Sales AI agent: from first message to order

An AI sales agent is a language model application that takes over the conversation with a buyer from the initial “hello, is this in stock?” to a created order in the CRM. What distinguishes it from a chatbot is the result: a bot replies, an agent drives the sale through to completion or honestly hands over to a human. Below is a step-by-step guide: how the agent guides the client at each stage of the dialogue, what it needs from you, how to distinguish a working product from a renamed bot, how much it costs, and how to launch in a week-long trial. End-to-end example: an online household chemicals and cosmetics store, Instagram Direct, 23:10.

Six steps of the dialogue with an agent in direct

At 23:10, a mother of two writes in the direct messages: “I'm looking for a hypoallergenic, fragrance-free laundry detergent for children's clothes. What would you recommend?” The manager is asleep. The agent takes over.

Step one, the response. The agent replies within 10–30 seconds, as stated in the Marichka AI description. Speed clinches the sale here: the client had another two tabs open simultaneously, and the first sensible reply wins.

Second step, clarification. The agent asks one question instead of five: “Washing machine or hand wash, and which volume is more convenient: up to a litre or a month's supply?” This is qualification, just without a questionnaire. The answer narrows the choice down to two or three items.

Step three, selection. The agent looks at the catalogue, which is compiled from website pages or a PDF or Excel file, and sees current prices and stock levels because they are updated automatically. They suggest two products with an explanation of the difference between them, and do not invent a third one that is not in stock.

Fourth step, objection. “It's a bit pricey”. The agent has the right to explain the price per single wash cycle, offer a smaller volume or remind about free delivery from a certain amount, if such a rule is in the knowledge base. It does not have the right to give a discount, and that is correct.

Step five, ordering. The agent asks for the city and branch, confirms the amount and creates the order in the CRM. In the morning, the manager sees the order card with the contact and comment instead of an open question.

Step six, follow-up. If the client goes silent after the third step, the agent returns to them the next day with a short message stating that the smaller volume is still available. It is this exact step that human managers most frequently skip in their daily workflow, whereas the AI agent never misses it. A public example of this scheme in action within a different niche: the food delivery service Box Catering generated over 500 leads with a processing time of up to 10 seconds and UAH 500,000 in revenue via an AI manager (the case study is published on the Marichka AI website).

What the agent needs from you: a knowledge base and boundaries

An agent sells precisely as well as their knowledge base is assembled. It consists of four parts.

Catalogue. The simplest way is to provide a link to the website: the agent will gather the names, descriptions, prices and availability itself. If the catalogue lives in a spreadsheet, a PDF or Excel file will do. It is important that price and stock updates happen automatically, otherwise in a month's time the agent will start selling things that are no longer in stock.

Shop rules. Delivery and payment terms, timeframes, returns, minimum order amount for free delivery, and which cities are not serviced. Everything that the female manager keeps in her head needs to be written down as text.

Tone and taboos. How to address the customer, whether jokes are allowed, and which words not to use. And a separate list of things the agent does not handle on their own: discounts above the limit, returns, complaints, wholesale enquiries.

Handover limit. When the agent calls a human and exactly how. In Marichka AI, a complex question goes to the operator along with the dialog context, while the agent itself replies solely from the knowledge base and does not make things up. The setup is described in an eight-step video guide and requires no code, but you will have to assemble the four parts of the knowledge base yourself: no one else knows your rules.

Text or voice AI sales agent

There are two classes of products on the market. The text agent lives in messengers and on the website. The voice AI sales agent makes and receives calls: confirms orders, calls through the database, and works as the first line of a call centre.

For a shop with sales via Instagram and website, the choice is obvious, and the audience itself suggests it: the buyer writes in Direct or in Telegram, and you need to reply where they write. A voice agent makes sense when a significant portion of orders come in via calls or when delivery needs to be confirmed by voice. Marichka AI works with text in Telegram, Instagram, Facebook, Viber, WhatsApp, as well as via a widget on the website; a separate tool is needed for calls.

Will an agent replace a manager

Short answer: in a small shop, it will replace the night shift and repetitive conversations, but not the person who resolves disputed situations. A longer answer is available from Gartner. According to a forecast dated 18 November 2025, by 2028 there will be ten times as many AI agents in sales as live salespeople, and at the same time fewer than 40 % of salespeople will say that agents increased their productivity. Analyst Melissa Gilbert explains this by the “value ceiling”: past a certain point, more AI no longer yields more productivity. Piling tools onto complex processes only exhausts people (Gartner, 18 November 2025).

The same release advises starting with data quality and a pilot with a clear objective rather than scaling up the number of bots. From this, a simple allocation rule follows for a shop. Agent: first response, catalogue selection, checkout, reminders. Human: returns, complaints, wholesale enquiries, any dialogue where the customer is annoyed.

How to choose the best AI sales agent

The best sales AI agent for your shop is the one that passes the test on your own dialogues, and the length of the feature list has nothing to do with it. Under the name “agent”, the market often sells ordinary bots; we have broken down how Gartner describes this phenomenon and where the boundary of autonomy lies in the article What are AI agents. Therefore, check instead of reading.

1. Provide a link to your website and see whether the agent compiled the catalogue itself or requires a spreadsheet manually.
2. Do you have an unscented product under one litre that can arrive in Lviv by Saturday?.
3. Please ask them to place an order and check whether it has appeared in your CRM.
4. Ask for a discount greater than permitted. The agent must refuse and transfer to a human.
5. Ask about an item you do not have. Correct answer: “we don't have that, there is a similar one”, without a fictional description.
6. Be quiet for a day and see if the agent comes back on their own.
7. Check which channels he works with using a single knowledge base, and whether there is role-based access for managers.

How much does it cost and how to calculate the payback period

Ready-made agents are sold by subscription with a message limit. At Marichka AI, the LITE plan costs €49 per month for 1,000 messages and two managers, STANDARD is €99 for 2,000, BUSINESS is €299 for 7,000 messages with external CRM and a personal manager, an additional thousand messages costs €25, and there is a 7-day trial.

Payback is calculated without complex models. Take three of your own figures: how many chats arrive per month outside of working hours, what proportion of them currently end without a reply until morning, and what the average order value is. Multiply them, and you will get the money that is sitting in Direct at night. Compare that with the cost of the subscription plan. Average figures for the market are intentionally omitted here: in cosmetics and in food delivery these numbers differ multiple times over, and someone else's figure will only throw you off.

Launch in 7 days of trial

A week-long trial is enough to get a “yes” or “no” answer on your own data, provided you don't spend it hesitating.

1. Day one: provide a link to the website or catalogue file, and check that prices and stock levels have pulled through.
2. Day two: write down delivery and payment rules, return conditions, list of prohibitions.
3. Day three: run through twenty real questions from Direct over the past month and fix the responses that didn't appeal.
4. Day four: connect one channel, preferably the one with the most night-time messages.
5. Day five: connect the CRM and check that the test order was created.
6. Day six: turn on the agent for the evening and night time, leave the manager during the day.
7. Day seven: read all the dialogues through the night and count how many of them led to an order.

A detailed breakdown of each day with configuration examples is in a separate manual how to test an AI manager in 7 days.

Before the trial, do one thing: export the chats from Direct for the last two weeks and note the time of each first message. If a third or more fall outside the manager's shift, you have something to hand over to the agent right now. If less than a tenth, start with daytime response speed, and the agent can wait.