What an AI agent really costs a small business

Short answer: building one costs real money, and running one usually doesn’t. Most articles on this get that backwards, because most of them are written by agencies selling the build.

The two bills

1. The build. Figuring out the job, connecting to your systems, writing the checks, testing it against real work. That’s where the hours go.

Published ranges, all from companies that sell builds. I couldn’t find a neutral survey, so take them with that in mind:

managed or configured agents, $3,000–12,000 setup (Advantech ITS);
local Columbus agencies, from $3,000 (The Automators) to $7,500 (AutomateNexus);
“typical custom agent,” $20,000–60,000 (Biz4Group).

2. Running it. Model fees, a bit of hosting, and upkeep.

A worked example: review replies

Say you get 300 Google reviews a month and want a drafted reply for each, which you approve before it posts.

Each reply is roughly 3,000 tokens in (the review, your tone guide, a few examples) and 500 out.
On Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output (claude.com/pricing, checked October 1, 2026):
- 300 × 3,000 = 900K input tokens, about $1.80
- 300 × 500 = 150K output tokens, about $1.50
That’s about $3.30 a month in model fees for short, plain replies. If you switch on a model’s extended reasoning, the thinking is billed as output too, so budget two to three times the output cost: call it $3 to $8 a month.

That’s the part people are afraid of, and it’s the cheap part.

So where does the money go?

Knowing what “working” means. If you can’t measure it, you can’t tell whether it’s helping.
Connections. Your review platform, your CRM, your inbox. Integration is most of the build.
The checks. What happens when it’s wrong. It will be, sometimes, and a company is responsible for what its AI says (Moffatt v. Air Canada, 2024, ABA summary).
Upkeep. Models change, prices change, your business changes.

How to not overspend

Start with one job, not “AI strategy.”
Run a pilot next to the person doing it today, and compare.
Get the monthly running cost in writing, in dollars.
Make sure you own the code and the accounts. If you part ways with the builder, it should keep working.

Where I come in

I build these for Columbus businesses: one job, measured, with checks. The first step is an AI Opportunity Assessment at $950. I map two or three workflows, rank them by return, and hand you a working prototype of one on sample data plus a written plan. If it’s worth running, a pilot at $4,500 puts one workflow beside the person who does it today for a few weeks, with measurement and a plain go or no-go report. Every dollar of the pilot credits toward the build if you go ahead.

The tools I build run on your own accounts, under your provider’s business terms. Your data doesn’t pass through a personal subscription of mine.

Email [email protected] with the job you’d like off your plate.