AI agent for 1C: MCP and plain-language requests

A manager writes: "Find stock for the item across three warehouses and prepare a reservation for the customer." The agent first clarifies missing parameters, then reads the permitted data and shows the result with links to 1C objects. For writing, it creates a draft or requests approval from the responsible employee.

Critical operations can be limited by amount, document type, department, legal entity, and execution time. The interface can be Telegram, a corporate messenger, a web chat, a voice assistant, or a panel inside the working app. The same controlled toolset can be connected to multiple interfaces and specialized agents.

MCP defines a consistent way to describe the tools an AI client can call. For 1C, we design an MCP server or an adapter over the system's available interfaces: HTTP and web services, OData, extensions, external processing, a bus, or managed file exchange. The 1C:Enterprise platform officially supports HTTP services, web services, and OData access; we choose the approach after checking the configuration and write requirements.

The security gate is a separate component of KT.Team architecture. The language model does not connect directly to the 1C database and does not receive tokens or credentials: it only selects the tool and forms the parameters. Before execution, the gate checks the identity of the user or calling agent, the available organization and department, the operation type, record scope, limit, repeat call, and whether approval is required. The agent acts within the rights of a specific 1C user, and every call and result are saved in the log.

MCP makes tools easier to reuse, but it does not replace business permissions or operational control on its own. The protocol specification requires validating input parameters, enforcing access control, limiting call frequency, confirming sensitive actions, and keeping logs; we record these requirements in the architecture and acceptance criteria. The same principle of controlled access is implemented across the entire company LLM & Security Gateway a unified gateway through which employees and agents work with language models.

An AI assistant usually means a question-and-answer mode: find data, explain a metric, suggest where to look. An agent is different because it performs allowed actions, preparing and posting documents through tools with permission checks and confirmation. The layer described here covers both modes: you can start with answers to questions and later enable operations.

AI accountant knows source accounting procedures, reconciliations, taxes, and exceptions in the accounting process. AI estimator Works with estimates, materials, substitutes, orders, shipments, and actual consumption. Both can use the same 1C MCP and shared security gate, but receive different tools, instructions, roles, and approval routes. KT.Team's AI solutions catalog is on the hub AI for business.

In the OSNO-VA financial agent case we moved calculation logic from spreadsheets into SQL, MCP, and executable skills. The agent calls verifiable tools and gets a reproducible result. We apply the same engineering principle to access to 1C.

The agent relies on the structure of the accounting system itself: 1C implementation, customization, and integration we treat as a separate direction. If most of the routine comes from exchanging documents with counterparties, start with EDI - 1C integration with Diadoc or 1C with SBIS, and the agent then works on top of the established document workflow.