Blog
What does AI cost in practice? Where the money really goes
· 4 min read
You ask for a quote for a new website and get a fixed amount back. You ask the same for an AI application and get a more nuanced answer: it depends on usage. For a business owner used to a fixed bill, that feels evasive. It is not evasive, it is how AI works. In this post we explain where the money goes in practice, without naming a figure we could not live up to anyway.
Why a fixed amount does not exist here
A licence for standard software is usually bought per user or per year: a fixed number, regardless of how intensively you use the system. AI models work differently. Large language models are billed per unit of usage, per token, not per subscription. OpenAI explicitly prices its API per million tokens, separately for each model category. The more text an application processes, the higher the bill, even if nothing in the configuration changes.
You see the same pattern, in a different form, with low-code platforms. OutSystems determines licence costs partly on the basis of the scale and complexity of the apps you build, measured in the number of screens, database tables and API methods, and on the number of users who actually use the app. This means that a proof of concept with a handful of test users and a few screens sits at a different scale from that same application once all employees use it daily and functionality is added. A supplier who promises you a hard total for an AI project in advance is therefore really making a promise about your future usage and the size of the app, which they cannot yet know.
That is no reason to leave AI aside. It is, however, a reason to budget differently: not as a one-off amount, but as a cost model that moves with what the application actually does.
The costs you do not see coming
The bill for the model itself is usually not the largest item. What is often underestimated is the work around it.
First there is integration work. An AI agent that reads invoices or summarises emails has to be connected to the systems where that data already lives: a mailbox, a CRM, an Excel file that has been the source of truth for years. That requires mapping out properly in advance which screens, roles and integrations are involved, and deciding for each part whether it goes into the first version. Skip that step and you discover halfway through the build that an integration is missing and still has to be built, with all the delay that entails.
Then there is testing, and that takes more time than with traditional software. An AI application that sometimes gives a wrong answer does not fail with an error message, but with a plausible-sounding yet wrong result. That is why it is sensible to prove the riskiest part first, separately, before you build the rest. That costs time up front, but it prevents a setback from surfacing only in the final week before go-live, when changing course is expensive and stressful.
And then there is the human in the loop: someone who checks the output before it goes out into the business. That is not a temporary measure until the model gets better; it is a structural cost item for as long as an AI system makes decisions that have an impact. We wrote in more detail earlier about how to set up that checkpoint in human in the loop for AI automation. The time someone spends on it belongs in the budget, not in the margin.
Maintenance does not stop at go-live
With AI, an application that works is not finished. Model providers change prices and model versions, usage patterns change as soon as employees trust the application and start using it more, and output that is good today may turn out differently tomorrow because an underlying model has been updated.
If you want to keep that under control, measure the usage and costs of AI-assisted work per build day or per period, instead of discovering only in the annual accounts that the bill has crept up. That way you know where time and money actually go, and you can adjust in time instead of being surprised afterwards.
It also pays to keep warnings and error messages in the build and test environment at zero every round, instead of letting them pile up. A stack of ignored warnings makes a new, real error invisible, and you then only discover that error at a customer or user, which is always more expensive than catching it beforehand. In addition, briefly record at the end of each work period what is done and what is still open: that prevents a following session from losing time working out where someone left off, and that time is also a cost item, even though it never appears separately on an invoice.
What does this mean for your business?
A few concrete steps to get a grip on AI costs without buying a false sense of certainty:
Do not ask for one total price for the whole project; ask for a cost model: what does a period of building cost, what does usage cost at a realistic volume, and how does that scale as the number of users grows.
Prove the riskiest part first, small and separate, before you have the rest built. That gives you an early, realistic picture of what works and what takes extra time, rather than only at go-live.
Include the human in the loop as a fixed item in the budget from the start, not as something you add later when things go wrong.
Measure usage and time from the first build day, per component, so that you adjust based on facts rather than on a feeling that the bill is suddenly higher than expected.
Would you like to know what such a cost model looks like for your situation, with which steps and which risks? Get in touch with us, and together we will look at where the money would actually go in your case.