What are AI tokens, and how do they become a monthly bill?

What it is, in plain words

Think of a translation agency that charges by the page, both for the pages you send in and for the pages it sends back, with the pages it writes priced higher than the pages it reads.

AI models work the same way, except the unit is a token instead of a page. A token is a small piece of text: a short word, part of a longer word, a punctuation mark. Every time your automation or chatbot sends text to the model, the provider counts the tokens going in (input) and the tokens coming out (output) and charges for both. On the provider price list I checked, prices are quoted per million tokens, with input and output listed separately and output priced higher.

Two consequences follow:

• You pay for what the model reads, not only for what you type. Instructions, documents and earlier messages sent along with a question all count as input.
• Cost scales with use, not with the build. A chatbot that costs a few cents a day in testing can cost much more once real customers use it every day.

A concrete example

Take a small support chatbot on a shop’s website. A customer asks, “When will my order arrive?”

What the model actually receives is much more than that one sentence:

1. The bot’s standing instructions: tone, rules, what never to say.
2. The pieces of your FAQ or policy documents that the bot looked up for this question.
3. The conversation so far, because the model does not remember it between messages.
4. The customer’s question.

All four are input tokens. The answer is output tokens. On the next message, the conversation so far is sent again, now one exchange longer. That is why long conversations cost more per message than short ones, and why a bot that pastes whole documents into every request costs more than one that sends only the relevant paragraph.

Usage also has a ceiling you set yourself. A nightly content job of mine, which uses an AI model to write social posts, hit the monthly spending limit on the account it ran on. It produced four posts and then ended with an error. A second job that would have used the same account stayed switched off, so the two would not share one limit.

Why it matters when you hire someone

A freelancer can build a chatbot or an AI step cheaply and still hand you something expensive to run. Three things go wrong when the buyer does not know how tokens work:

The bill lands on the wrong account. If the bot runs on the freelancer’s AI key, you cannot see the cost, and the bot stops when their account does. The key should be yours. The API keys explainer covers who should hold it.

Nobody measured the cost per conversation. Testing with ten questions says nothing about a month of real traffic. The acceptance line on the chatbot job page asks you to know the cost per 100 conversations for this reason.

The limit becomes an outage. A spending limit protects your card, but when it is reached, requests fail. For a customer-facing bot that means silence; for an overnight job it can mean half the work and an error nobody reads. Decide who gets alerted before the limit is reached.

The design choices that move the token bill (how much document text goes in with each question, how long answers are, which model is used for which step) are the builder’s. The questions below make them visible. What moves the price of a freelance automation covers the build side of the same decision.

What to ask the freelancer

A good answer uses numbers from your own test runs, not from a pricing page, and names the person who gets the alert.