How much does it cost to develop an AI agent: what makes up the price and real examples

The short answer is:development from scratch with integrations

September 2026

The main idea is simple: an AI agent is paid not once, but three times — for implementation, for working models, and for support. Most unpleasant surprises occur when only the first item is taken into account in the budget.

Three items of expenses

integrationhow models workinfrastructure and support

Integration

How models work

Infrastructure and support

Why are ""turnkey agent"" and "agent on the platform" different?

On the market, you can find offers from 50 thousand rubles to tens of millions. To compare them honestly, you need to understand what exactly is being sold: setting up a ready-made system, developing a separate solution, or an enterprise architecture with multiple agents.

not as answers, but in the scope of development

transfer difficult cases to a person

There is also a downside. The platform sets the framework. If a company wants to control every internal component, change the architecture at the code level, and store everything only in its own loop, the price inevitably increases. In this case, not only the agent is bought, but also the engineering freedom.

How much does the work of models actually cost?

The token

Russian models make it possible to pay for consumption directly in rubles. According to the open tariffs for September 2026, the benchmarks look like this:

GigaChat source: tariffs for legal entities of the Savings Bank, effective from 09/01/2026 for existing corporate clients; for new clients, the catalog is used from this date cloud.ru , where conditions may vary. YandexGPT tariffs need to be checked on the day of publication, because the price pages of cloud providers sometimes change their structure and availability.

Nuances are important in the practical estimate. GigaChat has a minimum payment of 600 ₽ in months with usage: if the actual consumption of Lite was a conditional 57 ₽, 600 ₽ can still be payable. The gap between Lite and Pro from Russian suppliers is about 4-8 times, so sending the entire stream to the older model without routing means overpaying. Simple questions can be handled by the younger model, controversial and expensive ones by mistake can be handled by the older one.

For Russian models, the entrance and exit often cost the same. For Western models, the output is usually several times more expensive than the input. This changes the arithmetic: long answers on Russian models are relatively cheaper, and RAG scenarios with more context are relatively more expensive.

RAGcontext architecture

In the measurement of nine real-life support dialogues published by mxchat.ai , there were about 7,200 input and 716 output tokens per conversation. At the same time, 48% of the input tokens were documents extracted from the knowledge base, 40% were system promptings sent at each step, and only 2.5% were user words. The conclusion is unpleasant, but useful: nine tenths of the bill is determined by the setup, not the traffic.

Saving on the model yields less than saving on the context. Caching system software can reduce consumption by 40-70%, and routing simple requests to a lower—end model by 50-70%. But agent-based scenarios, where one user request turns into 4-6 model requests, quickly increase consumption. Multi-agent systems can spend 10-15 times more tokens than a single agent.

A separate topic is access to foreign models from Russia. Russia is not included in the lists of supported countries by OpenAI, Anthropic, and Google, and payment gateways and intermediaries add significant margins to official prices. According to open prices, it can range from tens to hundreds of percent, depending on the model and access scheme. In addition, a ruble payment through a Russian account does not by itself prove compliance with data localization requirements.

What do they pay for repeatedly, and what do they pay for once?

A good estimate separates one-time and regular expenses. This protects against a typical mistake: the company compares two proposals for the implementation price, and six months later finds out that one includes knowledge base updates and quality control, and the other contains only initial configuration.

The one-time part answers the question "how to launch". The second one is "how to make the agent continue working after three, six and twelve months." Businesses change promotions, prices, regulations, schedules, inventory balances, application forms, and the API of external services. If the agent is not updated, it gradually begins to respond with last year's company to current customers.

Hidden costs that are not included in the estimate

TCO

A useful rule of thumb from corporate assessments is that for every ruble in the agent's "brains" there is about one and a half rubles for integration and its support. Therefore, a cheap model does not save the project if the agent is poorly connected to the data, does not know how to transfer controversial cases and requires manual analysis of every second dialog.

30-50% of the development cost per year

There is another article that is rarely compared to the API, salaries. The average salary of an ML engineer in Russia in 2026 is approximately 185 000 rubles per hand, with a typical range of 140-230 thousand rubles, and higher in Moscow. Even a small internal team is noticeably more expensive than model calls. Therefore, the question "do it yourself or give it to the contractor" should be considered not by the price of tokens, but by the full cost of the team, deadlines and risks.

There are situations when it is better not to order an AI agent. If the task is already being solved by regular CRM or marketplace functions, if the process is not described, if the knowledge base is contradictory, if the company does not have an owner of the result, the agent will not become an accelerator, but a beautiful overlay on the chaos.

Market benchmarks: how much is being asked in Russia and around the world

Below are not our prices, but market benchmarks from open reviews. They are needed to understand the scale of forks and not compare the incomparable.

Other reviews give similar but offset forks: RAG agent 150-200 thousand rubles, functional agent 600 thousand - 2.5 million rubles, multi–agent system from 10 million rubles. The spread itself is indicative: the market is not counting the "agent at all" yet, but a set of works around him.

In the world, the order is different: a simple agent — $ 8-25 thousand, an agent with a knowledge base — $25-80 thousand, a multi—agent project — $80-250 thousand, a corporate platform - from $ 250 thousand. One corporate support agent can cost $150-300 thousand without taking into account the cost of the models.

An interesting trend is the transition from paying for jobs to paying for results. Intercom charges $0.99 for a resolved request, Salesforce uses payment models for an agent's action or for a conversation. The logic is clear: if a product does the work instead of a human, it becomes less natural to sell licenses to people.

The agent's price is almost never determined by the model's price

Our prices

Softrestchat is a platform of our own development for creating AI agents for communication: text messages and voice conversations. On this platform, we design agent behavior, set up scenarios, connect communication channels, prepare a knowledge base, check the quality of responses, and then implement the solution into the company's workflows.

30 000 ₽

This cost is obtained due to the finished base. Softrestchat already has dialog management, channel connectivity, knowledge base storage, control screens, event logs, and basic analytics. The client pays for the configuration for a specific task and the implementation of the agent in their process.

8,000 credits5-10 ₽ per minute of conversation30 kopecks for the agent's reply

build an agent yourself5,000 ₽ per month

95,000 ₽ for one company workflow

It is correct to compare these prices with the launch on a ready-made platform and with the development of an individual solution for the task. If you need non-standard connections, a closed loop, complex decision-making logic, or multiple interacting agents, the cost is calculated separately.

How to calculate the payback and where lies are most often used in this calculation

The payback of an AI agent cannot be calculated using the formula "how many operators will we replace". It is more correct to look at the flow of requests, the proportion of tasks that the agent actually closes, the cost of human processing, and the total cost of ownership.

Annual savings = volume × percentage of requests actually closed by the agent
                 × (the total cost of treatment by a person − the cost of treatment by an agent)
− the total cost of ownership

The first mistake is to calculate by the average cost of treatment. A simple consultation and a complex escalation may differ in price by 5-10 times. If you take the average, you can both underestimate the benefits on simple issues and overestimate them on complex ones.

20–35%

The third mistake is the lack of a baseline. If before the implementation they did not measure how many calls there were, how long the call lasted, how much it cost to process, what proportion went into repeated contacts and escalations, it is almost impossible to calculate the effect. After launching, there is always a temptation to choose beautiful numbers.

The fourth mistake is to confuse retention with decision. The agent can have a long dialogue, but not close the task. For a business, it is not the number of messages or the percentage of "automatically processed" requests that is more important, but the proportion of cases that are actually resolved without harm to the customer.

According to industry guidelines, payback in 3-6 months is possible for medium-sized businesses with a ready-made knowledge base and a clear flow of standard requests. 6-9 months is more realistic for corporate implementation with multiple systems. With an honest accounting of all expenses for the first year, a period of 8-12 months often looks healthier than the promise of "it will pay off in a month."

What to do next

If you want to estimate the cost of an AI agent for your process, start by describing the task. We need the volume of requests, calls, or documents per month, typical scenarios, rules for transferring to a person, a list of integration systems, and an understanding of who will own the result within the company.

According to this description, you can prepare a calculation in three configurations: a minimum launch on the platform, a working version with integrations, and a maximum scenario with advanced logic. For each option, it is useful to immediately specify the implementation price, monthly expenses, terms and conditions under which the estimate will change.

a platform for creating AI agentswhich questions are repeated most often?

Автор Антон Амосов

Поделиться Скопировано!