AI pricing is unusually hard to reason about, and the confusion is not entirely accidental. The headline number is rarely the number you end up paying, and the gap tends to appear in month three rather than month one.
Here is where the money actually goes.
Usage pricing, and why it surprises people
Most AI services bill by volume of text processed, in units called tokens. A token is roughly three quarters of a word. You pay for what goes in and what comes back.
The trap is that costs scale with usage in a way software licences do not. Ten documents a day is trivial. Ten thousand is a different conversation. And usage tends to grow quietly once something works, because people start feeding it more.
Before committing, work out your realistic monthly volume, then model it at three times that. If the answer at triple volume is uncomfortable, you have found a constraint worth knowing about now rather than later.
Per-seat fees, multiplied properly
Many AI features arrive as an add-on to software you already pay for, priced per user per month. The individual figure always sounds reasonable. Multiply by headcount and by twelve before deciding.
Worth asking a second question too: does everyone need it, or do three people need it and everyone else was included because per-seat pricing made it awkward to be selective?
The cost nobody quotes
If someone has to check every output before it is used, the work has not been removed. It has changed shape, from producing to reviewing.
Sometimes that is a genuine win, because reviewing a draft is faster than writing one. Sometimes it is not, because verifying a plausible-looking answer takes longer than doing the task properly in the first place. Which one you have depends on the task, and it is worth measuring rather than assuming.
Getting the data ready
This is the line item most quotes leave out. If your information is spread across an inbox, three spreadsheets and a shared drive nobody has tidied since 2019, something has to organise that before any AI system can use it usefully.
That work is often larger than the AI implementation itself. It is also worth doing regardless, which is the one consolation.
A worked example
Take a business wanting to summarise incoming enquiries and draft replies. Say 400 enquiries a month.
- Usage cost at that volume is genuinely small, typically a few euro a month at current rates.
- Building and connecting it is a one-off cost, and the larger number.
- Someone still reads the drafts before they go out, at least at first.
The interesting result is that the running cost is rarely the problem at small-business volumes. The build and the review time are. Which means the right question is not “can we afford the AI” but “is this process worth the build”.
Questions worth asking any quote
- What does this cost per month at our actual volume, and at three times it?
- Is the pricing per seat, per use, or both?
- What has to happen to our data before this works?
- Who checks the output, and how long does that take?
- What happens if we stop paying? Do we keep anything?
A supplier who answers those plainly is worth working with. One who cannot is quoting on hope.
If you have a quote in front of you and want it read properly, send over what you have been offered.
Before committing to a figure, it is worth being sure you are pricing the right thing. Copilot, a general assistant and a custom build solve different problems, and paying for the wrong one is the most common way this budget gets wasted.
It is also worth knowing the patterns behind the rollouts that quietly stop being used, because the cost that hurts most is the one that bought nothing.
If you have a quote and cannot tell what it genuinely commits you to over three years, describe what has been proposed and you will get a straight read on it.
Common questions
Is the licence fee the main cost?
No, and treating it that way is why budgets go wrong. The licence is the visible part; the time to verify output, the person who owns it, and the data tidying it depends on are usually larger.
Is there a cheap way to test the water?
Yes. One named group, one named task, a monthly subscription, and an honest measurement of the whole task including checking. That is the cheapest real answer available.
What does a custom build typically involve?
A build cost, a running cost, and a maintenance cost, plus an owner with actual time. The right moment is when a subscription has already proved the task is worth solving.