How to Choose an AI Agent for Your Small Business's Repetitive Work

"Everyone is selling me an AI agent. How do I tell which one is actually worth it?" If your business carries the same work every day, you have probably asked this. Knowing how to choose an AI agent for a small business comes down to three questions most sales demos skip: who makes the decisions inside it, how much of every run goes through the AI model, and whether you pay for the work it does or for a subscription that keeps charging in a quiet month. But first, a word on what an agent actually is, because most demos make it look far smaller than it is.

An agent is not just a chatbot

When people hear "agent", they often picture a chat window that answers customers. Answering inquiries is one job an agent can do, but it is only one. An agent is software that takes over a piece of repeating work and carries it to the end, whether or not anyone is chatting with it.

• Insurance claims. A business preparing claims for insurance companies has no chat at all. There is a file, a set of steps that must happen on it, and a submission at the end. The client may never appear in the flow.
• Invoices. A fully internal agent checks incoming invoices, issues the outgoing ones and sends them, while nobody outside the office ever talks to it.
• Listings and ads. An agent publishes a business's listings or ads on the sites where it advertises and keeps them up to date.
• Triage. At a clinic, an agent sorts incoming referrals, faxes and lab documents and routes each one to the right doctor's queue.
• Professional software. An agent can work inside the tools a team already uses, such as a game studio's production software or an office's drafting and design software, carrying out repeating production steps.
• Client-facing work. Booking appointments, chasing missing documents for an immigration file, following up on a repair quote, answering a renter's question.

So when you shop for an agent, think about the whole job, not just the conversation. The three questions below apply to every one of these.

Who makes the decisions: your rules or the model

Every job has firm rules. A claim cannot be submitted until certain documents are in the file. An invoice above a set amount needs the owner's approval. A repair company asked "Can you do it cheaper if I book today?" does not give discounts, and price talks go to a person. Only certain services are offered, and certain cases always need the owner.

When the AI model makes those calls, the result can drift from day to day. It may "helpfully" offer a lower price or send a file that was not ready. And it can change overnight when the vendor updates or swaps the model. When those decisions run as fixed rules against the business's own lists, with no model involved, the outcome is the same every time and you can see why it happened. The AI is left with what it is genuinely good at: reading a document or a message and writing text from the business's own information.

That is the first sign of a good agent: logic decides, AI reads and writes.

Where the AI spends your money

Many agents "look at" a web page with an AI model: a portal, a supplier's site, a listing, a status page. To do that, the whole page goes to the model. In one measurement, a single ordinary public web page came to roughly 499,000 characters of content after cleanup. Even a compact snapshot of it was around 33,000. That cost is paid again on every single run.

A fixed reading rule, written once, does the same job each time for almost nothing. The expensive model work belongs in setting the agent up, not in every run. So "how much of this runs on the model, every time?" is the first cost question to put to any vendor. If the agent ends up costing close to an assistant's wage, it does not pay for itself.

Pay for the work, not for the subscription

Many agents are sold as a monthly subscription or per-seat fee: the same price whether the agent finished five hundred tasks or almost none. Often that fee covers only the tool, while the setup, the gaps and the follow-through stay with you.

When you hire a person, you pay for someone who does the job. Buy an agent the same way: it takes over the work end to end, and its cost follows what it finishes, such as claims submitted, invoices sent, listings published or appointments booked. A busy month costs more because more got done; a quiet month costs less. Ask: "If I have half the work next month, does my bill go down?" And watch for high monthly minimums, which quietly turn "pay per outcome" back into a subscription.

How to choose an AI agent for your small business: a checklist

1. Name the outcome first. "Every claim file completed and submitted", "every invoice issued on time", not "an AI assistant".
2. It does the actual job, and is paid for the job. The price follows work done, not a flat fee for a tool that covers "up to here".
3. Cost far below a person. Ask how much of each run goes through the model.
4. Do a 30-day cost model at your own volume. Your real monthly number of files, invoices, messages or bookings, plus platform fee, usage, per-outcome charges and seats. Check how much is fixed no matter how much work is done.
5. Rules decide, AI reads and writes. Which decisions does the model make on its own? Those should be your rules.
6. Same input, same result. What changes when the vendor's model changes?
7. Tailored over generic. An agent shaped around how you actually work usually costs less and gives you more control.
8. Approval and handoff. Where can you approve, stop or take over? Is anything binding gated behind a person?
9. Logs you can read. Every action, with the rule behind it, not just "the AI decided".
10. Setup burden and repairs. Who configures it, and who fixes it when a portal or form changes?
11. Your data, your exit. Records stay yours and exportable, on terms of 12 months or less. For health, legal or immigration files, ask where data is stored and who can see it.
12. Start narrow. One repetitive job, measured for a month. Then expand.

If you remember one line, make it this one: let logic decide, let AI read and write, and pay for work done, like a person you hired, not a subscription. Take this list into your next vendor conversation and see how many answers come back plain and specific.