AI tokens: what an agent consumes and how to predict it

AI plans are not counted in hours or number of emails, but in tokens. The word may seem abstract. Yet it describes a simple reality, and understanding it avoids two costly mistakes: paying for a token pool that is too large out of caution, or running out of a token pool that is too small by the second week. This course explains what a token is, what consumes them, and how to forecast.

IWhat exactly is a token?

Take the word "facturation". The AI does not see it as a whole: it breaks it into two or three pieces, depending on the model. Models have been trained mostly on English, so a French text is often broken into more tokens than it has words. Remember a prudent rule: count a little more than one token per French word, sometimes much more with proper nouns, product references or numbers.

IIReading also consumes tokens

This is the point that almost everyone forgets. When an agent reads your pricing grid, each piece of that grid is counted. When they read a received email, too. When they write their response, too. Reading consumes tokens, not just writing. Some models "think" additionally before answering, and this internal reflection also counts.

Direct consequence: an agent who writes three lines can consume a lot if they had to read fifty pages to write them. The length of the response tells you almost nothing about the cost. What matters is everything the agent has in front of them when working.

IIIOrders of magnitude: an email, a page, a long PDF

The exact figures vary depending on the model and language. Here are cautious orders of magnitude, for reasoning and not for billing.

These values add up with each passage: the mission, the useful documents, the day's items (received emails, overdue invoices), then what the agent writes. If you want just one idea to remember: it is the large documents reread with each passage that make the bill, much more than the emails sent.

IVThe four token pools of ABYS

On ABYS, each plan gives a token tank for the month. This tank is shared by all agents on the account: from 19 euros, you can have multiple agents working, and it is the amount of tokens that limits the work, not the number of agents. The account displays consumption throughout the month. A yearly payment option exists, with a discount.

VHow frequency changes consumption

The calculation fits in one line: monthly consumption = cost of one passage × number of passages. The cost of a passage depends on what the agent reads and writes. The number of passages depends on your plan's rhythm. An agent that runs every day makes about thirty passages per month; an agent that runs once a week makes four or five. An email received by the agent adds a passage.

VIConsume less without working less well

Before taking a larger reservoir, check if the agent is not reading too much. The levers are simple and do not degrade the work.

If, after these adjustments, the reservoir is still too tight, switch to the next plan. I advise you to decide based on one or two weeks of actual consumption, not an estimate made on the first day.