What is an AI agent.

What an AI agent is, in plain words

An AI agent is a program that is given a goal rather than a script. It reads what a person writes in free-form text, decides on its own which steps to take, works with the systems it needs and performs actions: creates a deal, sets a task, books a meeting.

The difference from a chatbot is fundamental. A chatbot follows a branch drawn in advance: step, button, next step. If a person writes something the author of the script didn't foresee, the bot gets lost. Instead of a branch, an agent is given a goal — for example, «find out what the client needs and book a site visit» — and a set of tools it may use. How exactly to reach the goal it decides itself, in every conversation.

How an AI agent differs from a chatbot

A chatbot executes a script, an AI agent pursues a result. Every other difference follows from that — in understanding speech, in memory, and in what happens with an unfamiliar question.

The practical consequence: a chatbot has a ceiling beyond which it starts to irritate, and that ceiling comes quickly. An agent's ceiling is different — it runs into the quality of the materials you gave it.

What kinds of AI agents there are

By task they fall into four kinds: selling, support, working and analytical. By design — into those that only answer and those that act: the first only talk, the second have access to the company's systems.

The selling agent

Takes the enquiry, works out what is needed, handles objections and carries it to the next step — a booking, a site visit, an invoice. This is the kind people usually mean when they say «sales robot».

The support agent

Answers questions about an order, delivery status and return terms. Works with a knowledge base and the client's history, escalates difficult cases to a person.

The working agent

Doesn't talk to clients at all, it does the internal work: fills in cards, prepares quotes, distributes tasks, collects documents. Nobody sees it, but it saves the most hours.

The analytical agent

Reads conversations and chats, scores them against criteria and turns them into figures. Speech analytics and quality control are built on this.

A separate division is by whether the agent can act. An agent that only answers is useful as an information desk. An agent with access to the CRM, the calendar and the stock replaces part of a person's work. The difference in implementation cost between them is noticeable, and this is exactly what to ask a contractor about.

What an AI agent is made of

Of four parts: a language model, tools, memory and rules. The model understands and formulates, the tools let it act, the memory holds the context, the rules set the limits.

The model

It is responsible for understanding what is written and for the wording. By itself the model knows nothing about your business and can do nothing beyond text.

Tools

Access to the CRM, the calendar, the price list, the stock. It is the tools that turn a conversationalist into a worker: without them the agent can only talk.

Memory

The context of the current conversation and the client's history. Thanks to it a person doesn't have to retell what they already said a month ago.

Rules and limits

What may be promised, which discounts are allowed, when to call a human, what not to discuss at all. Without this part the agent will sooner or later say too much.

Where an AI agent's knowledge of your business comes from

From nowhere, until you tell it. A language model knows the language and general facts about the world, but it doesn't know your prices, your terms or what you even sell. Knowledge of the business comes from the materials loaded into it.

It works like this: before answering, the agent searches your materials for the pieces relevant to the question and answers with them in front of it. So the quality of answers comes down not to how «clever» the model is but to whether the needed piece is in the base.

What is usually loaded into it:

• the price list, terms, lead times, guarantees;
• descriptions of services and products with specifications;
• frequent questions and good answers to them;
• real objections and what strong salespeople say to them;
• the rules: what may not be promised, when to call a human.

Hence an important consequence: if the agent often answers «let me check with a specialist», the problem is almost never the model. The problem is that the needed piece isn't in the base.

What an AI agent can do in sales

Take the first contact and the routine off people: reply instantly at any hour, work out what is needed, set the next step and keep the deal card without a salesperson.

• replies within seconds in messengers, on the website and on marketplaces;
• understands voice messages and photos, not only text;
• works out what the person needs and filters out irrelevant enquiries;
• books a meeting, a site visit or a call and sends reminders;
• creates the deal in the CRM and fills it in itself;
• comes back to those who went quiet — without a salesperson having to remember;
• hands a hot client over to a person together with the conversation history.

What a conversation with an AI agent looks like

From the client's side — like messaging an attentive salesperson who replies instantly. From the company's side — like an enquiry that created itself in the CRM with the details already worked out.

A worked example from the glazing trade:

• Client, 22:40: «Hello, how much would it cost to glaze a balcony?»
• Agent: «Good evening! Let's work it out. Is the balcony in a flat or at a country house, and roughly how long is it? If it's convenient, send a photo — that makes it more precise.»
• Client: sends a photo.
• Agent: «Thank you, I can see it — a standard three-metre balcony with straight glazing. Roughly 60 to 90 thousand depending on the profile. The exact figure comes from the surveyor, and the survey is free. When suits you — a weekday or the weekend?»
• Client: «Saturday morning, please.»
• Agent: books Saturday, creates the deal in the CRM with the photo and the parameters, and sets the salesperson a task to confirm the survey.

Note two things. The agent did not name an exact price, because it is not allowed to — that is a rule set during configuration. And it did not send the client away until morning: the conversation happened at 22:40, and in the morning the salesperson sees a booking rather than an unread message.

What an AI agent cannot do

It doesn't know what you haven't given it, it doesn't run complex negotiations and it doesn't fix a broken process. An agent speeds up what already works and exposes what doesn't.

• It doesn't replace negotiation. Leave hard bargaining, non-standard terms and work with a large client to a person.
• It doesn't invent facts when set up properly. But if there is little material, it will often answer «let me check with a specialist» — and that is a question of preparation, not of technology.
• It doesn't fix the process. If enquiries are lost because there is nobody to handle them, the agent will merely bring the client to the same wall faster.
• It doesn't work without data. Without a CRM the correspondence stays in the messenger, and a month later nobody remembers what was agreed.

AI agent, live salesperson or scripted bot

These are not competitors but different tools for different parts of the job. A scripted bot is cheap and fits simple repeating requests, an AI agent covers the first contact and the routine, a person is needed where negotiation and big money begin.

The working arrangement is usually mixed: the agent takes the first contact and everything that repeats, the person joins at the deal. An attempt to replace the whole sales team with an agent ends the same way every time — with the loss of large clients who need a human.

Where an AI agent pays off fastest

Where there are many enquiries, they are of the same kind, and the decision is quick: consumer services, appointment booking, quotes from parameters. Slowest where the deal is long and more than one person decides.

• Services with appointments — clinics, salons, car services. The client wants a specific time slot, and the agent closes that in two messages.
• Construction and refurbishment — windows, gates, stretch ceilings. Many evening enquiries and a repetitive first conversation: parameters, a price guide, booking a survey.
• Selling on marketplaces and in messengers — availability, lead times, delivery, specifications. The questions repeat by the hundred.
• Education and events — programme, schedule, price, registration. Spikes of enquiries on announcement days.

It is harder where a person's expertise matters from the first minute: complex B2B equipment, large projects, tenders. There an agent is useful not as a seller but as a worker — preparing documents and keeping the cards.

How long implementation takes

A simple scenario in one channel launches in a few days. A full agent with CRM integration and objection handling usually takes two to four weeks, and most of that time goes not on the technology but on collecting your materials.

Collecting the materials turns out to be the bottleneck almost every time: prices, terms, frequent questions and the answers that worked are scattered across people's heads and old chats. Until they are gathered, the agent has nothing to answer with.

How much an AI agent costs

The price is made of three parts: a one-off setup, monthly support and the charge for calls to the model. The spread is wide because «AI agent» is used to sell both an off-the-shelf chatbot and a system with integrations.

The right order is to count not from the price but from the losses. Take the number of enquiries a month, estimate what share goes unanswered for more than half an hour, multiply by your conversion and average deal value. That is the sum you are losing now. If it is several times the cost of implementation, the decision is obvious; if it is comparable, it is probably too early for you.

What to agree in advance: who owns the materials and the settings, what happens when you part ways with the contractor, and whether refining the answers after launch is included in the price. The last point matters most — without it the launch is not finished.

Five mistakes when rolling one out

Almost all failed rollouts look alike: the agent is launched without materials, without a CRM, without rules and without refinement after launch.

• Launch and forget. For the first weeks you have to read real conversations and correct the answers. Without that the agent stays for ever as it was assembled on day one.
• Give it no materials. Expecting sensible answers from an agent that wasn't given the price list and the terms is the same as expecting them from a newcomer on their first day.
• Pass it off as a human. It comes out quickly and spoils the impression more than the fact of a robot ever would.
• Let it handle every deal. Large and non-standard deals should be handed to a person straight away, by a rule rather than by hope.
• Pour traffic into a leaking funnel. If enquiries get lost after the agent, it will only speed up the losses.

What to ask a contractor

Five questions separate those who implement from those who sell a box.

• Does the agent only answer, or does it have access to the CRM and the calendar?
• What happens when it doesn't know the answer?
• Who owns the materials and the settings once the work is finished?
• Does the price include refining the answers in the first weeks?
• Where are the conversations stored and how is that handled legally?

An answer of «our AI will learn everything by itself» to the first question is a reason to be wary. It will learn nothing you haven't given it.

Personal data and security

Correspondence with a client is personal data, and handling it is regulated by law — the GDPR in Europe, 152-FZ in Russia. The operator is the company that collects the enquiries, not the contractor.

What this means in practice: you need a lawful basis or consent, a published policy and clarity about where the data is stored. A separate question is sending texts abroad: many models run on servers in other countries, and that is a cross-border transfer with its own requirements.

Ask the contractor where the conversations are physically stored and what data goes into the model. If there is no answer, that is a risk which will stay yours, not theirs.

How to tell whether you need an AI agent

It pays off where there are more enquiries than people can handle and where the first reply is slow. If there are few enquiries and answers are instant, there will be almost no gain.

A simple check — answer four questions:

• Do enquiries arrive in the evening, at night or at weekends?
• Does more than ten minutes pass between an enquiry and the first reply?
• Do salespeople spend time on the same questions over and over?
• Do clients who say «I'll think about it» disappear with no follow-up?

Two «yes» answers and the agent will pay for itself. None at all — deal with something else first: most likely you need order in the CRM or more enquiries, not an agent.

What we see in our own data

In one project the AI reviewed 211 conversations across 123 deals for four salespeople. The main observation is simple: conversations hold far more information than ever reaches the CRM.

The real reason for losing a deal almost never matches the one the salesperson put in the card — because in the card people pick whatever closes it fastest, while in the conversation what actually happened is spoken aloud. That is not a reproach to salespeople but a property of any manual reporting.

Hence the practical conclusion: the value of an agent is not only that it answers instead of a person. It also turns conversations into data that did not exist before it.