How much does it cost to build an AI agent?

Ask five vendors what an AI agent costs and you'll get five different numbers, because the honest answer depends on scope, not a single figure. Based on published vendor guides across the market, the working range is: a prototype runs $10k–$30k, an MVP $20k–$60k, and a single production-grade agent $20k–$80k. Complex, multi-agent systems run $100k–$500k+, and enterprise programs land at $100k–$200k+ per phase. Year-one, all-in totals typically fall around $160k for a lean build, $246k mid-market, and $400k–$450k for enterprise. These are market ranges observed across published vendor guides, not our own pricing — what any vendor quotes you depends on your specific scope.

What actually moves a project between those bands isn't headcount, it's how much of the surrounding system already exists. An agent that reads from one API you already trust and writes to nothing is a $20k–$40k job. The same agent writing to a CRM or ERP, with permissioning, an audit trail, and human-in-the-loop confirmation on irreversible actions, is a different project — the model call is a rounding error next to the integration and evaluation work around it.

The question keeps getting asked this hard because the buyer population has changed shape, and the honest measure of that is a statistical agency rather than a vendor forecast. The US Census Bureau's Business Trends and Outlook Survey put AI use between 17% and 20% of American businesses across December 2025 to May 2026, with a further 20% to 23% expecting to be using it within six months — and the rate climbs steeply with size, reaching 37% at firms with 250 or more employees and 32% at 100 to 249, against roughly 14% in retail trade and 39.7% in the information sector (US Census Bureau, AI Use at U.S. Businesses, May 2026). That spread is exactly why per-project quotes vary so widely: buyers aren't comparing a handful of similar boutique builds anymore, they're comparing everything from a scripted single-tool assistant to a multi-agent system with its own evaluation pipeline, and vendors price accordingly.

The build-vs-buy decision changes the number more than any of the above. Off-the-shelf agent builders and embedded copilots inside the SaaS tools you already pay for are now the default first option rather than an exotic one — most major business applications ship some version of an assistant in the box, and the marginal cost of switching it on is a licence tier rather than a project. That option is real and often the right one when the workflow is generic — drafting a reply, summarizing a ticket — because someone else has already paid the integration and evaluation cost and amortized it across thousands of customers. It stops being the right comparison the moment the workflow is specific to how your business actually operates: a claims process with your carrier's rules, a KYC flow with your risk thresholds, a service-drive workflow with your dealership's escalation logic. At that point the embedded agent's generic version either can't do the task or does it wrong in ways nobody notices until an audit, and the real comparison is the embedded fee against the custom-build numbers above, not against zero.

The number most budgets miss isn't the build at all — it's the run-rate. CloudZero's State of AI Costs report measured average enterprise AI spend at $62,964 a month in 2024 and projected that would “rise to $85,521 in 2025 (a 36% increase)”, with the proportion of organizations planning to invest over $100,000 per month “set to more than double, jumping from 20% in 2024 to 45% in 2025” — the 2025 figures being that report's forecast rather than a later measurement of it (CloudZero, State of AI Costs, 2025). Why an agent's run-rate outgrows a chatbot's is not a forecast — it is documented in the API itself. Model APIs are stateless, so every step of an agentic task is a fresh request that carries the whole conversation so far: Anthropic's documentation states that "as the conversation advances through turns, each user message and assistant response accumulates within the context window, and previous turns are preserved completely," and that "everything in the request counts toward the context window: the system prompt, every message in messages (including tool results, images, and documents), and your tool definitions" (Anthropic, Context windows). A task that takes six tool calls to finish therefore pays for a growing prefix six times, not once.

Published vendor guides put ongoing maintenance at 15–30% of the build cost per year on top of that, plus separate API spend ($100–$10k/month depending on volume) and hosting/infra ($200–$5k/month). A $60k MVP that looks finished on delivery day quietly costs another $9k–$18k a year in maintenance alone — before the accumulating per-task token consumption above even shows up on the bill. Budget the first year's total, not the invoice for the build.

Evaluation is the line item most quotes underscope, and it's the one that determines whether any of the numbers above hold once the agent meets real production traffic instead of a demo script. A golden dataset, a regression gate that blocks a release when accuracy drops below an agreed budget, and per-capability scoring are what turn 'is it getting better?' into a chart instead of an opinion — and none of it is optional once the agent's mistakes have a real cost. Vendors who quote a single number for 'the AI agent' with no separate evaluation-suite line item are usually folding it into the build estimate silently, which is exactly how a fixed-scope quote turns into a change order three weeks in. The decision rule: ask what the acceptance test is before the code gets written. If nobody can answer that yet, the honest next step is a short, capped discovery phase — not a fixed-scope commitment neither side can actually price yet.

Scope is what actually separates a $20k agent from a $200k one, and the shape of the work tells you which side you're on before a single quote arrives. An agent that reads one API you already trust and writes to nothing sits at the cheap end. A multi-assistant AI platform serving several domain-specific assistants with conversational memory, calendar integrations, and full prompt observability — the kind of build we shipped over six months with a five-person team — sits at the other end, because the integration surface and the evaluation work around each assistant is the majority of the bill, not the model call itself.

The practical test for any quote: ask the vendor to break out build cost, monthly run cost, and annual maintenance separately, and ask specifically how they priced the reasoning loop — the tool calls, retries, and self-checks a single task runs through before it produces an answer, each one re-sending the context that came before it. A quote with only one number hasn't scoped the whole project yet. If you're evaluating AI agent development company USA options for a production build rather than a demo, that same three-line breakdown is the fastest way to tell a scoped proposal from a placeholder one — and it's the same discipline behind our own AI development pricing approach: fixed deliverables, not a rate-card guess.