How much does it cost to build an AI agent?

The short answer: building an AI agent yourself costs $20 to $200 a month in model API calls plus a server for a modest internal workload, which sounds cheap until you add the engineering. A custom agent built by a contractor or agency commonly runs $10,000 to $50,000 to deliver, and the maintenance never ends. Building wins when the workflow is specific to your business and central enough to justify owning. Buying wins when your constraint is time, which at most small companies it is.

The visible costs, which are genuinely small

Start with the part that shows up on a credit card statement, because this is the part that makes building look obvious.

An open source orchestration tool like n8n can be self-hosted for no license cost. A small VPS to run it on is $10 to $40 a month. Model API usage for a modest internal workload, say a few hundred agent runs a month doing research and drafting, typically lands between $20 and $200 depending on which model you point it at and how much context each run drags along. Add a vector database or a scraping API if your workflow needs them and you are still comfortably under $300 a month for something real.

Compared to $149 to $500 a month for an agent that already works, that looks like a straightforward win. It is not, and the reason is that none of the above is the expensive part.

The invisible costs, which are the actual cost

The expensive part is engineering, and it arrives in four waves.

First, scoping. Turning "handle our inbound leads" into a specification a machine can execute is harder than it sounds and it is where most internal agent projects stall. You discover that half your process lives in someone's head and the other half has exceptions nobody documented.

Second, integration. Every system the agent touches needs authentication, rate limit handling and a data mapping that survives contact with real records. If the agent has to reach into a CRM, a mailbox, a billing system and two spreadsheets, you are not building an agent, you are running a data integration project with an agent on top, and you should budget it like one.

Third, reliability. A demo that works nine times out of ten is a great demo and an unusable production system, because the tenth run wrote something wrong to a system of record. Error handling, retries, human approval steps and an audit log are most of the actual build, and they are the parts nobody estimates.

Fourth, maintenance. APIs change, connectors break, models get deprecated, and edge cases surface for as long as the thing runs. Whoever built it becomes the only person who can fix it. At a fifty-person company with a platform team, fine. At a twelve-person company where that person is also your only ops hire, that is a real single point of failure.

What contractors charge, and why the range is so wide

If you hire it out, the market range for a custom agent is roughly $10,000 to $50,000 to deliver, and it is not unusual to see six figures for anything touching regulated data or a legacy system.

The spread is mostly about integration surface and reliability requirements. A single-workflow agent that reads one API and writes to one destination sits at the bottom of that range. An agent that touches four internal systems, needs an approval workflow and cannot be allowed to make a mistake in front of a customer sits at the top. The prompt engineering, which is what people imagine they are paying for, is a small fraction of the invoice.

Then there is the retainer. Custom software has an ongoing cost whether you formalize it or not, typically 15 to 20 percent of the build cost a year if you keep the contractor on, or a permanent claim on an internal engineer's time if you do not. Ask for that number before you sign the build, not after.

The honest test for build versus buy

Three questions settle it faster than any spreadsheet.

Is the workflow specific to your business? If it depends on your own data model, your own rules and a process no other company runs, no product will fit and you should build. If it is research, list building, reporting or inbox triage, that is a common job and someone has already built it better than you will.

Is it central enough to justify owning? Owning software means owning it forever. That is worth doing for a process that is core to how you make money, and a bad trade for a task that just needs to stop eating Tuesday afternoons.

Who actually has the hours? This is the one that decides most cases honestly. Most small teams that decide to build are really deciding to add a project to whoever is already furthest behind. If you cannot name the person, the calendar time and who covers it when they are on vacation, you are not going to build it.

The middle option people miss

Build and buy are not the only two choices, and the framing hides a useful third one: buy the general jobs, build the specific one.

Most companies have exactly one workflow that is genuinely theirs and a long list of ordinary tasks that are not. Buying an agent for the ordinary tasks costs a few hundred dollars a month and takes an afternoon. That frees your engineering time for the one workflow that is actually a competitive advantage, instead of spending it rebuilding a research agent that already exists as a product.

The mistake is spending three months building the generic thing and never getting to the specific one. If you want the decision laid out in full, we covered it in build or buy an AI agent. For what the finished products in each category cost, including what each vendor meters, see the AI agent pricing breakdown, and the category map is on AI agent software.