What does a custom AI agent cost a small business?

A custom AI agent for a small business has two prices: a one-time build and an ongoing run cost. At Nexos, builds start from $4,900 per fixed-scope project, a private model on dedicated infrastructure starts from $890 a month, and the audit that tells you which one you need is free. What pushes a project above those starting points is scope, the number of systems the agent has to touch, and how much work it does every month.

That's the short answer. The rest of this piece explains where the money actually goes, so you can tell a fair quote from an inflated one, including ours.

What are you actually paying for?

"An AI agent" sounds like a single thing you buy. It's really four things, and a quote that doesn't separate them is hard to judge.

Our own pricing follows the same split. The audit covers discovery. The Build tier covers design, integration and launch, and includes 30 days of post-launch support. The Run tier covers a dedicated private model if you need one. The pricing page shows all three.

What makes a build cost more?

The single biggest driver is how many systems the agent has to touch. An agent that reads one inbox and drafts replies into the same inbox is a small project. An agent that reads the inbox, looks up the customer in your CRM, checks stock in your inventory system, drafts a quote in your accounting tool and asks a manager to approve it is five integrations, each with its own logins, data formats and failure modes.

After that, in rough order:

• How messy the inputs are. Clean, structured data (a form, a spreadsheet with consistent columns) is cheap to work with. Free-text emails, scanned PDFs and "it's in Dave's notes" take more building and more testing.
• How many edge cases matter. The main path through a workflow is usually quick to build. The unusual 10% of cases (the refund that's also a complaint, the order split across two accounts) is where the hours go.
• Where a human needs to approve. Every approval point needs a place to show the draft, a way to accept or edit it, and a way to resume afterwards. We think those steps are worth building. They also add scope. We wrote about where to put them in How much should an AI decide on its own?
• What "done" means. An agent that drafts for a person to send is cheaper and lower-risk than one that sends on its own. Most first projects should be the drafting kind.

What does it cost to run every month?

Once it's live, an agent's running cost is mostly model usage: you pay the AI provider for the text the model reads and writes, measured in tokens. OpenAI, for example, publishes its API prices per million tokens, with input and output priced separately and lower rates for cached input and batch processing (OpenAI API pricing).

Which model you choose matters a lot. At OpenAI's standard rates as of September 2026, input for its flagship models ranges from $0.10 per million tokens for the smallest (gpt-6-luna) to $10.00 for the largest (gpt-6-astra), and output from $0.50 to $50.00 per million (OpenAI API pricing). That's a hundredfold difference for the same volume of work.

To show what that means, take an illustrative workload (these are assumptions, not client data). An agent handles 1,000 emails a month, reading about 3,000 tokens of context for each one and writing about 500 tokens back. That's 3 million input tokens and 0.5 million output tokens a month:

For a narrow agent, the model bill is often the smallest line in the budget. It grows with volume, with long inputs (whole contracts rather than short emails) and with agents that work through many steps per task. Two practical rules follow:

1. Use the smallest model that clears your quality bar, measured on your own examples. Save the expensive model for the steps that really need it.
2. Put a spending cap and usage reporting in from day one, so a loop or a spike in volume shows up as an alert, not as a surprise invoice.

When does a private model make sense?

Some businesses can't, or would rather not, send their data to a third-party API. Regulated records, client confidentiality and contract terms are the usual reasons. For those, the model runs on dedicated infrastructure: hosted and managed by us from $890 a month, or installed on your own hardware and quoted per install.

That changes the cost profile from metered to fixed. You pay the same each month however much the agent does. That's predictable, and at high, steady volume it can beat per-token pricing. At low volume, a hosted API is usually cheaper. We cover the trade-off in more depth in Private AI without a research team.

What ongoing maintenance should you budget for?

An agent connects to tools that change underneath it. APIs get updated, a field gets renamed in the CRM, a vendor changes its login flow. Your business changes too: new products, new policies, a new approval rule. Somebody has to notice and update the agent.

Budget for:

• Integration upkeep when a connected tool changes.
• Quality review. Look at a sample of what the agent did every so often, not only at whether it ran. Is your AI actually working? explains how to measure that.
• Model changes. Providers retire older models on their own schedule, so an agent should be built to swap models without a rebuild.

You can handle this in-house with good documentation, or put it on an ongoing maintenance and support retainer. Either way, a quote that has no maintenance line at all is incomplete.

How do you keep the total honest?

A short checklist to use on any quote, including ours:

• Are build and run costs listed separately?
• Is the scope a named workflow with a clear definition of "done", not "AI for operations"?
• Does it say which systems will be integrated, and which won't?
• Does it name the model (or model tier) and estimate monthly usage cost at your volume?
• Is there a spending cap and usage reporting?
• Does it cover what happens after launch: support window, maintenance, who owns the accounts and configuration?
• Could a cheaper fix (buying software or a simple automation) do the same job?

If you can't answer those, you don't have a price yet. You have a guess.

Where to start

The cheapest way to get a real number is to scope one workflow properly before building anything. That's what the AI-readiness audit is for. It's free, it ranks your workflows by return and effort, and the written report is yours whether or not you build with us.