Artificial intelligence (AI) arrived promising to cut costs and save time. But many companies open their invoice at the end of the month and get a shock: the AI bill grew, it didn't shrink. The reason has a technical name, token, and a practical effect nobody explained when you signed the contract: you pay for a lot of work the AI does behind the scenes that you never see on screen.
This post explains, in business terms, why the bill goes up, why it goes up even faster when you use AI agents, and what separates the company that controls this cost from the company that only discovers the damage after it's already done. At the end, a checklist to demand from your vendor.
What is a "token cost" and why is it invisible?
A token is the unit the AI uses to charge you. Think of it like the "pulse" in old telephone billing: every piece of text that comes in (your question, the documents, the company rules) and every piece that goes out (the answer) gets sliced into tokens, and you pay for each slice. The more the AI reads and writes to handle a request, the more tokens it uses, and the higher the bill.
The detail that catches everyone off guard: most of that consumption is invisible. When a customer asks "where's my order?", the AI doesn't answer in one shot. It consults the history, reads the internal rules, checks the status in the system, drafts a response, reviews its own draft, and only then displays the final sentence. You see a two-line reply. Internally, the AI read and wrote the equivalent of several pages. All that internal conversation gets billed.
It's like hiring a lawyer by the hour and only looking at the one-page petition they delivered. The petition is short, but you pay for the hours of file reading, case law research, and revision that nobody saw happen. With AI it's the same: what appears on screen is the tip of the billing iceberg.
Why does the bill go up if AI prices are falling?
This is the paradox that confuses managers. The price per token has been dropping fast. Gartner projects that the unit cost of AI could fall by as much as 90% by 2030, and that models will be up to 100 times more efficient than 2022 models. So why is the invoice growing?
Because three things move the bill at the same time, and only one of them is falling:
- Price per token: falling. Good news.
- Volume: rising. The more AI proves it works, the more people in the company use it, and the more it processes.
- Frequency: rising. AI stopped answering one question at a time and started working in cycles, repeating steps until it reaches a result.
The price drops 10%, usage triples, and the bill goes up. It's the same logic as a utility bill: electricity got cheaper per kilowatt, but you installed air conditioning in every room. The cost reduction is real, it just can't keep pace with how fast the company starts depending on the tool.
AI agents change the math (for the worse, if you don't control them)
There is a massive difference between a chatbot that answers questions and an AI agent, which is a system that executes tasks on its own, from start to finish. The chatbot reads and responds. The agent reads, decides, acts, checks the result, and repeats the cycle until the job is done. That autonomy is exactly what makes the agent valuable. It's also what makes the bill spike.
The numbers are clear: according to Gartner, an AI agent consumes between 5 and 30 times more tokens than a regular chatbot for the same task area. This isn't deployment overhead, it's the nature of the work. The agent "talks to itself" dozens or hundreds of times before delivering, and every one of those loops gets billed.
The risk few vendors disclose: a poorly configured agent can get stuck in a loop, keep retrying something that doesn't work, and burn through budget in minutes. Worse: you pay for the final correct answer and for every failed attempt that came before it. It's like paying a delivery driver for the successful delivery and also for every wrong turn they took. Without guardrails, an agent spends like a taxi sitting in traffic with the meter running.
The most expensive mistake: using the most intelligent model for the most trivial task
This is the number-one waste we find in companies. There's an entire family of AI models, from the simplest and cheapest to the most powerful and costly. The reflex of someone just starting out is to use the most powerful model for everything, "just to be safe." Result: you use a cutting-edge, very expensive model for a task that a simple model would solve at a fraction of the price.
Classifying whether an email arrived "angry" or "satisfied" is a cheap-model job. Deciding the renegotiation strategy for a complex debt is a robust-model job. Using the expensive model for both is like hiring a surgeon to apply a bandage: it works, but you're paying a fortune for something a nurse would handle. The golden rule isn't "use the most intelligent AI," it's use the cheapest AI that solves each specific problem.
That explains an uncomfortable figure from McKinsey: 9 out of 10 companies already use AI, but only 39% can tie AI to a real profit gain, and when that gain exists, it typically represents less than 5% of results. High adoption, low return. A big part of that gap is money burned on wasted tokens.
Cost governance isn't a spreadsheet, it's architecture
The wrong reaction is to treat AI cost as a finance problem, where someone looks at the invoice after the fact and complains. By the time the expense shows up in the report, the money is already gone. Controlling token cost isn't a job for whoever checks the invoice at the end of the month; it's a decision that needs to be embedded in the agent design from day one.
In practice, an agent with real cost governance has four things that a "slapped-together" agent doesn't:
- Right model for each task: the system automatically chooses the cheap model for simple jobs and the expensive one only for complex ones, with no waste.
- Spending cap and kill switch: a budget ceiling and automatic shutdown if the agent enters a loop, so the bill never spikes without warning.
- Reuse: information already retrieved once isn't re-read and re-billed every time, which cuts repeated consumption.
- Cost visible by task, client, and department: you see how much each interaction cost, not a single lump sum at the end of the month that nobody can trace.
Notice that none of this is "buy cheaper AI." It's architecture. It's deciding, at build time, how the money will be spent. A company that skips this step doesn't save money, it just delays the shock.
How Steply controls the hidden cost of tokens
Steply builds custom AI agents for each company's operations, and we treat token cost as part of the project, not as an end-of-month surprise. In practice, that means building the agent with intelligent routing between models (cheap for volume, powerful for exceptions), with a spending cap and automatic loop-protection, and with a dashboard that shows cost by task, by client, and by department.
The effect for your business is predictability. You stop having to choose between "use AI" and "control costs," because the control is built into the tool. Instead of discovering the damage later, you track spending in real time, know exactly where tokens are going, and decide with numbers in hand when to scale up, when to cut back, and when to change course. It's the difference between driving with your eyes on the dashboard and driving blind and hoping the tank doesn't run out.
If your company already uses AI and the bill keeps growing without a clear explanation, or if you're thinking about hiring an agent and want to avoid the hidden-cost trap, that's exactly the problem Steply solves before it shows up on the invoice.
Frequently asked questions about AI token costs
What is a token in artificial intelligence?
A token is the smallest unit the AI uses to process and bill text. Every piece of a word, command, or document excerpt becomes one or more tokens. You pay for tokens that come in (your question plus the context) and for tokens that go out (the answer). The more the AI reads and writes to handle a request, the more tokens it consumes and the higher the bill.
Why does the AI bill increase if model prices are falling?
Because the price per token drops, but usage volume and processing frequency grow faster. AI now works in cycles, repeating steps to complete a task, and more departments in the company start using it. The price reduction is real, but it gets swallowed by the increase in total consumption.
Why do AI agents cost more than chatbots?
Because the agent doesn't respond just once. It queries data, decides, acts, checks the result, and repeats the cycle until the task is done. According to Gartner, this consumes between 5 and 30 times more tokens than a regular chatbot. Every intermediate step is billed, even if you only see the final answer on screen.
How do you reduce AI token costs in a company?
The biggest gains come from four fronts: using the cheapest model that handles each task (instead of the most expensive for everything), setting spending caps and automatic loop-protection, reusing already-processed information, and measuring cost per task and per client. The control must be embedded in the agent build, not just in invoice review.
What is the risk of an AI agent with no cost controls?
A poorly configured agent can get stuck in a loop and burn through the budget in minutes, repeating attempts that don't work. Since billing is per use, you pay for the final correct answer and for every failed attempt. Without a spending cap and automatic shutdown, the bill spikes without warning.
Is it worth using the most powerful AI model for everything?
No. Simple tasks, like classifying the tone of an email, are handled by cheap models. Reserving the powerful model only for complex decisions cuts costs without losing quality. Using the most expensive model for everything is the most common waste and the biggest driver of inflated AI bills.