The question I get on almost every intro call is some version of "what does an AI agent cost?" And the honest answer is that the number people worry about (the build) is usually the small one. The number that matters is the monthly run cost, because that is the one that decides whether the agent survives its first budget review. So here is the short answer, then the actual bill, line by line.
How Much Does an AI Agent Cost? Build vs. Monthly
Keep two numbers apart. The build is a one-time cost: scoping the workflow, wiring it into your tools, and testing it on real work. The run cost is what you pay every month after that, and it is made of the four lines below. For a single-workflow small business agent, the monthly figure generally lands between tens and low hundreds of dollars. To compare quotes fairly, add the build to twelve months of run cost and look at the first-year total. And if you are wondering about cloud or server costs: a hosted workflow platform or a small self-hosted server is the orchestration line, not an extra one.
The Four Lines on the Bill
Line 1: Model usage. This is the per-request cost of the AI itself, billed by Anthropic or OpenAI based on how much text goes in and out. It is also the line people overestimate the most. A typical small business workflow (say, classifying and drafting replies to a few hundred inquiries a month) uses so few tokens that the model bill often lands in the single or low double digits monthly. As I covered in AI Got 60% Cheaper, per-token prices have been falling for years. Model usage only becomes a real line item when you process high volumes of long documents or run agents continuously.
Line 2: Orchestration. Something has to trigger the agent, pass it data, and route the output. That is a workflow platform like n8n, Zapier, or Make, or a small self-hosted setup if you have the appetite. Entry business tiers on these platforms run from roughly twenty to a few hundred dollars a month depending on volume and seats. Self-hosting n8n on a small server can cut that to server cost alone, at the price of maintaining it yourself.
Line 3: The integrations tax. The quiet line. Connecting the agent to your actual systems sometimes requires upgraded plans: the CRM tier that unlocks API access, the phone system add-on, the form tool's paid webhooks. None of these are AI costs on paper, but the agent is why you are paying them. Budget for one or two of these surprises.
Line 4: Maintenance. The line nobody budgets and everybody pays. APIs change, edge cases appear, prompts drift out of date as your business changes. Whether it is your time, an employee's, or a retainer, a working agent needs a few hours of attention a month. An agent with zero maintenance budget is an agent with an expiration date.
Rough shape, not gospel: proportions vary by workflow. The point is that the AI itself is usually the smallest line, and the human attention line is the biggest.
Three Realistic Tiers
- DIY single workflow: a chat subscription plus a starter automation plan. Tens of dollars a month, plus your own time as the maintenance line.
- Typical built agent: one or two wired workflows on API access with a proper orchestration tier. Low hundreds per month all-in, including a modest maintenance allowance.
- Heavy usage: high-volume document processing, voice agents on phone lines, or always-on monitoring. This is where model and telephony usage become real and the bill reaches four figures. If you are here, you should be measuring revenue per run, not cost per month.
How to Keep the Bill Boring
Three habits. Use the cheapest model that passes your quality bar, and reserve frontier models for the steps that need judgment; routing easy steps to cheap models can cut the model line dramatically. Set spending alerts on every API key from day one. And review the agent's transcript log monthly: the same log that catches quality drift also catches the runaway loop that quietly triples your token bill.
The Bottom Line
A single-workflow agent for a small business generally runs from tens to low hundreds of dollars a month, and the AI itself is the cheapest part. Budget the maintenance line honestly and compare the total against the hours the workflow currently eats. If the agent does not clearly win that comparison, do not build it. That math is the whole decision.
Sources
- Anthropic, API pricing (per-token rates by model tier). anthropic.com
- OpenAI, API pricing (per-token rates by model tier). openai.com
- n8n, plan pricing for cloud and self-hosted workflow automation. n8n.io
- Zapier, plan pricing for workflow automation tiers. zapier.com
Want the real number for your workflow before you build?
We scope the monthly run cost line by line for your specific use case, so the decision is arithmetic instead of hope.