If you own a small business, your inbox probably has three or four pitches for "AI agents" in it right now. One promises to answer your phones. Another says it will run your sales team. A third is a chat bubble with a new name. They all use the same word, none of them describe the same thing, and almost none of them tell you what it costs.
So here is the guide I'd want if I were the one buying. What an AI agent for a small business actually is, which jobs are worth handing to one, what you should expect to pay, and which pitches I'd skip.
What an AI agent is, compared with what you already have
There are three things people lump together, and the differences decide what you should buy.
A chatbot answers questions inside a window. It can tell a website visitor your hours. It usually can't see your calendar, your customer list, or what you quoted someone last week, and it can't do anything after the conversation ends. I wrote about the gap between a chatbot and real automation if you want the longer version.
A Zapier-style automation follows a script. When a form comes in, send this email. When an invoice is ten days late, send that one. It's reliable and cheap, and it falls apart the moment a message doesn't fit the script, because it can't read.
An AI agent sits in between and past both. It can read a message, work out what the person wants, look something up in your tools, take an action, and hand the situation to you when it's out of its depth. I've explained why I call that a digital employee, because the way you evaluate one is closer to how you'd evaluate a part-time hire than a piece of software.
The practical test for any pitch is simple. Ask what it can see and what it can do. If the answer is "it can talk," that's a chatbot. If it can check your calendar, write to your customer records, and book an appointment, that's closer to an agent.
The jobs that are worth giving one
Agents work best on jobs with high volume, clear boundaries, and low stakes for any single interaction. For a small business that usually means four things.
Responding to new leads. Someone fills out a form or misses your call, and the first reply goes out in a minute instead of a day. This is the most common starting point because the money is easy to see. I walked through the mechanics in my post on building a speed-to-lead system, and the Lead Response Agent is the packaged version of it.
Following up after the quote or the visit. Most small businesses send one follow-up, if that, and then the lead goes quiet. An agent can send the second and third, in your voice, and stop the moment the person replies.
Collecting what you're waiting on. Intake forms, signed contracts, tax documents. Nobody enjoys sending the third reminder, and an agent doesn't mind.
Dispatching and scheduling. For trades and field service, that means handling reschedules and "when is my tech coming" messages. I covered how that works in practice in the AI dispatcher post for HVAC and plumbing.
You'll notice what's missing. Strategy, pricing a complicated job, calming an angry customer, deciding who to hire. Those need judgment, and a good agent is built to recognize them and hand them to you with a summary already written.
What an agent costs
Nobody agrees on this number, and most vendors skip it, so here is how I'd think about it. There are two kinds of cost: getting it set up, and keeping it running.
Setup is where the range is widest, and it depends on the same things I laid out in what small business automation costs. How many tools it has to connect to, how clean your customer data is, and how many situations it has to handle. One agent doing one job against one calendar and one inbox is a small project. An agent that touches your booking system, your CRM, your phones, and your accounting is a bigger one. I won't quote a single figure here, because anyone who does before they've seen your setup is guessing.
Running costs are easier. Off-the-shelf tools from national vendors commonly land somewhere between a few dozen and a few hundred dollars a month, and the price usually climbs with how many calls, texts, or contacts you push through them. For my own work, the ongoing support I offer after launch is $500 a month, which covers monitoring, fixing things when a tool changes, and a monthly call. That's a published number, and you can see it on the AI agents page.
The way to judge any of these prices is against the job. A reasonable comparison is a part-time person for that task. If someone to answer after-hours calls and chase leads would cost you a couple thousand dollars a month before training and turnover, an agent at a fraction of that is a fair deal, as long as it does the job well. If the agent costs more than the leads it recovers, skip it. That's also why I'd start with a number from your own business: how many leads did you lose last month, and what is one worth?
Pitches I'd skip
A few patterns come up often enough that I'll name them.
The agent that "runs your whole business." Nothing does. Anyone promising that is selling a demo.
The platform built for agencies, sold to a business with eight employees. You end up paying for features you'll never open, and you still have to build everything yourself. If you're weighing tools like that, my post comparing Zapier and n8n is a decent place to see how the pieces differ.
The agent with no handoff. Ask what happens when it doesn't know the answer. If the honest reply is "it tries its best," keep walking. A good one escalates to a person, quickly, with context.
Anything that can't say what it was told not to do. An agent should have a short list of things it never promises: a price it hasn't been given, a time slot you haven't offered, a diagnosis. Without those limits, it will eventually say something you'd never say, and I covered the risks in is it safe to let AI talk to your customers.
When an agent is not worth it
This is the part vendors leave out. If you get ten leads a month and answer all of them within the hour, you don't need one. If your intake is a pile of sticky notes and a shared inbox nobody owns, an agent on top of that will produce a faster mess, and the first job is cleaning up the process. And if the work is mostly judgment, like quoting complicated projects, the right answer may be a better checklist, not software.
Agents also need watching in the first couple of weeks. You read what it sent, correct what it got wrong, and tighten the limits. Plan on that, or pay someone to do it.
How I'd choose one
Start with the place something gets dropped. Not the task you find tedious, the one that actually costs you customers. Put a rough dollar figure on it. Check whether it's repetitive and rule-bound. If it is, that's your first agent, and it should be one agent doing one job before you add anything else.
That's also how I run the audit: find the one gap, size it, and say plainly if the answer is "your current tools can already do this."
If you want to talk through what this would look like for your business, the audit is free and takes 30 minutes.