AI agent development for one company process

An agent differs from a chatbot not by its model, but by its authority to act. A chatbot responds with text, and that is where it ends. An agent receives a task, determines what data it lacks, retrieves it from your systems, performs permitted operations, and stops where human approval is required.

This leads to a practical requirement that usually determines project success. An agent needs access, not prompts: to 1C, CRM, email, documents, and regulations—with permissions, logging, and a control point. That is why we design agents around data exchanges and permissions, not dialogue flows: behind our integration layer are 77 published enterprise cases from 13 years of work with PIM, ESB, WMS, and 1C.

What this looks like in practice is shown in review of open agent platforms: the technical layer is now available to everyone and costs almost nothing; the difference between a working agent and a demo lies in integration, permissions, and process ownership.

An AI agent pilot is more often held up by approvals than by technology. Security teams and technical directors ask the same questions in every project, and they need answers before launch, not at acceptance.

Where the data goes. Requests to models pass through a logged gateway, so the question “what exactly left the perimeter” is answered by the log, not by guesswork. If data must stay within the perimeter, models are deployed on your infrastructure; if you have no suitable hardware, services hosted in CIS data centers are used.

What the agent can do independently. Access rights are granted on a least-privilege basis, separately for reading and writing. Material actions—posting entries, sending to a counterparty, or changing data in an accounting system—require confirmation from a responsible person. The agent prepares the decision but does not sign it.

How to investigate an incident. Every agent step—what data it requested, what it decided, and why—remains in the trace. Investigating an incident becomes a matter of reading the log, not reconstructing events from indirect clues.

Who will maintain it a year from now. The solution is built on open components with permissive licenses and runs in your infrastructure. Your IT department can maintain it, not just us: we build this requirement into the project rather than merely promising it. The engines used and the reasons behind them are explained in review of AI agent platforms.

The $3,600 estimate covers a typical process: analysis, deployment in your environment, data connections, security gates, testing with real data, and team training. The exact figure depends on three things: how many systems must be connected, how clean the input data is, and how many rule exceptions there are.

The remaining cost is ongoing operation, which must be calculated separately: models, the access gateway, and support. We calculate it per user-day, not per abstract million tokens, so it can be compared with the salary of the person currently doing the work manually. The full calculation with a calculator is in the analysis how much an AI agent costs and what makes up its TCO.

Each subsequent iteration costs less than the first: the environment, access, and team skills are already in place. That is why we do not charge a subscription fee—the next process starts when it is genuinely needed.

A note on operations: in September 2026, costs fell significantly. Anthropic cut prompt-cache read prices fourfold, which is the main cost for long-running agents that repeatedly reread regulations, data schemas, and work history. As a result, calculations based on first-half prices now overstate operating costs. The latest figures and daily operating-cost formula are explained in how much an AI agent costs and in 2026 model comparison.