What You Are Actually Paying For When You Buy AI

By John O'Connell Article, Peaks Perspective AI, Compliance, Future of Fintech, Risk Management, Technology, WealthTech

You signed a contract for an AI tool. The bill arrives measured in something called tokens, or credits, or messages. Last month it cost one number. This month it costs twice that. Nobody on your leadership team can tell you why.

You do not need to understand the engineering to fix this. You need to understand the meter. AI is now a line in your operating budget, and it does not behave like the per-seat software you have bought for the last twenty years. The firms that learn how the meter runs will scale AI with confidence. The firms that treat it like a flat subscription will be surprised by the bill and slow to commit because of it.

What is a Token?

A token is a fragment of text. It runs about four characters, or roughly three-quarters of a word.¹

A short, common word like “trust” or “fees” is a single token. A longer phrase costs more. “Wealth management” runs about four tokens, and “financial planning” about the same, because each word breaks into smaller pieces. As a rule of thumb, 100 tokens come out to roughly 75 words, so a short client email runs about 200 tokens and a page of a quarterly report about 500.²

The AI charges you for text going in and text coming out.³ Text going in is your question plus anything you attach to it. Text coming out is the answer it writes back. You pay on both ends of every exchange. Hold onto that. It is where the cost hides.

The Asymmetry that Breaks Budgets

Here is the part your CFO will miss. Input and output are not priced the same. With Anthropic’s Claude models, output costs five times more than input.⁴ The direction of the work matters more than the volume of it.

Picture two tasks. In the first, you hand the AI a 40-page document and ask for a 2-page summary. The machine reads a great deal and writes very little. In the second, you give it a 1-page prompt and ask it to generate a 10-page document. The machine reads very little and writes a great deal.

The second task costs more. Often far more. And it costs more even though the first task puts more total paper on your desk. You are paying for what the machine produces, not for what it consumes. Reading is cheap. Writing is expensive.

This has a direct consequence for how you plan. Any workflow built around generation runs the meter hard. Examples include drafting policies, producing client reports, writing market commentary. Any workflow built around summarizing or extracting runs it light. The same tool can be cheap or expensive depending entirely on which direction you point it.

Why the Real Workflow Costs More than the Demo

The hidden multiplier is context.

In a back-and-forth session, the AI does not remember your last message for free. The entire prior conversation often gets sent again, and charged again, on every new turn. Turn ten is not one message. It is the whole transcript processed ten times over. Attach the same client file to every question in a session and you pay for that file every time.

This is why a tool that looked cheap in a pilot becomes a budget problem at scale. It is also why agentic automation, the capability vendors sell hardest, is the most expensive pattern you can run. An agent that works through a multi-step task re-reads its own context at every step. The meter runs the entire time.

How Vendors Repackage the Meter

You will rarely see raw tokens on an invoice. Microsoft bills Copilot Studio usage in Copilot Credits at one cent per credit.⁵ Other vendors bundle usage into messages or bury it inside a seat tier. The currency changes. The logic underneath does not. A more complex task that re-reads context and calls outside systems burns more of whatever unit you are buying.

Watch the words “unlimited” and “included.” They carry conditions. Microsoft includes a 25,000-credit monthly allowance with each Microsoft 365 Copilot license, but that allowance does not roll over, and once it is gone, unlicensed users stop getting answers unless you have attached a pay-as-you-go meter behind it.⁶ Read what the allowance actually covers before you treat it as free.

The Questions You Should Be Asking

Stop asking what the tool costs per user. That is the wrong question for a technology priced by consumption. Ask three better questions:

What does one representative task cost, from the question you ask to the answer you receive?
What happens when you exceed the included allowance, throttling or overage billing?
Which of your intended uses are long-output or agentic, and therefore the expensive ones?

The firm that asks these questions controls its AI spend. The firm that does not will discover the answer on an invoice.

Variable Cost, Fixed Discipline

This is a variable operating cost. Treat it like cloud compute or postage, not like a software license. You do not need to count tokens. You need to know which of your workflows runs the meter hardest, and you need to budget before you deploy, not after.

The technology is not the hard part. Understanding what it costs to use the technology is important. The discipline to educate your teams on what it costs and building a budget is hard. Build it now, while your AI spend is still small enough to study. It will not stay small for long.