AI costs for businesses: the guide

The cost of an AI project has three layers. Tools and subscriptions are usually the smallest and most predictable part: from a few dozen to a few thousand euros a month, depending on volume. Implementation — designing the solution, connecting it to your CRM, invoicing or email, testing and going live — is the largest share of the initial investment. Maintenance completes the picture: the first months always bring adjustments.

Some costs are usually left out and spoil the promised return: cleaning up messy data, training the team and giving habits time to change. None of them is large, but ignoring them means the spreadsheet ROI never shows up in reality.

When an application uses an AI model through an API, it pays per token. A token is a chunk of text: about 0.75 words on average, and a million tokens is roughly 750,000 words. You pay for two kinds: input tokens (the question, instructions and context you send) and output tokens (the generated answer).

Output almost always costs about five times as much as input. That is why long answers, conversation histories resent with every message and unnecessary context are what make the bill grow.

The estimate rests on four variables: input tokens per request, output tokens per request, requests per month and the model's price. A support chatbot with 300,000 requests a month, 500 input and 300 output tokens, on a model at $1 input and $5 output per million tokens, costs about $600 a month.

Volume dominates: a simple use case with many calls can end up costing more than a heavy one with few. Add a 20% to 30% margin for peaks and always present a range, together with the volume it is based on.

The biggest lever is the model. The price per token varies by more than a hundred times between the cheapest and the most expensive model, and within one provider the gap between models matters more than the gap between providers. Simple tasks such as classifying, extracting data or answering common questions rarely need the top tier.

Then come the providers' own mechanisms: prompt caching, which makes reads of repeated context up to 90% cheaper, and batch processing, with a 50% discount for work that doesn't need an instant answer. Shorter prompts, limits on answer length and monitoring usage per feature complete the list.

Cost only makes sense next to what you save. For an automation the maths is simple: hours per month × the real cost of an hour gives the current cost; automation typically removes 70% to 90% of those hours; the project cost divided by the net monthly savings gives the months until it pays for itself.

To prove the return to leadership, record the baseline before you start, translate results into time, cost and revenue, separate one-off from recurring gains and attribute results honestly. And don't forget the cost of doing nothing: a team that keeps working by hand pays it every month.