AI agents for small business: when they pay off

Agent, AI step or plain automation

The question is who decides the next step: rules you wrote, or the model. The more the model decides, the more checks you need around it.

Where an agent earns its cost

Agents pay off when the work needs reading, judgement and a choice between several next steps, and when a person can review what matters. Typical use cases in a small business:

• First replies to customer questions, using order history and your policies, handing over to a person when unsure.
• Sorting an inbox or contact form: what the request is, who should handle it, and a draft answer.
• Questions about your own documents and procedures for the team, with links to the source.
• Researching a new lead or supplier before someone calls them.
• Checking a document against your rules and listing what is missing.

Where automation is the better choice

If the steps are always the same, a model deciding them adds cost and risk without adding value. Keep it as a plain automation when the process moves money, when every record must follow the same rule, when you need an exact audit trail, or when the input is already structured, such as a webhook from Stripe or a form with fixed fields.

A common good design is mixed: fixed rules for the flow, and one AI step where a person used to read something. For invoice entry, for example, the model extracts the fields, plain checks confirm the totals add up and the supplier exists, and anything that fails goes to a review list.

What makes an agent safe to run

Before an agent touches real customers or data, it should have these in place:

• Only the tools it needs, read-only by default.
• Human approval for anything that sends, pays, deletes or changes a price.
• A log of every step, so you can see why it did what it did.
• A test set of 30–50 real past cases with known right answers, run before launch and after every change.
• A cost limit per task and a clear hand-over to a person.

What it costs

There are two costs: building it and running it. Running costs are mostly model usage, billed per token by the provider. For a task that reads a few pages and writes a reply, that is usually cents rather than dollars, but agents that loop through many steps cost more, which is why a per-task limit matters.

With us, one AI workflow or agent in production usually costs $6,000–15,000 to build and takes two to four weeks, including the test set, logging and review steps. Model usage is billed to your own account, and we estimate it before the build.

How to start

Pick one process that costs real time today. Write down how many items it handles a month, how long each takes and how often mistakes happen. Collect 30–50 real examples with the right outcome. Start with an AI step and human approval, measure it against those numbers, and give it more freedom only where the results hold up.

Frequently asked questions