An AI agent for small business is started with a single task that is repeated daily in correspondence, and with a single channel where clients write the most. Next, you need to decide what the agent resolves by itself and what it always passes on to a person, and enable it gradually: first at night, then at weekends, and only after that during the day. This procedure makes it possible to see mistakes on a few dozen dialogues while they cost nothing.
Let's break this down using the example of a small leather shoe workshop. There are four people working there: the owner, two craftsmen, and a manager who runs the Instagram page and handles orders in the CRM. Half of the buyers write in direct messages, while the rest come from the website. They ask the same things: what size to get for a foot length of 24.5 cm, how long manufacturing takes, and whether a different sole colour is possible. The manager replies until seven in the evening, but messages keep pouring in until midnight.
Contents
- First week – sort through your own correspondence
- What task should be given to an AI agent for a small business first
- What the agent decides itself and where it calls upon a human
- How to avoid paying too much for your text message allowance and tariff right from the start
- Knowledge base from your best answers
- The switch-on schedule covers the night, the weekend and then the working day
- Who manages the agent after launch
- Frequently asked questions
Section: AI for business
First week – sort through your own correspondence
The most common mistake at the start is to begin with choosing a platform. First, open your own chat history. Take the last hundred dialogues from Instagram and the website and sort them into piles: questions about sizing, production time, payment and delivery, custom orders, and complaints.
For the workshop, the picture usually turns out uneven. Two or three topics take up most of the dialogue, and the responses in them are almost identical. The rest is varied and requires a master. At the same time, indicate what time each message arrived and whether the conversation ended with an order.
Small businesses are already using generative AI on a mass scale, at least in the US. According to a U.S. Chamber of Commerce report (August 2025, research with Teneo), 58 % of small businesses used generative AI in 2025. In 2024, there were 40 % of these, compared to just 23 % in 2023. This refers to generative AI for any task, without breaking it down into client-facing and internal ones. An AI agent for small businesses that talks to customers on its own is the next step. It is worth preparing for it using your own figures, and a hundred dialogues will provide them in a single evening.
What task should be given to an AI agent for a small business first
The first task for an AI agent in a small business has four characteristics. It is repeated every day. The answer to it is already written down somewhere: on the website, in a size chart, in a manager's templates. The mistake is easy to spot and fix. And the result can be counted: how many dialogues were closed, how many orders were created.
In the workshop, questions about size and production time fall into this category. The size chart is linked to foot length, the timeframe depends on the model and the queue, and both answers are given by the manager ten times a day. A custom order based on a client's sketch does not fit here. That requires a master craftsman to look at the photo and say whether such a last can be made.
A standard chatbot for small business with “Sizes” and “Delivery” buttons would resolve some of these issues. The problem is that a female customer writes in free text: “I have a wide foot, 25 cm, for winter with a thick sock, what should I get?”. Buttons do not answer that. The agent reads the question in full, looks for the answer in your data, and can take the conversation all the way to an order. More details on, what is the difference between an AI agent and a regular bot, we wrote separately.
What the agent decides itself and where it calls upon a human
Agent permissions are written before launch, on a single page. Without this, every non-standard conversation turns into a lottery: either the agent makes things up, or drags a human into something they could have handled themselves. For the workshop, the breakdown looks like this.
| The situation in the correspondence | The agent themselves | Passes to a person |
|---|---|---|
| Price, availability of a ready-made pair, colour | matches the catalogue | – |
| Size by foot length | selects by grid, specifies width | if the foot is beyond the net |
| Production time and delivery | names by model and conditions | if the female client asks to hurry up |
| Checkout | gathers data and creates an order in the CRM | – |
| Discount, negotiation, exchange | – | immediately, with the whole history |
| Bespoke tailoring from a sketch | accepts photos and wishes | for the master's evaluation |
The second column determines how much work the agent will take off the female manager. The third, how peacefully you will sleep. At the start, it is better to hand over more to the human than necessary. It is easy to expand an agent's rights once you already see how they are coping. Regaining the trust of a female client to whom the agent promised a discount without your knowledge is considerably harder.
Such a table is worth showing to the whole team. The fitters need to know that the photo with the sketch will be sent to them with a note from the agent, and that they must reply to it themselves.
The handover must include context. Marichka AI, the AI sales manager, hands over a complex query to the operator along with the chat history, so the manager doesn't have to ask in the morning which model and what size. If the customer already named everything except the sole colour in the 23:40 chat, only one message remains for the morning.
How to avoid paying too much for your text message allowance and tariff right from the start
The price of most services depends on the number of messages, so calculate your volume before choosing a pricing plan. Take an ordinary week without promotions, add up the incoming messages from all the channels you plan to connect, and multiply by four. Separately, check with the service how it counts: only customer messages or including agent replies as well.
When an AI agent for small business works only in Instagram at the initial stage, a junior tariff plan may suffice. In Marichka AI, it is designed for 1,000 messages per month. If there are not many night and weekend dialogues, the volume for the first month will be less than a full month, since the agent does not yet work during the day.
The first 7 days in Marichka AI are free. That's enough to get through the first two stages of the activation guide and see the agent responding to your real customers at one o'clock in the morning.
Knowledge base from your best answers
The agent knows precisely what you have given it. Therefore, the knowledge base is built from answers that are already being sold, and the description from the main page of the website is not enough for this. The main page is written for someone seeing the brand for the first time. A female buyer in direct messages has already chosen a model and is asking about her foot.
For the workshop, this is a size chart with foot length in centimetres and advice for wide feet. Production times for each model and an explanation of what they depend on. Care instructions for smooth leather and suede. Prepayment and delivery terms, and exchange rules, which read identically on the website and in the profile bio. And twenty to thirty real questions from that hundred dialogues, along with the manager's answers, that resulted in an order.
Marichka AI takes data from website scanning, PDF and Excel files, and automatically updates prices and inventory. It responds exclusively using this database. If the database does not contain the production lead time for a new model, the agent will not make it up. The query will be escalated to a human. That is why gaps in the database become apparent on the very first night: these are the conversations that get passed on to the manager.
Check the database yourself before letting clients in. Ask the agent the ten most difficult questions from your correspondence, including the one where the female client is between sizes 37 and 38. If three answers out of ten have to be corrected, update the database and repeat.
The switch-on schedule covers the night, the weekend and then the working day
The safest place to take your first steps with an AI agent is where nobody is responding right now. An agent's mistake at two in the morning costs less than silence until the morning, and during the day there is a manager nearby who would have answered anyway. The sequence for the workshop is as follows:
- A week on nights from 19:00 to 09:00 in a single channel. The owner reads all of the agent's chats every morning.
- Next week, extra days off are being added. Only dialogues that have been handed over to a person and those that ended in an order will be read now.
- The agent comes on during the day, and the manager handles only forwarded conversations and individual orders.
- A second channel (website or Telegram) is connected using the same knowledge base once the first month has passed without any surprises.
At each stage there is one rule: do not move on while the dialogues still contain answers you would be ashamed of. How to break down a test week day by day, we discussed in the article with daily test week plan.
Who manages the agent after launch
In a small company, a single person is responsible for the agent, and ideally it is the person who used to handle clients. In the workshop, that is the female manager. Once a week, she reviews the dialogues forwarded to the person and adds any missing answers to the database. When a new model comes out or deadlines change, updating the database becomes part of the model launch, just like photos for Instagram. If no one takes this work on, an AI agent for small business will start answering based on the old price list in two to three months, and female clients will notice that before you do.
In a month, compare it with that first week when you were counting messages. How many night-time chats now end in an order. How many conversations the agent brought to a close by itself and how much it handed over. How much time the manager has freed up for individual orders, which are the ones that bring the workshop the most. An AI agent for small business proves its worth when these figures are visible in your CRM, without any generic promises.
AI for small business does not require a large project: a single task, a single channel, documented agent permissions and a gradual rollout cover the first month. Start today with messaging. Open your direct messages, scroll through the last hundred chats and note how many of them came in after seven in the evening and concerned sizing or delivery times.
Ready to connect Marichka to your channel?
Marichka AI will help you automate communication, reply to clients 24/7 and increase sales.
Frequently asked questions
How much does an AI agent cost for a small business?
Such services operate on a monthly subscription, and the price depends on the number of messages. With Marichka AI, the lowest tier is designed for 1,000 messages per month, and the first 7 days are free. Before choosing a plan, calculate how many messages you receive in a typical week and multiply that by four.
Do you need a programmer to run an AI agent?
For a typical shop or workshop, no. The agent is set up in the dashboard without code: a catalogue is uploaded or a link to the website is provided, a channel is connected, and the responses are checked. Marichka AI has an 8-step video guide for this. A programmer might only be needed for a custom integration with a proprietary accounting system.
What happens if the agent doesn't know the answer to a customer's question?
He has to hand the conversation over to a human. Marichka AI responds only from the knowledge base, so for questions outside of it, she does not invent an answer, but passes the dialogue to a manager along with the entire chat history. Such handovers should be reviewed weekly: each of them suggests what to add to the base.
How long does it take to launch an agent in a single channel?
Technical connection is faster if the catalogue and delivery terms are already gathered in one place. Verification with real customers takes longer: it is wise to give the agent at least a week of night and weekend dialogues before turning it on during the day.
