An AI agent for business automation is best suited to tasks where customers write in their own words every day, and the answer is already in your data: product selection, enquiries about availability and price, and placing orders based on correspondence. Recurring tasks that always follow the same pattern are handled by a standard rule in the CRM. Decisions regarding payment and warranties remain with a human.
Let’s look at an example. There’s a small bike shop in Lviv with a repair workshop; let’s call it “Pedal”. It sells via Instagram and its website, and manages orders through a CRM system. The staff consists of two sales assistants and a mechanic. In April, when the city switches to cycling, their Direct messages fill up in a single evening: what frame size to choose, whether this cassette will fit an old rear derailleur, and when to collect a bike from the repair shop. The owner wants to put someone in charge of this, and that “someone” could be an agent. The question is what to entrust to them and what not to.
Contents
- An agent, a rule in the CRM and a person: three different actors
- Five questions to ask about a task before assigning it to an AI agent
- What should you tell an AI agent first when automating your business?
- Tasks that are best not to assign to an agent at the outset
- Business automation using an AI agent after the first month
- Frequently asked questions
Section: AI for business
An agent, a rule in the CRM and a person: three different actors
Automation in small businesses involves three key players, and confusion between them costs money. A rule in a CRM system is triggered by an event: a status change, a payment received, or a deadline passed. An AI agent reads messages, however they are phrased, searches for a response in your data and continues the conversation. A human intervenes when money, reputation or an unusual situation is at stake.
Take two messages received by “Pedal” in the space of one evening. The first: “Is the bike ready yet?” The answer depends on a single field in the repair card. Once the mechanic has marked the card as “ready”, the CRM can automatically send a message to the customer – no AI is needed here.
Secondly: “I’m 172 tall, have long legs, cycle around town and a bit in the woods, and have a budget of up to twenty-five thousand – what would you recommend?” There’s no basis from which to derive an answer here. You need to understand the enquiry, cross-reference it with the size chart and product range, clarify a couple of details and suggest two or three models currently in stock. This is precisely the sort of task that AI in business automation handles better than any button could.
If you send the first message to an agent, you’ll use up your message limit when a simple trigger would have sufficed. If you try to handle the second one with a rule, you’ll end up with a tree of buttons, which the customer will navigate away from on the third step. For more details on this, how to tell an AI agent apart from a normal bot, we wrote separately.
Five questions to ask about a task before assigning it to an AI agent
List everything the team does manually during the week, and put each task through a five-question checklist. A task that receives five “yes” answers goes to an agent first. Those with three or four “yes” answers can wait for the second round.
- Does it happen every day, at least a few times?
- Does the client phrase it in their own words, without a set format?
- Is the answer already in your records: the catalogue, the size chart, the delivery terms, or the workshop price list?
- Will an agent’s mistake be inexpensive, and will it be easy to spot and rectify?
- Can the results be measured: in terms of conversations, orders or consultants’ hours?
The second question filters out tasks for CRM rules: if the input is always the same (event, status, date), an agent is not required. The fourth question filters out tasks for a human. The cost of a mistake regarding the frame size is a single clarification the following day, whilst a false promise of a free warranty repair will result in you having to pay for the repair yourself or a customer who feels cheated.
The fifth question is often overlooked, and that’s a shame. A task whose results aren’t immediately apparent in figures will, a month later, spark a debate over whether it “helps or not”, and there will be no clear answer.
Sometimes, two tasks will both receive five “yes” votes at the same time, and it’s not clear which one to pass on to the AI agent first. In that case, look at the time. Choose the one where more messages come in after the shop has closed: the agent isn’t standing in for anyone there; they’re covering the hours when no one else is available, and any mistakes made during those hours are easier to deal with, as the alternative would be silence until morning. During the day, consultants carry on working as usual and review the night-time conversations in the morning.
What should you tell an AI agent first when automating your business?
In “Pedali”, three things score five “yes” votes, and all of them relate to conversations with customers. This aligns with how large companies utilise agents: according to a Google Cloud study (September 2025, 3,466 executives from 24 countries), 52 per cent of respondents say their organisation is already actively using AI agents, and they are most often tasked with customer service and managing the customer experience – 49 per cent gave this as their response.
Selecting the right frame size and checking spare parts compatibility
This is the most common question asked via direct message at a cycle shop, and the one least suited to a drop-down menu. The customer states their height, sometimes their leg length, riding style and budget, often in a single message full of mistakes. The sales assistant cross-references this with the manufacturer’s size chart and what’s in stock.
It’s a similar story with spare parts. “Will an 11-speed chain fit my 10-speed cassette?” The answer can be found in the product descriptions, provided compatibility is specified there. The agent needs this specific table, and it should be finalised before launch, as without it they’ll have to pass the query on to someone else.
April evenings and weekends
The second task is the same consultation, but at times when there are no consultants available. Marichka AI, the AI sales manager, responds within 10–30 seconds at any time of day or night. It gathers product information by scanning the website and PDF or Excel files, and updates prices and stock levels itself. It responds solely based on this database, so it won’t invent a model that isn’t in the catalogue.
For “Pedali”, it’s April: at 10.30 pm, someone decides to buy a bike ahead of the May bank holidays and sends a direct message. By the next morning, they might already be in a different shop.
Orders from correspondence straight into the CRM
Once the customer has made up their mind, the agent collects their name, telephone number and delivery method, and creates an order in the CRM. In the morning, the consultant sees a completed customer profile containing all the correspondence. This third task is a logical continuation of the first, so they are linked together: a conversation that ends with “I’ll take it” shouldn’t have to wait until the morning.
If there are still unanswered questions from the previous evening in your shop before opening, that’s exactly where you should start the test.
Tasks that are best not to assign to an agent at the outset
Simply being able to assign a task to an agent doesn’t solve anything in itself. An AI agent for business automation has a pricing structure based on the number of messages and requires a knowledge base, so there’s no point in giving it work that can be handled by a rule already in your CRM. Here are two groups that, at “Pedala”, remain outside the system, at least in the first few months.
Repair status and maintenance reminders: a set of rules is enough
The message “Your bike is ready; you can collect it by 19:00” is triggered by a status change in the repair record. A reminder about seasonal servicing six months after purchase is triggered by the date. Both work as simple rules in a CRM or email marketing service, and are more reliable there: they do not depend on how an agent interprets the context.
An agent will be needed when a customer responds to such a reminder by asking: “What does the service pack include, and how much does it cost for a mountain bike?” At that point, the conversation moves from the trigger to a free-form discussion, and the agent refers to the workshop’s price list in the knowledge base.
Warranty, returns, haggling: only people
The customer says that the frame cracked after a month of riding. We need to look at the photos, check the manufacturer’s terms and conditions, and decide whether this is covered by the warranty. Making a mistake either way will cost money. Wholesale discounts for cycling clubs and price negotiations for second-hand bikes should also be discussed here.
In conversations like these, an agent acts as a data collector: they’ll ask for a photo, the order number and the date of purchase, and pass all this information on to a technician or the owner. At Marichka AI, complex enquiries are forwarded to an operator along with the chat history, so the operator doesn’t have to ask the customer to repeat what they’ve already written.
Business automation using an AI agent after the first month
Once the first task is running smoothly without needing daily adjustments, you can move on to the next one. The procedure for “Pedals” is as follows.
| Task | Who is responsible for | When to connect |
|---|---|---|
| Choosing the right size and model, availability, price | AI agent | the first week, starting with evenings and weekends |
| Orders from correspondence in the CRM | AI agent | including a consultation |
| Repair status, maintenance reminders | rule in the CRM | at any time, regardless of the agent |
| Enquiries about prices and booking an appointment at the workshop | AI agent | The second month that the workshop’s price list has been in the knowledge base |
| Warranty, returns, wholesale, price negotiable | human, agent is collecting data | always |
Large companies have seen a noticeable increase in the number of agents over the past year. According to the KPMG Global AI Pulse survey for the third quarter of 2026 (2,131 executives from 20 countries), the proportion of companies where employees actively use AI agents has risen from 25 per cent to 34 per cent since the start of the year. The same survey also reveals a less encouraging figure: Only 12 per cent of companies consistently compare the benefits achieved with their AI expenditure.
A small shop doesn’t need a separate team to carry out this calculation. Before you start, make a note of three figures for a typical week: how many messages arrived after closing time, how many of those were only replied to in the morning, and how many orders from those conversations appeared in the CRM. After a month, note down the same figures again and compare them with the service fee. Regarding the total Priorities for automation in a small shop There is a separate breakdown, but here we are only talking about the part that the agent takes on, and in the shop this primarily involves correspondence.
An AI agent for business automation pays for itself on a single, quantifiable task, and only then should you assign it a second one. This evening, open your Direct messages, list the last twenty questions from customers and run each one through the five questions above. The one that scores five “yes” answers will be the agent’s first task.
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 many tasks should an AI agent be given at the start?
One, or at most two, simple tasks – for example, advising on a product and processing an order within the same conversation. That way, within a week or two, you’ll see where the agent is going wrong and be able to measure the results. Add new tasks once the first one is running smoothly without needing any morning corrections.
Do you need a programmer to set up an AI agent?
Not for a typical shop. The chatbot is set up in the dashboard: you provide a link to the website or a file containing the catalogue, connect the channel and check the responses. Marichka AI has an 8-step video tutorial for this. You may only need a developer if your accounting system does not have a ready-made connection to the service.
When is a standard rule in a CRM better than an AI agent?
When an action is triggered by an event and is always the same: the order status has changed, six months have passed since the purchase, or payment has been received. The rule will work the same way a thousand times and won’t use up the agent’s message limit. An agent is needed when a customer writes in their own words and the answer needs to be found in your data.
What should you do if an agent has given the customer the wrong answer?
First, identify what was missing from the knowledge base: an out-of-date price, a missing size chart, or an unaddressed scenario. Add this information and ask the agent the same question again. If the error relates to money or a guarantee, add this topic to the list of issues that the agent should immediately pass on to the customer.
How much does an AI agent cost for a small business?
Such services usually charge a monthly fee, which depends on the number of messages. With Marichka AI, the entry-level plan costs €49 per month for 1,000 messages, and the first 7 days are free. Before you choose, count the number of incoming messages you receive in a typical week and multiply that by four.
