How Business Agents Work in Enterprise Systems

Roman Korotaev Head of Enterprise Automation at Cinimex Let us define the term. A business agent is an automated software participant in a company's actual business process. It has a specific objective and access to the required source information through a limited and controlled set of tools. The business agent analyzes the data it receives and produces a verifiable result.

What Is a Business Agent?

Unlike a chatbot or an AI assistant, a business agent manages a task from its initiation to a defined point of completion. It applies domain-specific rules, tracks the state of the process, collects missing information, records the reasoning behind its conclusions, and, when necessary, involves a human to resolve intermediate issues.

Most importantly, critical calculations and access-control mechanisms remain the responsibility of conventional software algorithms.

What Is a Business Agent Used For?

Imagine a sequence of actions in a business process that cannot be fully defined in advance. For example, the system may need to collect data from different sources, take the context of that data into account, determine the next step, handle exceptions, and explain the result to a colleague.

Can this be automated using conventional software code? Yes, but doing so would require developers to define a large number of branches, exceptions, and rules in advance. The result would either be intentionally limited automation covering only strictly defined parts of the process, a need to redesign the process around the automation, or a conclusion that automation is not economically viable.

This is where an AI-powered business agent can find its niche. It can act as a connecting layer within the process, combine heterogeneous stages, request missing information, identify deviations, and suggest possible solutions. Ultimately, it can provide a human employee with a prepared conclusion supported by reasoning and recommendations.

For businesses, this can reduce the amount of manual analysis, accelerate decision-making, and make complex business processes more transparent and manageable.

How Is a Business Agent Structured and How Does It Work?

Let's look at how such a business agent can be organized using Anthropic's Month-End Closer as an example. The solution helps accountants perform month-end closing procedures.

It is important to note that Anthropic's business agent is not designed to operate directly inside an accounting system or replace its underlying algorithms. It is intended to work in an external environment, while being able to connect to an accounting system via API.

The Month-End Closer is a set of several independent agents, each with its own rules and capabilities. When a user initiates the month-end closing assistant, the main agent orchestrator takes over.

It receives an initial data package, including information about the organization and the closing period, requests a preliminary balance sheet, account balances and turnover for the period, data from previous periods, and other relevant information. The data can be retrieved via API from the accounting system or loaded from Excel files stored on the user's computer.

Based on this information, the orchestrator creates a work plan. This may include determining which accruals need to be reviewed, which accounts require reconciliation of balances and transactions, where significant deviations exist, and so on.

The main work cycle then begins. The orchestrator assigns tasks to its subordinate sub-agents. The Month-End Closer uses three of them: 

1. Ledger Reader — an AI agent for reading source documents from the user's folder. Its functionality is strictly limited to reading and extracting data in a predefined format.

2. Rollforward — a calculation sub-agent that uses the preliminary balance and verified data from source documents to prepare an accrual register, draft accounting entries, explanations of balance changes, and comments on deviations. For example, it can reconcile the opening account balance with transactions, reversals, and other changes, and then verify whether the resulting figure matches the closing balance in the general ledger.

If part of a discrepancy remains unexplained or there is insufficient data to reach a conclusion, the result is returned to the orchestrator along with a list of unresolved issues. The orchestrator can request additional information, ask the sub-agent to perform another check, or escalate a specific question to a human.
This means that the process flow is not rigidly predefined. The next step is selected based on the results of the previous one.

3. Poster — the final sub-agent responsible for structuring and formatting the results. It saves the final Excel report to disk, after which the document is passed to a human for review and approval.

Two points are particularly important here.
First, the sub-agents have clearly separated access rights to data. This prevents individual agents from going beyond the scope of their assigned tasks.

Second, the role of the orchestrator is critical. It coordinates the work of its sub-agents in a non-linear manner until a sufficient result is obtained. This is one of the key characteristics of agent-based AI execution and is what distinguishes a business agent from a simple chatbot or AI assistant.

Moving to 1C

A similar approach could be implemented in 1C. An agent orchestrator could manage the overall process, while specialized sub-agents would access a limited set of tools provided by the application solution. They could retrieve register data, analyze documents, identify deviations, prepare draft adjustments, and generate a final report.
The purpose of the business agent would not be to replace the application logic of 1C, but to connect its individual capabilities into a single managed process, independently select the next steps, collect the evidence supporting its conclusions, and provide the user with a prepared and verifiable result.

This is precisely the implementation approach we have adopted in our own AI agent solution for 1C.

Conclusion

The future of business agents lies in the emergence of a new intelligent layer within enterprise information systems — one capable of working with data, accounting objects, and application logic.
Such a layer will be able to independently organize work with information, select the next actions, and involve a human whenever a decision is required.
For 1C, this approach is particularly promising. An agent integrated into the application solution could work not only with aggregated metrics and reports, but also directly with documents, accounting objects, and the business-process logic of the company.