What AI actually costs inside an automation.

When I say "we'll put an AI model on reading the invoices", people picture a subscription of hundreds of euros a month. In reality, a call that reads one document costs a fraction of a cent. With the models I use this year, a page of text costs less than a text message. At a few hundred documents a month, the AI bill is usually smaller than Monday morning's coffee.

That doesn't mean the project is cheap. It means the money goes elsewhere.

Where the money actually goes

The build. Writing a good instruction for the model. Testing it on 50 real invoices, not three pretty ones. Catching the odd cases: two-page invoices, crooked scans, amounts in euros, a supplier who puts the total in the header. Building the validations around it. That's most of the price, and it's the same work as any automation, plus a little.

The mistakes. One misread amount that ends up in the books costs more than a year of AI calls. That's why I spend time on the safety nets, not on choosing the model. A cheaper model with good validation beats an expensive model without it.

The maintenance. Models change, suppliers change their invoice layout, a new supplier shows up with a strange PDF. Someone has to look in the "to check" folder now and then and adjust. With me that's the maintenance plan, and it's the reason it exists.

When AI is more expensive than a rule

When the data is already tidy. Putting a model on reading a CSV is like hiring a translator for a text that's already in English. A rule does the same job in milliseconds, for free, identically every time. Run 10,000 times a month, the difference shows in money and in nerves.

How I work it out for a client

1. Volume. How many documents, emails or photos a month?
2. Gain. What does an hour of the person who reads them now cost, and how many hours is it?
3. Risk. What does a mistake that slips through cost?
4. Calls. How many AI calls per document, at what price? Usually the smallest number in the whole table.

If the volume is small and the risk is high, a good person with a good form may well be cheaper than any automation. I'll tell you that too, because otherwise it costs you twice: once for the build, once for the disappointment.