How much does it cost to implement an AI agent: three engagement models

Three engagement models

The pricing page fixes three formats. The choice depends on the task, the data and the infrastructure requirements.

* The stated cost is the cost of the work. The necessary infrastructure and subscriptions are paid by the client independently. We help with the choice and the analysis.

A subagent performs one function inside an existing process — for example, sorts incoming emails or compiles the minutes of a meeting. The launch includes installation on your computer, basic configuration for the task, the passport and the work journal. An agent leads an independent scenario from the task to a verifiable artifact and includes integration with sources, team training and 30 days of support. Custom development is a project from scratch: assessment, architecture, development, integration, documentation.

What is included in the launch price

Regardless of the format, every launch includes three base components:

• The agent passport. Eight fields are documented: task, input, process, output, limits, permissions, perimeter and verification. It is the agreement about what the agent does and what it does not do.
• The work journal. Every stage is recorded: inputs, decisions, sources. The result is traceable back to the source data.
• Human control. No stage passes silently: disputed decisions stop and wait for confirmation.

Local deployment is a separate option: the environment and data flows are agreed for the task, including any required external connections.

That is why the "implementation price" is not just software. It is configuration for your data, control rules and documentation by which the team can work without the developer.

How to estimate payback honestly

Before launch, fix the result criterion — the site follows the same principle: without the task, the input data, the expected artifact, the acceptance criteria, the limits and the decision-maker the price is not discussed.

A working payback estimate is arithmetic on your process, not a promise:

1. How many hours a week the process currently takes from the employees.
2. What part of the work is typical and can be handed to the agent.
3. How much one hour of this work costs the company.
4. After what time the savings cover the launch price and the rework.

It is important to separate the calculation from the forecast: the calculation is your arithmetic, the forecast is a hypothesis that must be confirmed after the first launch. The project principle directly excludes unconfirmed promises of payback.

When the implementation will not pay off

Honest boundaries matter more than beautiful numbers. An agent implementation is risky if:

• The process is not yet stabilised: the rules and exceptions change every week.
• There is no input data: the agent works on materials, not on guesses.
• Every decision requires expertise, and a human will have to redo the agent's work anyway.
• The task is one-off — for it a chat or manual work is cheaper.

In these cases it is right to start with an assessment or a demo launch, not with a purchase.

Where to start

The demo launch is free: first watch the agent's result on the demonstration scenario, and make the launch decision after watching. Then the route is simple: choose an agent or a subagent in the catalog, fix the result criteria following the method, discuss the perimeter and access following the security principles — and only after that fix the price.

Read also: Cloud, team or local perimeter: how to choose the execution environment · Who needs an AI agent: four roles and their result