- Provider tokens: depends on traffic and which model. A small internal agent can run under $50/month. A customer-facing agent at scale runs into the thousands. - Vector DB (if used): $20–200/month depending on storage and query volume. - Compute (Lambda or container): $10–100/month for small to medium scale. - Observability and eval reruns: $20–100/month for the LLM-traceable observability stack.

Cost analysis

AI agent build cost.

AI agents are easy to demo and hard to put into production. Here's what the production-ready version actually requires (and costs), versus the prompt-and-pray prototype that breaks the moment a real user touches it.

Build cost (one-time)

A production AI agent build is typically a 3–4 week engagement. €25.000 covers: provider selection (Claude / GPT / open-source), use-case scoping, custom tool definitions, API integrations, vector DB or RAG when needed, an eval suite for behavior regressions, prompt-injection defenses, output validation, and a production deploy with logging and traces.

Skip any of those and you have a prototype, not a production agent. Skipping evals in particular is the most common reason agents quietly degrade after launch.

Run cost (monthly)

· Provider tokens: depends on traffic and which model. A small internal agent can run under $50/month. A customer-facing agent at scale runs into the thousands.
· Vector DB (if used): $20–200/month depending on storage and query volume.
· Compute (Lambda or container): $10–100/month for small to medium scale.
· Observability and eval reruns: $20–100/month for the LLM-traceable observability stack.

Where most agents waste money

Three places: oversized models for simple steps (use a small model for routing, a large model only for the hard step), no caching of repeated tool calls, and no rate-limit guardrails (one runaway loop can burn a four-figure bill in an hour). All three are configuration, not architecture, and all three should be in place before launch.