Last month, I calculated the cost of an AI application using OpenAI that appeared to cost 2 cents per document. A bargain, I thought. Then I added oversight, corrections and integration. It was no longer a bargain.
That is not a calculation error. It is the bill nobody adds up.
The price you get from OpenAI, Anthropic or your software supplier is the easiest number in the entire story. It is on an invoice, you can look it up and you can compare it. That is exactly why everyone focuses on it. And it is exactly why the business case is so often wrong: the number that is easiest to find is usually the smallest part of what AI really costs you.
What total cost of ownership actually means
Total cost of ownership, or TCO, is an old concept from IT. The idea is to count not just the purchase price, but everything a system costs throughout its lifetime. For a car, that means not only the purchase price but also insurance, maintenance, fuel and depreciation. AI works the same way, except the hidden items are larger and harder to see.
I like to divide the costs into two piles. One pile appears on an invoice. The other does not.
The invoice contains the tokens, subscriptions and API costs. That is the visible pile. It is relatively honest, although how it is distributed across teams and applications is often a mystery.
The invisible pile is larger. It includes integration: someone has to connect the AI to your existing systems, and that connection has to be maintained. It includes data work: preparing, cleaning and making documents accessible. It includes human oversight: someone checking the output before it reaches a customer. It includes remediation: what does it cost if a model makes something up and that error carries through into a customer letter? And at the very bottom is your dependency on the supplier, which only acquires a price when something goes wrong.
The cost everyone forgets: oversight
Of all those hidden costs, one is consistently underestimated: human oversight.
This is the item I notice most often because it runs directly counter to the promise used to sell AI. The promise is that AI takes over work. In practice, AI relocates work, and the relocated work is often less visible and therefore less likely to be accounted for.
The argument I often hear is that AI takes over the work, so the person can go. In practice, work shifts more often than it disappears. Someone still has to assess the output, especially when it informs a customer interaction or a decision. Checking takes time, and that time does not appear on any AI invoice.
I see this clearly in insurance. Take a model that pre-sorts claims. On paper, it saves manual assessment. But with claims, oversight is precisely what you must not cut, because a wrong decision affects a customer and sometimes the regulator. So you keep a human in the loop. AI makes the work faster, not free. Anyone who bases the business case only on token costs is comparing half the cost side with the entire benefit side. The fact that AI is not always the answer is exactly what happens when this calculation turns out differently than you hoped.
Why the invoice itself keeps moving
Even the visible pile is less fixed than it seems. Most AI services charge by usage, and usage is not constant.
The price per token is falling rapidly. According to the Stanford AI Index 2025 (opens in new window), the price for GPT-3.5-level performance fell from roughly 20 dollars per million tokens in November 2022 to 0.07 dollars per million tokens in October 2024. That is a decline of more than 280 times in just under two years. If you look only at that figure, AI appears to be becoming almost free.
Yet the bill is rising at most organisations. The reason is an old pattern: when something becomes cheaper, you use so much more of it that your total spending rises. This is the Jevons paradox I described earlier. A cheap model invites longer prompts, more context and agents that repeat themselves. The unit price falls by a factor of ten while consumption increases by a factor of one hundred. For your TCO, this means you cannot calculate the cost side once and call it done. Usage changes every quarter.
Four places where the hidden bill lives
When I calculate an AI business case, I review four items that are almost never included in the first version.
1. Integration and maintenance
An AI tool in isolation does very little. The value emerges when it is connected to your systems, and that connection is not a one-off job. Models change, APIs change and your systems change. Someone has to keep up. Include that cost, even when it looks like a "quick job".
2. Data work
AI is only as good as the data you put into it. Preparing, cleaning and making documents accessible is work, and it is work you often underestimate because it is tedious. In many projects, this is the largest hidden cost, larger than the model itself.
3. Oversight and remediation
Covered in the previous section, but it belongs on every list: the time required to check output, plus the cost when an error slips through. For anything that affects a customer, this is not optional.
Follow me on LinkedIn
4. Dependency on your supplier
The most expensive item appears on no invoice: what happens if your supplier raises its price or changes its model? Many tools are a layer on top of an OpenAI or Anthropic model, and you often do not discover whether you are buying a wrapper until the price rises. That dependency is a cost, even if you cannot yet express it in euros. I explored this further in my piece about vendor lock-in and the kill switch.
And the cost even the supplier does not measure
One hidden cost runs so deep that the supplier itself does not publish it: energy consumption. To understand the sustainability side of your AI bill, you would want to know how much energy a model uses. That is almost impossible to determine because the major providers deliberately keep their distance from independent measurements. Until the EU AI Act enforces that transparency, this remains a blind spot in every complete TCO calculation. I mention it because an honest bill also names the items you cannot yet fill in.
A calculation that shows the difference
Let me make this concrete with an illustration. The numbers below are illustrative, not drawn from research, but their proportions match what I see in practice.
Imagine an application that processes documents. The invoice says 2 cents per document, and you process ten thousand per month. That is 200 euros. On the invoice. Done, you might think.
Now for the rest. Building and maintaining the integration takes a few days each quarter from someone with the right skills. The data work needed to make the documents readable took weeks upfront and keeps returning. And for ten per cent of the documents, the output is not entirely correct, so someone checks and corrects it. Allow a few hours per week for that oversight. Those hours do not appear on any AI invoice, but they are real, and over a year they easily exceed the 2,400 euros in token costs.
The lesson is not that the application is bad. It may be perfectly good, even with the full bill. The lesson is that you only see the full bill when you create it, and that the 200 euros on the invoice gives you a false sense of cheapness. I am convinced that most disappointing AI business cases are not failed technology. They are good applications assessed against an incorrect and far too low cost estimate.
How I calculate an AI business case
No complicated model. Three rules already remove most of the hot air.
First, count both piles. Put the invoiced costs and the hidden costs side by side before you name a benefit figure. If the hidden pile remains empty, you have not looked carefully enough.
Second, calculate with growing usage, not today's usage. An application that now costs 200 euros per month can cost many times more after a broader rollout because more people will use it more often. Who uses what determines your bill, and that depends on which users you have in your organisation.
Third, give the calculation an expiry date. AI TCO is not a sum you do once. Repeat it every quarter because prices, models and usage all move.
Why suppliers keep the invoice low
There is a reason the visible costs look so low. Suppliers compete on the number you can compare most easily, which is the price per token or per seat. They want to keep that number low because it helps them win the comparison. The costs they leave with you, integration, oversight and data work, do not appear on their price list because they are your costs, not theirs.
That is not malicious intent; it is how the market works. But it means you can never derive an honest TCO from a price list. The price list is optimised to look cheap. Your bill only emerges when you add up what the supplier deliberately leaves outside its number. Anyone who fails to do that compares apples with part of an apple, then chooses the provider best at looking inexpensive rather than the one that is actually inexpensive.
Why this is a board-level question, not an IT detail
You might think TCO belongs in procurement or IT. With AI, it belongs at the board table for a simple reason: hidden costs affect your margin, and dependency affects your continuity. Those are not technical issues. They are governance issues.
The broader picture, from controlling usage to measuring returns, is in my guide to AI costs and ROI. This post covers one part of that picture: adding up the bill honestly before approving it.
I began with an application that appeared to cost 2 cents per document. Once I counted everything, it was no longer 2 cents. That is not an argument against AI. It is an argument for counting the entire bill, because the cheapest application on the invoice is not always the cheapest one in your organisation.
Sources
* Decline in the price per token for GPT-3.5-level performance (from 20 to 0.07 dollars per million tokens): hai.stanford.edu (opens in new window) * Why so many AI projects fail to achieve measurable results (MIT NANDA, via Fortune): fortune.com (opens in new window)