Are your AI costs rising too? Then you're probably consuming more than you think.
The price per token has been falling for years, and yet the bill keeps climbing. That's not an error in your accounting, but a direct consequence of how AI applications work. This page explains where consumption leaks away and which levers you can pull.
Why do AI costs rise while tokens get cheaper?
Because consumption grows faster than the price falls. A language model remembers nothing between calls, so all instructions, retrieved documents and earlier steps are sent again every time. A twenty-step task pays for that context twenty times. Control comes from making consumption visible, pruning context and routing work to the lightest suitable model.
- Consumption grows faster than the price per token falls
- Repeated context is the biggest hidden cost driver
- Output and reasoning tokens weigh more than input
- Steer on cost per task, not on the monthly total
The paradox: cheaper per token, more expensive per month
Anyone following the rates of the major model providers has watched them fall for years. Anyone following their own AI spend has watched it rise. Those two lines don't contradict each other: they measure different things. One measures the price of a unit, the other the number of units you consume.
And that number is growing fast, because every improvement you make to an AI application costs tokens. You give the model better instructions, let it search your own documents first, give it access to your systems, have it check its own work. Each of those is a sensible step. Together they mean that one user question becomes a series of calls under the hood, each carrying a larger package of context along with it.
On top of that, consumption behaves differently from a licence. A subscription is predictable: you know what it does per month. Consumption responds to behaviour — to a busy week, to a team discovering a new workflow, to a process failing more often than expected. Without visibility, you only find out afterwards what happened.
The good news: this is almost never a technically unsolvable problem. In practically every AI application we open up, a substantial share of the consumption sits in work nobody misses once it's gone.
Where consumption leaks away
Six mechanisms explain most of a rising AI bill. They rarely show up in a dashboard, yet they sit in nearly every application.
01
Context that travels along at every step
A language model remembers nothing between two calls. Everything it needs to know, you send again each time. An assistant that handles a task in twenty steps therefore ships that accumulated history twenty times. It isn't the first question that's expensive, it's the twentieth.
02
Instructions and tool definitions that always ride along
The system instruction, the behavioural rules and the descriptions of every system the AI may reach: those blocks are usually fixed, yet they travel with every single call. Invisible in your interface, heavy on your bill.
03
Search results that arrive too broadly
When a solution first searches your own documents, the temptation is to retrieve generously: more fragments, larger chunks, extra overlap. That lowers the odds of a missed answer, but you pay for every extra fragment on every single question.
04
Output weighs more than input
What the model writes counts several times heavier than what you put in, at virtually every provider. Models that 'think out loud' before answering also produce intermediate steps you never see but do pay for. An instruction like 'be thorough' is therefore a cost decision.
05
Repeats, retries and loops
A failed call that is automatically retried, an agent circling because one step won't complete, a user rephrasing three times: none of that is logged as an 'error'. It disappears into the total and quietly grows along with it.
06
The heaviest model as the default
During the build you pick the strongest model, because then everything works. That choice tends to stay in place afterwards, including for the simple routing work: categorising an email, summarising a form, filling a field. Heavy tools for light work is the most expensive habit there is.
Six levers you can pull
You don't have to choose between cheap and good. In practice these six interventions reinforce one another — and together they deliver far more than any of them alone.
Route by difficulty
Not every question deserves the heaviest model. Let a light model handle the routine work and escalate only when the task genuinely calls for it, or when the light model itself signals it can't cope. This is usually the largest saving and the least visible change for users.
Reuse what doesn't change anyway
Fixed instructions, standard context and documentation can be reused instead of paid for from scratch every time. The condition is that you structure your request so the fixed part comes first and the variable part last. An afternoon of structural work, with an effect on every call after it.
Queue the work that can wait
Not everything has to answer within a second. Overnight enrichment of your CRM, cleaning up a document archive, reporting: that work can wait. Providers reward it with a markedly lower rate per processed batch.
Prune what you send along
Summarise long conversations instead of dragging them along in full. Retrieve less but more precisely from your own sources. Drop tool descriptions the AI never uses in this process anyway. Less context often means a better answer too: models lose sharpness in an overcrowded window.
Put limits in the software, not in a memo
A maximum number of steps per task, a ceiling per team, an alert when consumption deviates from the pattern. Limits that live in the software work; limits that live only in policy are discovered at the final invoice.
Measure per task, not per month
One monthly total tells you nothing. Cost per handled request, per processed document or per resolved ticket tells you everything: which application pays for itself, which one is running away, and where scaling up makes sense.
SME or enterprise: the gain sits in a different place
The mechanism is the same, the biggest saving is not. Start where your organisation is leaving the most on the table.
SME
Clean up what you already have
- Licences you pay for people who barely open the tool.
- Three assistants side by side doing largely the same work.
- Automations consuming in the background without an owner.
- One person keeping an eye on consumption, instead of nobody.
Usually a matter of sessions, not months. The gain is in the overview: knowing what you buy, who uses it and what it returns.
Enterprise
Steer on what is spread out
- Attributing consumption to teams and applications instead of one collective line item.
- Ceilings and alerting in the software, not only in policy.
- A cost check before a new application goes into production.
- Agreements on which model is the default and when you escalate.
Here the individual intervention isn't the problem — the problem is that nobody sees the whole. Control appears the moment every team sees its own consumption and is held to it.
The question isn't "what does it cost", but "what does one task cost"
A monthly figure says little. It doesn't tell you which application pays for itself and which one is quietly burning money. The moment you trace consumption back to one handled request, one processed document or one answered customer question, the conversation becomes concrete.
Because then you can make the only comparison that really matters: what does this task cost with AI, and what does the same task cost without? If that comparison works out well, a rising bill isn't a problem but a sign that you're scaling. If it works out badly, you know exactly which application to address — not that "AI is expensive".
That calculation is also why we don't work with a price list. The value of an AI solution lies in what a task costs relative to what it returns, and that differs per process and per organisation. What does hold everywhere: if you don't measure that ratio, you're steering on instinct.
Read more about the investment in a project on what AI implementation costs. How agents work and why they consume so much is explained in the knowledge base under AI agent and RAG. Still at the start? Have a look at getting started with AI.
How we bring your token costs under control
Not a major clean-up operation, but four steps that make your consumption visible and keep it that way.
-
Make visible where it goes
first stepWe break your consumption down by application, team and task. Without that picture every saving is a guess. More often than not, the bulk of the bill turns out to come from a handful of processes.
-
The easy money first
short lead timeRedundant context, unused connections, repeated work that could be reused, and the too-heavy model on light work. Interventions users don't notice and that you see back immediately.
-
Guard quality at every change
ongoingCheaper must never mean worse. We define upfront what a good answer looks like, measure that with each adjustment and roll back anything that hurts quality. Saving without a yardstick is postponed trouble.
-
Lock it in so it doesn't return
structuralCeilings, alerting and a fixed cost check for every new application. That way control becomes a property of your organisation instead of a project you did once.
Curious where your consumption leaks away?
The AI Readiness Scan maps which applications you run, what they deliver and where you lack control. A concrete starting point for a conversation about cost.
Take the AI Readiness Scan
Frequently asked questions about token costs
What exactly are tokens?
A token is a piece of text: often a word, sometimes part of a word or a punctuation mark. Language models don't charge per question but per token, counting both what you put in and what comes out. The moment you use AI inside a process rather than in a chat window, that count becomes the number that determines your bill.
Why is my AI bill rising while rates are falling?
Because your consumption grows faster than the price per token drops. Applications get smarter by sending more context: longer instructions, retrieved documents, conversation history, intermediate steps. Every improvement costs tokens. You only feel the cheaper unit price once you also steer on the number of units you consume.
Why are AI agents so much more expensive than a chatbot?
A chat question is one round trip. An agent works in steps: plan, look something up, consult a system, verify, adjust. At every step the entire accumulated context travels along again, including all intermediate results. Analysts therefore see agentic tasks consuming a multiple of a single chat question — the range varies widely by task, and it is exactly that unpredictability that makes limits important.
Is this relevant for a smaller company too?
Yes, but the attention goes elsewhere. In smaller organisations the money usually sits in licences nobody uses, in three assistants side by side doing largely the same work, and in automations consuming in the background without anyone watching. That is often cleared up in a few sessions, whereas larger organisations are more likely to face consumption spread across many teams.
Will quality drop if I steer on cost?
Not if you do it carefully. Pruning context often sharpens answers, because the model has less noise to weigh. The risk lies in blindly switching to a lighter model. That's why we first define what a good answer is, test every change against it, and roll back whatever doesn't hold up.
When does running your own model on your own infrastructure make sense?
Only once your volume is predictable and large, or when your data absolutely cannot leave your walls. Below that threshold you mainly pay for people and maintenance rather than for compute. For most organisations, using existing models more intelligently is a faster and safer route than self-hosting.
How do I stop it getting out of hand again in six months?
By putting control in the software rather than in agreements. Ceilings per application, alerting on deviating consumption, and a fixed cost check before a new application goes live. Anyone who keeps tracking cost per handled task sees a problem coming instead of discovering it on the invoice.
Is your AI bill climbing faster than you expected?
Book a no-obligation conversation. We look at what you run, where the consumption comes from and which interventions will have the fastest effect in your situation.