The short answer
$35,000 to $120,000, and the range is that wide for reasons that are predictable rather than mysterious.
A scoped proof of concept against real data, with evaluations and guardrails, starts at $35,000 and takes four to six weeks. A production agent with full integration, observability and handover is typically $70,000 to $120,000. Multi-agent systems start around $90,000.
Almost nobody publishes these numbers. That is the main reason this article exists.
The build itself is the cheap part
Wiring up a model API takes an afternoon. If someone quotes you $8,000 for "an AI agent", they are quoting you the afternoon.
Everything that makes it dependable is the actual build:
Retrieval quality. Getting the right context in front of the model is usually harder than the model call. Chunking strategy, embedding choice, reranking, and knowing when retrieval returned nothing useful.
Tool use and orchestration. What the agent is allowed to do, in what order, and what happens when a tool fails mid-sequence.
The evaluation suite. A graded test harness with real cases, run on every change. Without it you have no idea whether today's prompt change made things better or worse — you have a feeling.
Guardrails. Input validation, output constraints, refusal behavior, cost and token ceilings. The ceiling matters more than people expect: an agent in a retry loop can spend real money in minutes.
Human-in-the-loop review. On the decisions where being wrong is expensive. Designing where the human sits is a product decision, not a technical one.
Observability. Logging, tracing and alerting on failure modes — not just uptime. Uptime tells you the agent responded. It does not tell you the response was garbage.
The running costs nobody quotes
Three, and they surprise people:
Inference. Varies enormously with prompt size, context strategy and volume. Worth modeling before you build, because the architecture that is cheapest to build is frequently not the cheapest to run.
Re-evaluation. Models get deprecated. Providers change behavior on the same version number. The eval suite has to be re-run, and something has to be done when it regresses.
Drift. Prompts that worked in March stop working in September because your data changed, your users changed, or the model did. This is the cost that gets budgeted at zero and then isn't.
What makes it more expensive
- Dirty data. Almost always the single biggest multiplier.
- Integration count. Each system is its own auth, rate limit and failure mode.
- The cost of being wrong. A support-triage agent and a financial-recommendation agent need very different guardrails, and the second costs considerably more.
- Regulatory constraints. Audit trails, data residency and explainability are engineering work, not paperwork.
When you should not build one
Genuinely: when a well-configured off-the-shelf tool does the job, when nobody can define what a wrong answer looks like, or when the underlying process is broken. An agent on top of a broken process produces failures faster.
We will tell you if that is our read. It costs us the work and saves you the project.