AI agent for small business: where to start

The owner of a small company usually knows an agent would help, but not where to begin. The list of ideas is long: client replies, reports, purchasing, email, advertising. You cannot take everything at once, and picking at random means disappointment a month later.

It is better to start with one process and test it on your own data. Below are three signs of a good first process, some examples, and a way to tell whether it worked.

A good candidate answers “yes” to all three questions.

1. It repeats. The same thing is done every day or every week: collecting numbers, answering the same questions, checking new listings. One-off tasks are not a good start.
2. It can be measured. There is time it takes or money that passes through it. If the result cannot be compared with anything, you cannot tell whether it helped.
3. A person confirms the risky steps. The agent prepares, suggests and explains. An email, a message in someone else's chat or a change to data goes out only after your “yes”.

The third point removes the main fear: that the agent will do something without asking. In CapyAgent the confirmation is built in, and each approval is one-time.

Below are four tasks the agent does today. Each one can be a first step.

• Morning digest for Bitrix24. At 8:00 a message arrives in Telegram: who was not called back yesterday, which requests have been waiting for more than a day, where the source is not recorded. Details are in the article about the morning digest. In Bitrix24 the agent only reads.
• Tender search. On weekdays the agent checks your filters in Tenderplan, reads the first documents of the procurement and sends the cards to Telegram. The module is paid, price on request. How it works is described in the article about tender search.
• Answers to routine questions. The agent lives in Telegram: an employee writes to it in plain language and gets an answer. It remembers people and tasks, so the context does not have to be explained again every time.
• Reports and notes. A voice message or a meeting recording is turned into text, and the agent compiles notes from it. It prepares a regular report on a schedule.

Other tasks grow from the same logic. For example, short-term rental: prices, guest replies, cleaning. The case is described in the article about an AI manager for short-term rentals. Some of its steps are still planned, and they are described that way there.

Take a sheet of paper and list everything the company does by hand more than once a week. For each line, answer three questions:

• how many hours a week it takes;
• what happens if the agent makes a mistake, and who will notice;
• where the agent will get its data: from the CRM, email, a spreadsheet or a chat.

Start with the line where the hours are many, the cost of a mistake is small, and the data already sits in one place. Usually that is a digest or sorting incoming requests, not correspondence with clients on your behalf.

If you are torn between two options, choose the one where the agent only reads and sends you the result. It is easier to go back there and fix things.

Agree on the numbers before launch, not after. Three are enough for a small business.

• Time on correspondence and data collection. Measure one week before the agent and one week after. A rough count of hours in a table is enough.
• Response speed. How long a client or an employee waits for an answer: in minutes and hours, not “it got faster”.
• Money. How many requests were not lost, how many purchases were found, and what the agent itself costs. The cloud is 1,590 RUB a month, the model about 2,000 RUB, together about 3,600 RUB.

If after a month the hours have not dropped and the money looks the same, the process was chosen badly. That is a normal outcome of a trial: take the next line from the list.

Once the first process works, add a neighboring one. The CRM digest shows which requests get lost, and the next step can be to have the agent sort through email. Tender search gets a full analysis of a procurement behind a button.

Expand one task at a time and check the same three numbers each time. That shows which step gave a result and which one only added messages to the chat.

You can start for free: the app for Windows, macOS and Linux downloads at no cost, and the server install is a single command. For yourself and within your own company the program is free. The cloud is needed when results must arrive even if the computer is switched off.

Pick one process by the three signs and test it on your own data. Download the app if you want to try it yourself, or look at the cloud if you need results without the computer switched on. The other tasks for a company are collected on the page AI agent for business.