Concepts
What can AI agents do for a small business?
The useful question is not what AI agents can do — it is which jobs are worth handing over, at what cost, and which ones will quietly go wrong. This page answers those three, with the real numbers.
One thing to be straight about first: these are capabilities, not case studies. I built this product and I am not going to describe a customer outcome I do not have. What follows is what the software does and what it costs, so you can judge whether it fits your work.
Jobs agents are genuinely good at
| Job | Why it suits an agent | Tools it needs |
|---|---|---|
| Read incoming messages, classify and route them | Judgment on varied text, which rules handle badly | None, or Slack/email to notify |
| Draft a reply for a person to approve | Drafting is cheap; sending is the risky part, so keep the human | None, or send_email |
| Turn forms and emails into structured records | Structured JSON output makes the result machine-readable | http_request to your system |
| Watch something on a schedule and speak up | Runs without you; costs a fraction of a cent per check | http_request plus Slack, Discord or Telegram |
| Book and confirm appointments | Reading intent from a message, then acting on a calendar | Google Calendar |
What it costs in real money
Agents are billed in tokens, so the cost tracks the size of the job. A worked support-ticket example comes to $0.0144; a real logged run that also escalated to a human $0.0204; a check that fails early $0.0024. The dearest agents shipped here reach about 11 cents.
So a workflow checking something every fifteen minutes is roughly 96 runs a day — a few cents. Source: two runs costed step by step, readable without an account.
The free plan needs no card and covers 100 runs or $0.50 of usage a day, whichever comes first. That is enough to find out whether an agent does your job properly before paying anyone.
What to keep away from an agent, for now
Anything that moves money without review
Issuing a refund, paying an invoice, cancelling a subscription. Have the agent prepare it and a person approve it. The cost of a wrong draft is a second of reading; the cost of a wrong payment is not.
Work where the same input must give the same output
Agents are not repeatable by nature — the same ticket can produce differently worded replies. Fine for prose, wrong for anything another system parses. Use structured JSON output, or a deterministic step.
Anything nobody will ever look at
This is the real risk for a small team, and it is the failure I hit myself. Source: a run of mine looked up two orders, got 403 twice while the tool result said "ok": true, and wrote to the customer as though it had read them. It finished COMPLETED and cost $0.0204. A person reading only the status would have seen a success.
If nobody would notice the agent being wrong for a month, do not automate that job yet. Automate the one where somebody notices the same day.
A sensible first month
- Pick one job that happens several times a week and is annoying rather than critical.
- Build the agent so it drafts and a person sends. Read the first twenty runs.
- Only then give it a tool that acts on its own, and only that one tool.
- Put it on a schedule once you have seen it be right without you.
Next
- Build the first one — see Get started in 3 steps.
- Decide whether an agent is even the right tool — see AI agents vs automation tools.