An AI business assistant is a large language model-based application that understands messages written in any words and performs a task at your request or a customer's request: it answers, selects a product, and writes text. Unlike a chatbot, it does not need a script. A button-based bot guides a person through branches that someone has drawn in advance, whereas an AI assistant reads the question as a whole and looks for the answer in your data.
The difference is easiest to see with a real-life example. Imagine an online shop for dog and cat food and accessories. Sales go through Instagram and a website, and you manage orders in a CRM. ChatGPT is already writing posts, and you once had a button-based bot on Telegram that you abandoned: customers were writing outside the buttons.
Contents
- One direct message and two different replies
- What is called an AI assistant and where its work ends
- An AI assistant for yourself and your clients
- How an AI assistant differs from a chatbot when a customer writes outside of the buttons
- How does the assistant know your product range and why does it sometimes make things up
- Five things to gather before connecting
- How much does an AI assistant cost for business and how to test it in a week
- Frequently asked questions
Section: AI for business
One direct message and two different replies
Tuesday, 21:50. A female customer writes in direct messages: “Good evening! The Maltipoo puppy is 4 months old and has a chicken allergy. What dry food would you recommend and will you be shipping tomorrow?”
The button bot responds to this menu: “Select section: Catalogue / Delivery / Payment / Contact manager”. A female customer clicks “Catalogue”, receives a link to a website with a hundred feeds and a filter that lacks a “chicken-free” option. The second question, about tomorrow's dispatch, the bot simply didn't see. After that, everything depends on her patience, and at ten o'clock at night patience wears thin.
A language model assistant reads the same message and sees three conditions in it: puppy, small breed, and chicken allergy. It searches the catalogue for puppy food without chicken protein and suggests two options, for example lamb and salmon, with the price and pack size. It answers regarding delivery based on your delivery rules, if they are recorded in the knowledge base. At the end, it asks whether to place the order. Marichka AI claims a response time of 10–30 seconds for its assistant. The button bot replied even faster, but only the second response leads to a purchase.
Mentally scroll through your conversations over the past week. How many of them started with a message that fits on the “Delivery” button?
What is called an AI assistant and where its work ends
In its guide, IBM defines an AI assistant as an application that understands natural language commands and performs tasks for the user via a conversational interface. It also draws a line there with an agent: an assistant is reactive and acts upon your request, whereas an agent is predominantly proactive, planning steps towards a goal on its own and using the tools available to it (IBM Think).
In practice, the boundary is fluid. The Ukrainian example that everyone saw is Diia.AI. On 1 September 2025, the Ministry of Digital Transformation announced that this national-scale AI assistant was advising and already providing ready-made services: even back then, income certificates could be obtained via chat, and the service was in open beta testing. In other words, an assistant that started out by giving answers was granted the right to perform an action.
The same thing happens with the assistant for your shop. As long as it only responds using the catalogue, it is an assistant in the narrow sense. When it creates an order in the CRM and calls a manager, it is already taking on part of an agent's work. If you want to understand this boundary more deeply, read what are AI agents and where is the boundary between them and assistants. When choosing a service, the name carries little weight. Look at what the assistant has access to and what it is permitted to do.
An AI assistant for yourself and your clients
By “AI assistant”, small business owners often mean different things. Some talk about a chat where they write posts and replies to suppliers themselves. Others mean a virtual assistant who communicates with customers instead of a manager. You likely already have the first option. According to Microsoft and LinkedIn (Work Trend Index, May 2024, survey of 31,000 people across 31 countries), 75 % of knowledge workers use AI at work, and 78 % of them bring their own tools without waiting for the company to organise it.
Customer support works by different rules, because their mistake is seen first by the buyer.
| Personal assistant | Customer assistant | |
|---|---|---|
| Who is writing to him | you or your team | buyers on Instagram, Telegram, on the website |
| Where did you get that knowledge from | the general model and what you pasted into the chat | catalogue, prices, stock levels, delivery terms |
| Who sees the mistake | you, before using the text | customer, immediately |
| When does it work | When did you open the chat | round-the-clock, without your involvement |
| What do I need to prepare | nothing but a good request | knowledge base and human handover rules |
The table shows why a personal chat cannot simply be “put on direct”. It is not connected to the channel, does not know your stock levels and has no rule for when to go quiet and call a human.
For a pet food shop, the difference is noticeable right on the first evening. The internal assistant helps to quickly write a post about a new batch of chicken-free food. The customer assistant answers those who read this post at 23:00 and want to place an order.
How an AI assistant differs from a chatbot when a customer writes outside of the buttons
A scripted chatbot and an AI assistant can sit in the same channel and have the same avatar. Where they differ is in how they process messages. The bot looks for a match with a pre-set option: a button, a specific word, a command. The assistant parses the meaning of the phrase and constructs a response from what it finds in your materials.
Where a button-based bot is still appropriate
A scripted bot handles narrow and predictable tasks well. For example, providing a tracking number by order number, booking a grooming session in a free slot, or showing the warehouse's working hours. Here, the customer is already thinking in terms of buttons, and the bot costs less and makes no mistakes in text because it doesn't write any.
Where does it break
Problems begin when a question consists of several conditions. “Dachshund harness, chest girth 48, so it doesn't press under the legs” doesn't fit into any branch. The same goes for surzhyk, mistakes, voice messages converted to text, and two questions in one message. The bot responds to such a request with a menu, and the person who wanted to buy has to explain everything a second time. The assistant handles such a message like any other.
If most of your direct messages look like questions about Maltipoos, you need an assistant. If they are dominated by “where is my parcel”, a button-based bot might just be enough.

How does the assistant know your product range and why does it sometimes make things up
A language model on its own knows nothing about your warehouse. It can talk convincingly about puppy food in general, but doesn't know that the 1.5 kg packs of salmon ran out on Saturday. That is why a customer assistant works with a knowledge base: before each response, it searches for the required fragment within it and builds its answer upon it.
At Marichka AI, the knowledge base is built without a programmer. The service itself reads website pages, accepts a catalogue in PDF or Excel, and automatically updates prices and stock levels. The assistant answers exclusively from this base. If the answer is not there, it hands the chat over to an operator along with the context, so the manager doesn't have to ask the client about the breed and age again in the morning.
Without such a restriction, the model sometimes fills the gap with plausible fabrication. In our example, it might “confirm” the hypoallergenic nature of the food, which the manufacturer did not claim. The mechanics are discussed in more detail in the article about how, Why is a language model confidently wrong?. This leads to one practical question for choosing a service: how does the assistant behave when it does not know the answer.
Five things to gather before connecting
The assistant responds as well as your materials are complete. Before the test, gather:
- Up-to-date catalogue with prices, stock levels, pack sizes and remaining inventory: file or link to the website.
- Delivery and payment terms as explained to customers, including dispatch days and exceptions.
- Twenty real direct message questions from the last month, along with the awkward ones.
- List of topics that the assistant does not handle independently: returns, complaints, discounts, wholesale orders.
- The name of the person to whom he is handing over the dialogue, and the hours when they are available.
The third point takes the most time, and it is also the most useful. When the pet shop owner sits down to write out questions, she usually finds several topics to which she does not have a written answer herself, for example, about switching food within the same brand. Such an answer will have to be written down anyway, with or without an assistant.
How much does an AI assistant cost for business and how to test it in a week
A ready-made AI business assistant is usually sold on a subscription basis, with the price depending on the number of messages. At Marichka AI, the LITE plan costs €49 per month for 1,000 messages and two manager seats, STANDARD is €99 for 2,000 messages, and an extra thousand costs €25. Custom development of an assistant is calculated differently: a separate budget for creation and monthly support after launch.
You can test it free for the first 7 days, which is enough for a fair trial. Download the catalogue, send the assistant your twenty questions from Direct and see where it answered accurately and where it handed over to a human. Then connect one channel, most often Instagram, and leave the assistant on for the evenings. At the end of the week, count how many dialogues it drove through to an order in the CRM and how many it handed over to the manager.
An AI assistant wins over a chatbot where your clients write in their own words, and offers no advantage where the “Delivery” button is enough for them. If you recognised your direct messages in the message about the maltipoo with an allergy, start with a week-long test using your own questions. After that, you'll know in numbers exactly how many evening conversations the assistant takes off your hands.
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
Is a virtual assistant a person or a programme?
Both meanings are live. In the labour market, a virtual assistant refers to a remote helper who manages email and calendars. In conversations about AI, it refers to a program based on a language model that responds to messages. Before comparing prices, clarify which of them is meant.
Can you set up ChatGPT to reply to customers in direct messages?
The chat itself for personal use cannot be connected to Instagram Direct and knows nothing about your stock levels. For customers, you need a service that connects to the channel via the official API, takes facts from your catalogue and hands complex dialogues over to a human.
How much does an AI assistant cost for a small business?
Ready-made services operate on a subscription basis. With Marichka AI, the LITE plan costs €49 per month for 1,000 messages, STANDARD costs €99 for 2,000, and the first 7 days are free to test. Custom development is priced differently: a separate budget for creation and monthly support.
Will an AI assistant replace a sales manager?
No, it takes away some of the work. The assistant handles repetitive questions and evening chats, while complaints, returns, discounts and non-standard orders are left for a human. A properly configured assistant passes such a chat to a manager along with the message history.
Can one assistant work on Instagram, Telegram and a website simultaneously?
Yes, if the service supports these channels. Marichka AI connects to Instagram, Facebook, Telegram, Viber, WhatsApp, and a website widget, and responds across all channels from a single knowledge base.





