How much an AI agent costs: what determines price and TCO

The market answers the question "how much does an AI agent cost" with a range from tens of thousands to tens of millions of rubles. The range honestly reflects the spread of tasks, but you cannot plan a budget from it: it contains four different cost items, each with its own mechanics and payment schedule.

A model subscription or platform license covers only one of these items - inference. The effect appears when the agent is connected to operational systems - 1C, ERP, CRM - operates within the rights of a specific user, and leaves a trace in the operations log. This surrounding infrastructure is paid for separately from the license, and in terms of total cost of ownership it is what determines the final amount: TCO is made up of creation, operation, and risks minus economic benefit - we broke down the formula in the article on TCO for IT projects.

Below is what each cost item consists of, where to get verifiable figures, and what inflates the budget. The cost structure for your configuration is calculated by AI automation calculator: number of processes, automation share, token volume, and inference setup.

Tokens are the only cost item that can be calculated before launch with model-level precision. In LLM comparison 2026 we use the active workday formula: base = 1M·P_in + 0.5M·P_out, meaning one million input and half a million output tokens per active user per day. Claude Opus 4.8 costs $5 / $25 per 1M tokens (input / output), or about $17.50 (about RUB 1,400) per day. All figures are estimates as of 27.06.2026, with an exchange rate of 80 RUB/$.

Integration with a large system estate. Each system in the agent setup is a separate tool with its own permissions, tests, and error handling. The budget grows with the number of systems and fastest of all with write operations: a confirmed write requires an approval route, limits, and logging, while reads and reports are cheaper.

No API. If a system has no programming interface, access is built through exports, changes to the system itself, or interface automation. This is a separate project with its own budget, and its cost can exceed the cost of the agent itself.

Personal data. It cannot be sent to external models without anonymization - GDPR. This adds a mandatory element to the setup: LLM gateway, which anonymizes data before it reaches the model and restores real values after the response. This covers both development and operations.

Requirement for a local model. On-premise removes external data transfer, but adds capital expenses and operating costs for the setup. Economics here depend on GPU utilization: a heavily loaded setup makes each token cheaper, while idle hardware makes it more expensive than the cloud. We calculate the configuration and payback after the load profile; you can compare setups, including RU cloud under GDPR, in calculator.