For most of the last two years, I priced AI the way I’ve priced every piece of software in thirty years of building technology: a number per seat, budgeted once, revisited once a year. It’s the only pricing model most IT leaders have ever had to plan around.
Then I started building AI systems that ran at real usage instead of pilot usage, and that instinct didn’t survive contact with an actual bill. A workflow that ran a handful of times a day during testing runs hundreds of times a day once a team adopts it. The per-token price never moved. The bill did, and it climbed in multiples, not percentages.
I’d made the mistake almost everyone makes. Software has a seat price. You buy access, use it as much or as little as you want, and the invoice barely notices. AI doesn’t work that way. Every query, retrieval and step an agent takes to finish a task consumes tokens, and tokens are metered like electricity, not sold like a subscription.
I learned how AI economics works by watching the shift from the inside. I was building a system that combines retrieval, workflow automation and human review to support a marketing team’s daily work. The finance conversations that came with moving from pilot to daily use are what made it click.
Here’s the disconnect that blindsides budgets. Most IT leaders know roughly what they’re paying per API call or per seat. Fewer know what they’re actually spending, because those two numbers move independently.
The mental model that fails first
Seat-based pricing trained a generation of IT leaders to treat software cost as fixed. Add ten users, the license line moves in ten predictable increments. Budget it once a year and move on.
AI breaks that model. Two employees on the same seat can generate wildly different bills depending on what they ask the system to do. Someone summarizing a short email uses a fraction of the tokens someone running a multi-step research task across several documents does. It’s the same license, but completely different cost.
I’ve watched this happen more times than I can count. The deeper a task goes, the more tokens it burns, and that’s true whether the task is trivial or genuinely valuable. Cost tracks the depth of the work, not a judgment about who’s using the tool well.
Nobody explains that part upfront. Adoption is supposed to be the win. In token economics, adoption is also the thing that drives the bill up.
And here’s the twist that makes it counterintuitive: per-token prices have fallen fast. Stanford’s 2025 AI Index Report found that inference cost for a system performing at GPT-3.5’s level dropped more than 280-fold between November 2022 and October 2024. Look only at the price sheet and you’d swear AI got cheaper. Then look at what teams are actually doing about it. The FinOps Foundation’s State of FinOps 2026 report found that 98 percent of FinOps teams now manage AI spend, up from 31 percent two years earlier, and they named it their top forward-looking priority for the year. That kind of urgency doesn’t gather around technology that’s getting easier to forecast.
Where the surprises hit
The surprises show up in three specific places. I’ve had a direct hand in all three while scaling systems from pilot to production.
The first is agentic workflows. A simple prompt and response might use a few thousand tokens. An agent that plans a task, retrieves documents, calls a tool, checks its own output, and retries when something looks off can burn ten times that for a single request. Every one of those steps gets billed. When I moved a workflow from single-shot generation to a multi-step process with retrieval and review built in, token use per task jumped in a way the original cost model never accounted for. Nothing was broken. The system was just doing more, and more is exactly what usage-based pricing charges you for.
The second is context growth. Retrieval-augmented generation pulls source material into every query so the model has something accurate to work from. The more documents you retrieve, the more context you feed in, and every token of that context gets billed on top of the question itself. A well-tuned RAG system retrieves exactly what’s needed. A loose one retrieves everything that might be relevant. The difference between those two shows up on the invoice.
The third is the success penalty, and it’s the one that catches leaders off guard, because it looks like good news until the invoice says otherwise. A pilot gets approved on light usage and a small budget. It works. People like it. Usage climbs faster than anyone modeled, because that’s what adoption looks like when a tool is genuinely useful. Most cost models don’t leave room for growth that fast while the project is still labeled a pilot.
What cost governance actually looks like
None of this means AI adoption should slow down. It means cost has to become an architecture decision instead of a financial afterthought that surfaces once the system is already in production. Four disciplines carry most of the weight.
Route models by task complexity. Not every task needs a frontier model. Routine summarization, formatting, and basic classification run fine on smaller, cheaper models with almost no quality loss, which frees the frontier model for the work that genuinely needs that level of reasoning. This single change has bent my cost curve more than anything else I’ve tried.
Monitor at the workflow level, not the account level. Your total monthly AI spend tells you almost nothing. Knowing that one specific workflow accounts for sixty percent of it tells you exactly where to look. Granular tracking built in early is the difference between explaining a cost increase to leadership in one sentence and spending a week on forensics.
Treat agentic pipelines like any production system that can run away from you. That means retry limits, timeout logic and a hard stop when a task loops longer than expected. A recent piece on why most agentic AI projects stall before they scale makes a related point: governance, not the model, becomes the real constraint once these systems move from demo to production. Cost control is a direct extension of that same discipline. Skip the circuit breaker and the risk isn’t just a bad output. It’s an uncapped bill.
Plan for price change instead of betting against it. Current pricing is still partly subsidized by vendors competing for market share, and that won’t hold. GitHub’s move to usage-based billing for Copilot this year previews where the rest of the market is going. Build your cost model on the assumption that per-unit pricing normalizes upward, not that it stays this generous.
These surprises don’t come from a vendor overcharging you or a system malfunctioning. They come from applying a twenty-year-old cost model to a technology that was never built to work that way.
You don’t need to fear usage-based AI pricing. You need to stop budgeting for it like a subscription. Getting ahead of this doesn’t mean spending less on AI. It means knowing what each dollar bought before finance asks, and fixing the architecture instead of the budget.
The number at the bottom of the invoice matters less than whether you can explain every line above it.