Article
People ask this the way they would ask the price of a car, expecting one number. The honest answer is a range, and the range is wide because "agent" now describes systems that differ by a factor of ten in scope. A narrow agent that does one job inside a product you already have runs $15k–$40k. A platform where agents are the product, with multiple tools, permissions per user, and a review workflow, runs $80k–$200k+. Both are called agents. Only one of them is what most people picture.
What the number actually buys
The model call is close to free to build; it is a function call with a prompt. What costs money is the harness around it: the tool integrations, the permission model, the logging, the stop conditions, and the evaluation set that proves it works before a customer finds out it does not. On a typical narrow agent, the model integration is a week of the schedule and the harness is the other three to seven.
Quotes that look cheap usually priced the model call and skipped the harness. The project then either stalls in review, because nobody can answer "what happens when it is wrong," or it ships without answers and the cost shows up later as an incident.
The variables that move the number most
How many tools it calls. Each tool is an integration, a permission decision, and a failure mode to handle. Three tools is a different project from ten.
Whether it writes or only reads. A read-only agent that drafts a summary is forgiving of mistakes. An agent that modifies customer data needs approval flows, audit logs, and reversibility, which is most of the remaining budget.
How much of the surrounding system exists. An agent bolted onto a product with clean APIs and an existing auth model is cheaper than one that needs those built first. We often find the agent is the easy 30% and the plumbing around it is the rest.
How rigorous the evaluation needs to be. An internal tool for five people can launch on a spot-checked test set. A customer-facing agent needs the kind of evaluation discipline that catches regressions before customers do, and that set does not build itself.
The cost people forget: running it
Build cost is one number; running cost is another, and it does not stop after launch. A support agent handling a thousand conversations a month, with a mid-size model and reasonable context, typically runs $200–$1,000 a month at today's pricing. That sounds trivial until usage grows tenfold and nobody put a routing and caching strategy in place, at which point the running cost can exceed the build cost within a year. Ask any vendor for both numbers, not just the first.
What we quote against
We scope agent work the same way we scope everything else: a written list of the tools it calls, what it is allowed to do unsupervised versus what needs approval, and the test set that will prove it. That list is what turns "how much does an agent cost" into a number we can defend, usually as part of the AI Integration workflow, which prices the harness explicitly instead of hiding it inside "development." Our AI Agents service page has the fuller breakdown of what's included at each tier.
The question to ask a lower quote
If a proposal is a third of everyone else's, ask what it does when the agent calls a tool with the wrong arguments, and ask to see the test set. Silence on either is the missing 70%, arriving later as a change order.