Your team asks AI things all day: draft this email, summarize this document, write a reply to this customer. Every one of those answers has a price. It is small, it is invisible, and it adds up faster than anyone expects.
That price is counted in tokens. If you understand tokens, your AI bill stops being a mystery. This guide gets you there in five minutes, no technical background needed.
A token is a piece of a word
AI does not read whole sentences the way we do. It breaks text into small chunks called tokens. A token is roughly three-quarters of a word in English, so 100 words is about 130 tokens. Every word you send to AI gets counted, and every word it writes back gets counted too.
You pay for both directions: what goes in, and what comes out. Paste a 20-page contract and ask for a one-line summary? You paid mostly for the 20 pages going in.
Numbers count too, and so do formatting, code and punctuation. A spreadsheet pasted into a chat is tokens. The polite "thank you" at the end of a conversation is tokens. Small ones - but the meter never sleeps.
Think of it like your electricity meter
You do not think about kilowatts when you turn on a light, but the meter counts them and the bill arrives monthly. Tokens work the same way. Nobody on your team sees the meter while they work - it just runs. That is exactly why AI costs creep up quietly: the meter is real, but nobody is watching it.
One difference makes it worse. Your building has one electricity meter. AI gives every person their own meter, on every tool, and those readings never gather in one place by themselves. Reading them together is the whole trick.
Why some tasks cost 100x more than others
Three things drive the price of any single AI task:
- How much text goes in and out. A short question with a short answer is cheap. Long documents, long conversations, and long reports are not.
- Which model does the work. AI comes in sizes. The biggest models can cost 10-30x more per token than the small ones. Same question, very different price.
- Conversation history. In a long chat, the tools resend earlier messages with every new question so the AI keeps context. Message 40 in a conversation quietly costs many times more than message 2.
Let's price a month, using real numbers
Take the number from the top of this page: one everyday task - "draft a quote for this client" - runs about 1,240 tokens, roughly $0.04. Now build a company month out of it.
- One person: say a dozen small tasks a day. That is about $0.48 a day, or roughly $10 a month.
- A 50-person company: about $500 a month - if every task stayed small and ran on a right-sized model.
- The same workload, done badly: route everything through the premium model at 10-30x the per-token price, let conversations run long so history is resent with every message, and that $500 of work can bill like $5,000-15,000.
Same questions. Same answers. A 10-30x difference in what the meter counts. That gap - not the $0.04 - is what your invoice is made of. Our guide to matching the right model to every task digs into exactly this choice.
And this is before waste enters the picture. Apply the 30-40% waste finding to whatever your company actually spends, and the takeaway is blunt: roughly a third of a typical AI bill is habits, not work.
How this shows up on your bill
Most companies pay for AI in one of two ways, and often both at once:
- Seats. A fixed monthly price per person (ChatGPT Business, Copilot, Claude for Work). Predictable, but you pay whether the seat is used or not.
- Usage. Pay-per-use, through a platform or add-on. Scales with real activity, which means it also scales with waste.
Either way, the number on the invoice is the sum of thousands of small token counts nobody looked at. Industry research finds that 30-40% of enterprise token budgets are waste - not because AI is a scam, but because nobody was watching the meter.
How to read your own AI bill this week
You can get from mystery to rough map in a few hours, without any tooling:
- List every AI line item. Ask finance for anything in the last three months that looks like an AI vendor - subscriptions, pay-per-use platforms, tools with AI add-ons. Most companies find more lines than they expected.
- Split seats from usage. For each line, note whether it bills per person or per token. They fail differently: seats waste money by sitting idle, usage wastes money by running hot.
- Check seat activity. Admin pages show who logged in. Across companies, 10-20% of paid seats typically sit idle, and trimming them is usually worth 5-15% of the bill. We wrote up the details in the hidden cost of idle seats.
- Ask the model question. One message to your team: "which model do you use by default?" If the answer is "the best one, always", the 10-30x gap from earlier is quietly sitting on your invoice.
- Spot the token hogs. Look for the long-running chats and the huge pasted documents. One or two habits usually explain a surprising share of the total.
"Tokens are cheap - why should I care about $0.04?"
Fair pushback. A single task really is pocket change, and treating every prompt like a taxi meter would be a miserable way to work. Nobody should do that.
But the bill is not one $0.04 task. It is thousands of them, times the number of people, times the model multiplier, times resent conversation history - every working day. Small numbers with multipliers attached stop being small. That is how 79% of companies ended up overshooting their AI budgets: not one big purchase, just compounding pocket change nobody added up.
The answer is not to make people count tokens. It is to have something count for them - which is what AI cost management means in practice: the meter gets read automatically, and humans only look when a number moves.
What to do with this knowledge
- Find out what you actually spend per month, across all tools. Most companies cannot answer this today.
- Ask which model your team uses for everyday work. If everything runs on the premium one, you are overpaying.
- Watch for long chat sessions and huge pasted documents - they are the token hogs.
- Or skip the manual work: Optimize connects to your AI accounts and does all of this automatically.
If you want the fast version of that last point: connecting is about 15 minutes of work, read-only, and the first report arrives within 48 hours - a map of spend per tool, per team and per person that this guide has hopefully made readable. And if you want the cautionary tale of what happens when nobody reads the meter at all, see the startup that found out from the invoice.