AI agents for small businesses: six concrete cases, what they cost, how to build one
An AI agent is a program that runs a business task end to end with an AI model: it reads an input, decides within a set frame, acts in your tools and hands the rest to a human. For small businesses: six examples, what they cost, and how to build one that holds up.
Every week, someone in your company opens the same emails, copies the same information from one tool to another, prepares the same documents from the same data. That work has no value; it has a cost, and it has a replacement.
An AI agent is a program that runs a business task end to end with an AI model: it receives an input (an email, a form, a file), decides within a set frame, acts in your tools and hands to a human whatever it must not decide alone. It is not a chatbot you question.
Which AI agents for a small business? Six concrete cases
Six examples of tasks an agent can take on in a small business, given as illustrations. Each is built as an ordinary request in the queue, not as a project. They are ranked by difficulty, and the first is almost always the right starting point.
01
Triage and pre-answer support
The agent reads every incoming email, classifies it (order tracking, billing, bug, sales), answers alone the questions whose answer is in your data (“where is my order”), and forwards the rest to a human with a summary and a proposed reply. For example, order tracking questions, often the most frequent, no longer need to be opened by a person.
02
Qualify inbound requests
A contact or quote form comes in. The agent enriches it (company, size, sector), checks it is not an existing customer, emails the two missing questions, and creates the record in the CRM with a score. Sales only call the requests that are ready.
03
Prepare quotes
From a qualified request and your price grid, the agent prepares a draft quote in your tool, with the lines, options and usual terms. A human reviews, adjusts, sends. For example, a quote that took an hour to prepare now only needs a review, and the usual lines are no longer forgotten.
04
Chase unpaid invoices
The agent watches due dates, sends reminders on your schedule and in your tone, adapts the message to the customer's history, and flags to a human the cases outside the frame (a dispute, a strategic customer, a high amount). It never blocks an account on its own.
05
Sync and enter data
Website orders into accounting, customer records between two tools, email attachments into the right folder. The agent reads formats that classic connectors do not understand (a PDF, a free-form email) and enters data where it belongs, with a log of what it did.
06
Summarise and monitor
Every morning, a summary of what happened: open tickets, customer reviews, mentions of your brand, moves at three competitors. Written, sorted by importance, with links. A two-minute read instead of an hour of tabs.
03
How much does an AI agent cost?
Two cost items, to keep separate. Building: it is development, so at Figue one or more requests in your queue, included in the subscription. A first triage agent is scoped and built in a few requests; a quoting agent, which touches your business tools, takes more. Running: the calls to the AI model, billed by the provider, in your name, like any third-party cost.
On that second item, the order of magnitude often surprises in the right direction. Based on the model providers' public price lists, checked on 22 September 2026, reading and classifying an email costs a fraction of a cent with a model suited to sorting; a support agent handling a few hundred emails a day generally spends less per month than a ticketing tool subscription. What costs is badly scoped agents that reread the same document ten times, or call the most expensive model for a sorting task.
How do you build an agent that holds up in production?
The difference between an impressive demo and an agent that has been running for six months without anyone thinking about it comes down to five steps, only one of which is AI.
That is exactly how a request goes at Figue: scoping with you, build by an engineer with AI to speed things up, review, preview, approval, production. An agent is a deliverable like any other, with a little more attention on the guardrails.
Not with the most ambitious case. With the one that costs the most boring hours today, whose rules are clear, and where a mistake would not cost much. In most small businesses, that is support triage or qualifying inbound requests.
Drop it like any other request: who does this task today, how long it takes, what makes it painful, and how you will know the agent is doing the job. We scope together, build, observe, widen. The second agent is always easier than the first: the data is connected, the guardrails are understood, the trust is there.
“The best first agent is the one everyone is happy to get rid of the task for.”
What is the difference between an AI agent and classic automation?
Automation follows fixed rules on structured data: if the field equals this, do that. An agent reads and understands free-form inputs (an email, a PDF, a conversation) and decides within a frame. We often combine both: the agent reads and classifies, the automation executes.
Can an agent get it wrong?
Yes, like an employee in their first month. That is why it has guardrails, a log, and an observation period before acting alone. The remaining errors are rare and visible; the ones that would cost a lot are out of its scope by design.
Does my data go to the model provider?
The agent sends the model what it needs for the task, no more. In production, it goes through your account with the chosen provider, in your name and under its terms: read them, especially on the use of data for training. We choose with you the provider and the level of data sent; some cases are handled with anonymised data.
How long for a first agent?
A support triage agent is scoped, built and observed in a few requests in the queue; their duration depends on how clean your data is and how fast you answer during scoping. We do not promise a date, we show progress.
Do I need a Figue subscription to get an agent?
An agent is a development request like any other, so yes, it fits in Figue Unlimited, with the scoping, review and maintenance that go with it. Model call costs stay in your name, like any third-party cost.
