Articles · Choosing

Bot, assistant, agent: three rungs, and a rung zero nobody mentions

12 October 2026

The owner of an electronics repair centre was certain: staff were slow to reply to customers, so the firm needed a bot. The offer was already on the desk — 150,000 KZT plus 25,000 KZT a month. Then someone looked at three months of records. During opening hours, half of all enquiries got a reply within 9 minutes. The losses were somewhere else: all 494 website requests had arrived with no one assigned to them, and 231 of them never got a reply. That does not take a bot. It takes a CRM rule, for a one-off 20,000 KZT.

This firm is invented: it comes from our sample reports, which we use to test our own audit agent. The mistake, though, is a familiar one. Bot, assistant and agent are three different rungs, and what separates them is not the name but what the software is allowed to do. Take the lowest rung that does the job. And the lowest rung of all is not AI.

Three different things are sold under one word

In October 2026 we went through the Bitrix24 partner directory for Kazakhstan: of 225 partners, 39 already sell "their own AI agent" or a bot. The cheapest bots start at 5,000 KZT a month. What each partner means by the word, the directory does not say. It could be a script with five buttons, or a program that writes to your customers on its own.

So the name in the proposal can be skipped. Look at one thing: what the software does by itself. If it says what was written for it in advance, it is a bot. If it answers in its own words from your records, it is an assistant. If it changes something in your systems or sends something in your name, it is an agent. The higher the rung, the more the preparation costs, and the more a mistake costs.

Rung zero: a rule in software you already have

A bot vendor has little reason to bring it up: there is almost nothing to sell on this rung. At the service centre, the CRM assigned WhatsApp and Instagram requests to the receptionists automatically. Website requests were created with no one assigned and could only be seen through a special filter; 215 of them were still sitting unprocessed. A rule — every request has an owner, and the owner gets an alert if there is no reply within 15 minutes — closes that gap with no model involved.

Who it is enough for. Firms where work gets lost at the joins: a request with no owner, a task with no deadline, a call nobody is obliged to make. The sign is simple: your records contain enquiries that nobody answered.

What a mistake costs. Almost nothing: a rule can be switched off in a minute. But it has a price worth knowing in advance. It adds work rather than removing it: the website requests will now have to be answered — up to 21.7 hours a month at this firm.

Rung one, the bot: it says only what was written into it

A bot follows a script: buttons, stock phrases, a message at a set time. It does not understand a question; it recognises one. The same service centre was offered an "is my order ready?" bot for 40,000 KZT plus 15,000 KZT a month. For rung one that is honest work: one question, and the answer is already in the system.

Who it is enough for. Firms where one question or one message comes round dozens of times a week: order status, a reminder the day before an appointment, the address and opening hours. The condition is to count first. The report says as much: no order-status bot until the firm has counted those questions for two weeks.

What a mistake costs. A customer asks something the script does not cover and goes round the buttons in circles. It is irritating, but a bot makes no promises. It needs one way out: a "get me a person" button, and a person who actually turns up when it is pressed.

Rung two, the assistant: it answers by itself, but only from your records

An assistant reads a question however it is phrased and answers in its own words. It takes the answers from what the firm has written down: the price list, the terms, the frequent questions. The assistant on our website works this way. It has 40 knowledge cards. What is not in the cards it does not make up; it passes the question to a person. It does not quote prices. After every change to the cards it sits an exam of questions recorded in advance.

Who it is enough for. Firms that get messages when there is nobody to answer them. At the service centre, 41.6% of enquiries come in during the evening, at night or on a Sunday, and the reply goes out in the morning, typically 837 minutes later — about 14 hours. Of the enquiries made during opening hours, 42.5% end in an order; of the rest, 20.5%. That is the place for an assistant: the questions vary, the answers are in the price list, and the staff are asleep.

Here too, though, the cheap check comes first. Evening customers may simply be a different kind of customer, and a link is not yet a cause. The report recommends a month of evening and Sunday cover, done from a phone, for 60,000 KZT. If the share of enquiries that end in an order goes up, the assistant has work to do.

What a mistake costs. An assistant can answer confidently and wrongly: quote a price that does not exist, promise a date. The customer will arrive at the front desk holding that promise. So an assistant needs written-down knowledge, an "if I don't know, I hand over to a person" rule, and a check after every change. It also needs a list of what is off limits: at the service centre that is diagnosis and any price not on the price list, complaints, warranty claims and refunds.

Rung three, the agent: it does not just answer, it acts

An agent is given access to the firm's systems. It reads requests, sets tasks, drafts an email or an invoice, writes to a customer who has gone quiet. Our sales agent works like this: it finds companies, drafts an email and puts it in front of a person for approval. Without a person's yes, the email does not go.

Who needs it. Firms where the work runs to several steps across several systems, and what comes in is free text, different every time. At the service centre nobody follows up with customers who go quiet: the log does not contain a single second message from a member of staff. Keeping an eye on every request and following up in approved wording is work for an agent. But only once there is a rule — "we message the following day" — and people have tried it themselves. Put an agent on top of a mess and it makes the mess faster.

What a mistake costs. An action. An email has gone, a record has changed, a customer has been quoted a figure. That is harder to undo than a clumsy sentence. So before launch an agent gets a list: what it does itself, what it does only with a person's approval, what it never does. How to draw one up is in our article on who answers for an agent's mistake.

The question is not "bot or agent" but "what will this software be allowed to do, and who will notice when it gets it wrong".

Climb when the rung below stops coping

The report gave the service centre an order of steps, not a purchase. First, someone responsible for the website, and the CRM rule. Then a month of evening cover. As for the bot: come back to the decision in a month, with numbers in hand.

The report also says why the 150,000 KZT bot should wait. The vendor has not said what share of conversations the bot resolves on its own, and it is not clear whether the firm could cancel. The bot does not answer the phone or book devices in. None of that means the bot is bad. It means it was offered before anyone knew which rung was needed.

Three numbers and one evening will show you your rung

1. Open the system where your requests are kept and pick out those with no one assigned or with no reply at all. If there are any, you are on rung zero, and that is where to start.
2. Work out what share of last month's enquiries came in outside opening hours, and how long they waited for a reply. At the service centre in the example it is 41.6% and about 14 hours. If yours looks similar, try a month of evening cover, then come back to the assistant.
3. Take the last fifty customer messages and sort them into three piles: the same question with the same answer; varied questions whose answers are in your price list and rules; things a person has to decide. The first pile is work for a bot, the second for an assistant, and the third is not for software.

If the biggest pile is the third, you do not need software yet. You need a person who has time for it.

All articles