What is the cost of an AI agent?
The price of an AI agent rarely looks like a single digit in a commercial offer. We disassemble the bill in parts: what is paid once, what is paid every month, what determines the amount for the model and which three items of expenditure are most often forgotten to budget.
Briefly
- The AI agent bill consists of five parts: development, model work, hosting and channels, support, and the work of your employees. The latter is almost never considered, and it is comparable to the rest.
- The one—time part is scenarios, integrations and a knowledge base. It is several times larger than the first month of operation, so the payback is considered a year, not a month.
- The monthly bill for the model depends not on the number of employees, but on the volume of the text: how many requests, how many documents the agent reads before responding, and how long he remembers the conversation.
- Savings on the model are almost always returned by mistakes that people correct. First, achieve quality, then reduce the cost.
- If there are fewer than a few hundred requests per month, the agent will not pay off: the constant part of the expenses does not depend on the load.
Five items of expenditure
When the contractor calls the price one amount, it is impossible to argue with it: it is not clear what is inside. Decompose it into five parts — and you can immediately see what is included in the project, and what you will have to pay extra later.
| Article | When is paid | What does it depend on |
|---|---|---|
| Development: scenarios, integrations, knowledge base | Once, 2-8 weeks of work | The number of scenarios and systems that the agent contacts: accounting system, cashier, messenger, website |
| The work of the model | Every month, based on the fact of requests | Text volume: number of dialogues, length of answers, how many documents are read before each answer |
| Hosting and channels | Every month, a constant amount | Server or cloud, database, proxy for messenger, monitoring |
| Support and improvements | Every month or by the hour | How often the price list, schedule and regulations change; how many new scenarios are added |
| The work of your employees | The first 4-8 weeks are tight, then point-by-point | Who collects the knowledge base, reads the dialogues and tells where the agent answered incorrectly |
Which of these is one-time, and which is forever
Only the first line is unique, and this is the main trap when comparing sentences. The contractor who gave the price lower could simply not provide support: a month after the launch, the agent's price list becomes outdated, the schedule changes, a new service appears — and each such change becomes a separate account. Ask not "how much does the bot cost", but "how much will the twelfth month of its operation cost".
Example of calculation
The numbers below are a training example to show the order of magnitude and logic, not the price list. Substitute your values: you know the number of calls from correspondence and call logs, and the time for one call is from monitoring the administrator during the day.
| Parameter | The value in the example |
|---|---|
| Requests per month | 1500 |
| Average time of an employee's request | 4 minutes |
| Percentage of requests that the agent closes itself | 40% |
| Time saved | 1500 × 40% × 4 min = 40 hours per month |
Then it is considered simple: 40 hours are multiplied by the cost of an employee's hour along with taxes, and the monthly bill for the agent — model, hosting, and support is deducted from the result. The remainder divides the one-time part of the project and gives a payback period in months.
Two places where the calculation usually lies
- The share of closed requests is taken from the contractor's presentation. It needs to be measured on your data in the pilot: in one scenario, in one unit, in two weeks.
- Saved hours count as money. They become money only if another job is put on the free time or the next employee is not hired. If the administrator just got less tired, this is also a result, but it will not be in the profit report.
Ready-made AI solutions from the catalog
These solutions have already been assembled and are being implemented faster than development from scratch: they understand both the scope of work and the monthly part of the bill.
AI seller in Telegram
The AI bot guides the buyer through the stages of the sale: finds out the task, budget and necessary parameters, selects products from your catalog and gives them to the client's manager, who already understands what he is buying. The stages and rules of communication are configured without a programmer.
Request a KPAI-Knowledge base and consultant
The AI itself collects the "question — answer" database from your site and from the history of work chats, updates it when pages change, and responds to customers using this database. If necessary, the answer first goes to the employee for approval.
Request a KPAI control of work chats
The bot reads working Telegram groups, decrypts voice messages and sends a report to the manager every day: what was discussed, what tasks are not closed, how the team works. If the client waits longer than normal for the manager's response, the bot raises the alarm.
Request a KPAI-the manager's secretary
A personal AI assistant in your Telegram: parses mail and weeds out phishing, finds tasks in work chats, searches for contracts and invoices by meaning, prepares letters, tables and documents and reminds you about business. Does not send anything without your "yes".
Request a KPWhen an AI Agent doesn't pay off
There are tasks where it is pointless to count the payback — the agent will not return the investment at any price of the model. Check your task for four signs.
- Few calls. The constant part of the costs — hosting, channels, support — does not depend on the load. On twenty requests per month, the cost of one response is significantly higher than the cost of a person's minute.
- Each appeal is unique. If there are no recurring questions, the agent has nothing to pick up: he will become an extra link in front of the same employee.
- The answer costs more than the error. Medical contraindications, refunds, contract disputes — here we need human responsibility, not speed.
- There is no data that the agent could access. If the schedule lives in the administrator's head, and the price list is in three different files, first the data is put in order, and only then the agent is connected. Otherwise, you will have to pay twice.
If the task has been verified, the next step is not development, but a pilot on one scenario: it gives a real share of closed requests and a real bill for the model.
How do we consider the project budget
We look at requests and data
We analyze correspondence and calls over the past months: what questions are repeated, how many of them, how long the answer takes. We look at where the agent will get the facts from: accounting system, price list, schedule, regulations.
The number of requests, the proportion of recurring requests, and the list of systems to integrate
We count two sums
A one—time part — scripts, integrations, knowledge base preparation - and a monthly part: model, hosting, support. We calculate the monthly amount based on your volume of requests, and not on the market average.
Project budget and bill forecast for the twelfth month
We check on the pilot
We run one scenario on real requests and measure the share that the agent closes without a person, and the actual cost of the model. If the figures do not agree with the calculation, we adjust the amount of work before the entire budget is spent.
Confirmed payback or reasoned rejection of the project
Let's calculate your budget for an AI agent
Write down how many requests you receive per month and in which system the data lives. We will return with two amounts — one—time and monthly - and an honest answer as to whether the agent will pay off on your volume.
- We will respond within a working day
- Let's clarify the volume of requests and the system
- We will send you a payment with a one-time and monthly part
Is it more convenient in the messenger?
+79262103289Employees respond from 9:00 to 23:00 Moscow time, every day
How can we help
AI agents
Not a demo chatbot, but an assistant on your data: from 1C, Access Control systems, CRM and regulations. Responds to guests, prompts employees and undertakes typical operations.
YandexGPTGigaChat1ССКУДChatbots
Chatbots for the website and messengers that respond to guests based on your data and transmit the conversation to an employee when a person is needed.
YandexGPTGigaChatTelegram1С
System integration
Turnstiles, ticket offices, 1C, CRM and the site begin to exchange data themselves. Employees stop transferring numbers with their hands.
1ССКУДREST APIЭквайрингFrequent questions
How much does an AI agent cost for business?
The account consists of five parts: the development of scenarios and integrations (one—time), the work of the model (monthly, depending on the number of requests and the length of responses), hosting and communication channels, support and improvements, as well as the hidden part - the time of your employees to check the answers in the first weeks. The one-time part is usually several times larger than the first month of operation, so it is pointless to count the payback for one month.
What does the monthly AI agent bill depend on the most?
Depends on the amount of text that passes through the model. It is influenced by three things: the number of requests, how many documents the agent reads before each response, and how long he remembers the conversation. An agent who uploads the entire forty—page schedule to each answer costs several times more than an agent who searches the database and takes two necessary paragraphs - with the same response quality.
Is it possible to make an AI agent cheaper if you take a simpler model?
Sometimes yes, but the savings are often imaginary. A simple model makes mistakes more often, and someone corrects every mistake with their hands: the administrator rewrites the answer, the client writes a second time, the manager analyzes the complaint. The correct procedure is as follows: first, achieve acceptable quality on a strong model, measure the proportion of calls that the agent closes himself, and only then try the model cheaper, comparing both for this proportion.
How to understand that an AI agent has paid off?
We need two figures before the launch: how many requests per month and how many minutes one request takes an employee. After the launch, the percentage of requests that the agent closed without a person is counted, and how long the people returned. If an agent closes 40% of requests for 4 minutes at 1500 requests per month, this is 40 hours of work per month. Then you compare it with the invoice amount — and you get a payback period in months.
Which is more expensive: to develop your own agent or to take a ready-made service?
A ready-made service is cheaper at the start and more expensive to operate if there are a lot of requests: you pay for each workplace or for a package of messages. Your agent is more expensive to develop, but his account grows more slowly and does not depend on someone else's tariff schedule. The border usually runs along integrations: if an agent needs to write to your accounting system, make requests and check payment, ready-made services run into restrictions, and development pays off faster.
Do I have to pay for an AI agent when no one is using it?
For the work of the model — no, it is charged on the fact of requests. But the permanent part remains: the server or cloud where the agent lives, communication channels, monitoring and support. Therefore, an agent that processes twenty requests per month has a very high cost of one response: before launching, it is worth checking whether there is enough volume to start automation at all.
What expenses for an AI agent are most often forgotten to lay?
Three: preparation of the knowledge base (someone has to collect and clean up the regulations, prices and answers), the work of a person to parse dialogues in the first weeks and improvements after the first month - there are always scenarios that were not in the task. According to our projects, these three articles together are comparable to the development of the agent itself, and when they are not laid, the project looks like a failure on budget, although the agent himself is working.
Read also
What to automate first: calculation instead of intuition
Many automation and AI projects do not have a measurable effect because they start with the wrong task. We analyze the research, prioritization techniques and give a formula according to which you can select the first task in two weeks.
How to test a bot idea before development
We are analyzing how to figure out in a few weeks whether customers will use the bot: an interview without prompts, a demand test, a sample calculation, a price using the Van Westendorp method and requirements for consent to personal data.
Digital Employee: what the bot will take from the administrator
The administrator of a fitness club or thermal spa complex answers the same questions all day long. We are sorting out what can be given to a digital employee for September 2026, what to leave to people, in which messenger to launch the bot and how to accept payment in it according to the law.