How much does it cost to build an AI agent?
Short answer: The cost of building an AI agent is driven by scope, integration surface and reliability requirements, not by the model. An agent doing one job against one system is a matter of weeks. One that acts across several systems of record, with approvals and audit trails, runs to months, and the integration work is where the schedule actually goes.
“How much to build an AI agent?” has the same honest answer as “how much to build a house?”: it depends on what you are building. But the cost drivers are predictable, and knowing them helps you budget for something that actually ships rather than a demo that dies in a slide deck.
What drives the cost
1. Scope: one job or many? A single-task agent (answer from these docs) is a fraction of the cost of a multi-agent system that routes across several capabilities and calls your tools. Be honest about which you need.
2. Integrations. An agent that only chats is cheap. An agent that reads your database, calls your APIs, and writes back to your systems of record is where real engineering time goes, because each integration needs auth, error handling and testing.
3. Data and retrieval. If the agent needs your knowledge, you need a retrieval layer: ingestion, chunking, embeddings, evaluation. Messy source data is the most common hidden cost.
4. Production hardening. This is the difference between a prototype and a product: persistent memory, observability, guardrails, human-in-the-loop review, CI/CD and deployment. It is also where most cheap builds cut corners, which is why they fail in production.
Why the cheapest quote is the most expensive
A $500 prototype that impresses in a demo but hallucinates, forgets context, and cannot connect to your systems is not cheap. You pay for it again when you rebuild it properly. The real question is not “what is the lowest price” but “what will run reliably in front of my users.”
How to budget
- Prototype / proof of concept: small, fixed scope, to validate the idea.
- Production build: the real number, driven by integrations and hardening above.
- Retainer: ongoing improvement, evaluation and new capabilities once it is live.
The best way to control cost is a tight, well-defined first scope. Ship one valuable workflow to production, prove the value, then expand.
A worked range
Numbers are useless without their assumptions, so here are the assumptions.
A single-purpose agent against one system. Answers questions and takes one or two bounded actions through an API that already exists and is documented. Four to eight weeks. Most of it is the tool layer, the authorisation checks and the evaluation set rather than the agent loop.
A multi-step agent across several systems. Holds state, calls three or four tools, survives a failure halfway through. Three to five months, and the range widens because it depends almost entirely on how cooperative those systems are.
An agent inside a regulated process. Everything above, plus an audit trail that reconstructs a decision years later, an explainability requirement, and a review pathway. Add half again, and add a sign-off cycle that is calendar time rather than engineering time.
The spread between those tiers is not model sophistication. It is how many systems the agent touches and what happens when it is wrong.
What makes a quote wrong
Four things get left out consistently.
Integration reality. Quoted from a diagram, discovered in the code. The API that exists has a rate limit that makes the design unworkable, or needs a field the agent has no way to know.
The evaluation set. Without one you cannot tell whether a change improved the agent. Building it is real work and it is the first thing cut from a competitive quote.
The exception path. Every agent has actions it cannot complete. Who sees them, with what context, and how fast they resolve. Omitting this does not remove the cost, it moves it to your operations team.
Running cost. Inference is usually the smaller line and falling. The persistent costs are whoever owns the evaluation set and whoever clears the exception queue.
Cheaper ways to get the same outcome
Worth asking before commissioning anything.
If the task is one step between two systems with clean APIs, use an automation platform and stop. If the task is answering questions from documents, that is retrieval rather than an agent, and it is substantially cheaper to build and to run. If a product already does this for your industry, buy it; the cheapest agent is the one nobody has to build.
An agent earns its cost when the work genuinely requires deciding what to do next based on what just happened, across more than one system, at volume. That is a narrower set of problems than the current marketing suggests.
Tools referenced
- OpenAI pricing
- Anthropic pricing
- LangGraph
EpochC builds AI agents and multi-agent orchestration end to end. See the clinical multi-agent case study for what a delivered system looks like, or contact us for a straight number on your scope.