The most common question after “can this be automated?” is “what will it cost to run?” — and most vendors answer it vaguely. This guide gives you the framework we use in every proposal.
AI system costs split into three buckets: model usage (per-token API costs), infrastructure (hosting, storage, queues), and maintenance (the human attention that keeps quality high as inputs drift).
The three cost buckets
Model usage scales with volume and prompt design. A document-extraction workflow processing 500 documents a month typically costs tens of euros in API fees — not thousands. The trap is unbounded workflows: chat interfaces with no context limits can surprise you.
Monthly cost estimate (500 docs/month) --------------------------------------- Model usage (Claude) €25–60 Hosting (VPS + Redis) €20–40 Monitoring included Maintenance retainer from €250 --------------------------------------- Typical total €300–350/mo
Infrastructure is usually the smallest line. Most automation workloads run comfortably on a modest VPS with Redis — the hyperscaler bill only appears when someone architects for imaginary scale.
Ask any vendor for a written monthly cost estimate before you sign. If they can’t produce one, they haven’t designed the system yet.
Estimating before you build
To estimate your own project before talking to anyone:
- Count the monthly volume of the task being automated.
- Estimate tokens per item (a page of text ≈ 500 tokens).
- Multiply by current model pricing — then double it for safety.
- Add €20–50/month hosting unless you have unusual requirements.
Running costs are a design decision, not a surprise. Systems built with cost budgets stay cheap; systems built without them drift expensive. Either way, you should see the number before development starts.