The biggest unknown in an AI integration is not the build cost but the running cost. A simple way to estimate the monthly bill before you commit.
When we discuss an AI feature with a client, the build cost is usually easy to settle. The question that makes everyone nervous is: what will it cost per month? That can be estimated, and before you commit.
The formula
Most language models bill on the amount of text processed, separately for input and output. The monthly cost comes together like this:
calls per month × [(average input tokens × input token price) + (average output tokens × output token price)]
Exact pricing varies by model and provider; cached input and some additional features, for example, may be priced separately. You can estimate the number of calls and expected workload, while actual token usage is best measured.
1. Number of calls
Page views do not matter; what matters is how often the feature actually runs. For document processing that is the number of files uploaded per month. For a support assistant, the number of incoming questions. It is usually far lower than you would first guess.
2. Text size
This is where estimates go wrong. The user's question is short, but what we send the model is not just that: it includes the system prompt, the context and often several pages of retrieved document excerpts. With a RAG-based search the input can be ten times what the user typed.
3. Measure, do not guess
That is why we usually build a narrow prototype first: run twenty real cases and look at the actual consumption. From there the monthly cost is no longer an estimate, it is multiplication.
What actually reduces it
A smaller model for routine work. Classification, tagging and short summaries rarely need the strongest model.
Caching. You should not pay again for the same repeated question.
Shorter context. More precise retrieval gives the model fewer but more relevant excerpts.
Cost monitoring. A monthly budget, alerting and, where supported, a spending limit so a runaway loop cannot produce an unexpectedly high bill.
When is it worth it?
The question is never whether AI is expensive, but expensive compared to what. If a process eats twenty working hours a month and this removes half of them, the decision is easy. If it would save two hours a month, the build will never pay for itself, and it is better to say so up front.
