integration, the roughly 42% of AI initiatives that get abandoned before production. Those don't fall with token pricing. Waiting for a cheaper model doesn't fix any of them. The lever that moves cost-effectiveness is getting the people you already pay to actually use what you already bought.

If the biggest AI costs are idle seats and unused capability, then the cheapest way to improve your return isn't a procurement negotiation; it's adoption. Training the people you already pay, on the tools you already bought, converts dormant licenses into the productivity you were promised, and it's a fraction of the spend already sitting in those seats. That's true across every role, and it scales as a team effort rather than a per-person one. Look at the full AI cost, sticker plus hidden, as part of your AI transformation, and the math points the same way every time: the model is the cheap part, and the people are where the money is won or lost. What it costs to train a team is small next to what idle seats already cost you.