MMOOZZGG.com

How to build an AI agent

AI agent for accounting: what it does and what it cannot sign

5 min read6 sources

An AI agent for accounting sorts incoming documents, checks their details, prepares reconciliations and answers staff questions about accounting. It should not post documents or sign reports: under 402-FZ, the CEO organizes the accounting and assigns it to the chief accountant, another official or a firm under contract.

On this page · 5 sections

A developer is needed

Which accounting work the agent takes

Accounting has a lot of work where a person reads a document and copies data from it: invoices, acts, delivery notes and letters from counterparties. These are tasks for an agent. Decisions on taxes and entries stay with the accountant.

Where teams usually start: sorting incoming documents, reconciliation with counterparties, answers to staff about payments and business trips, and reminders of payment deadlines under contracts.

To answer questions about internal rules, the agent gets a knowledge base: the accounting policy and the rules on business trips and bonuses. It works by the RAG principle: it finds the right passage in your documents and answers from it with a reference to the clause.

A report for the CEO on money and deviations from plan is built by another agent with its own access. It is described on the analyst agent page.

Incoming documents and their details

Under Article 9 of the accounting law, every business transaction is recorded in a primary document. It has seven required details: the title, the date, who drew it up, the content of the transaction, the quantity or amount with units, the positions of the people responsible and their signatures.

The agent checks an incoming act or delivery note against this list and compares the amounts with the invoice and the contract. A document without a date or a signature, or with an amount that does not match the invoice, goes to the accountant with a note on what exactly is wrong.

The agent first turns scans into text, and on a poor scan it can get a digit wrong. How data extraction works and where it breaks is covered on the document agent page.

Accounting documents are kept for at least five years after the reporting year, under Article 29 of the same law. So the agent works with copies and links to the archive, and it is not allowed to delete or change the originals.

Reconciliations and staff questions

The agent prepares a reconciliation statement with a counterparty from an export of settlements: it matches payments and shipments for the period, marks the differences and writes a draft letter to the counterparty. The accountant checks it and sends it.

The accounting team gets staff questions every day: when the salary will come, which documents an expense report needs, how to arrange a business trip. The agent answers from the accounting policy and internal rules and shows the clause the answer is based on.

If the documents have no answer, the agent does not invent one but passes the question to an accountant. This rule goes into the agent’s instructions from day one and is tested on test questions before launch.

Employees’ salaries, passport details and sick leave are personal data. Under 152-FZ, Russia’s personal data law, it is collected and stored in databases in Russia, and a notice goes to Roskomnadzor, the Russian data regulator, before any transfer abroad. How we work with such data is described on the page about data security.

Connecting to 1C

While the agent works with exports, you can build it in a builder without a programmer. When it must read account balances itself, create draft documents and match them with the bank statement, you need a connection to 1C, the Russian accounting system, and a developer.

The connection is made through an API, the official channel for data exchange between programs, or through MCP. MCP is a protocol that Anthropic opened on 25 November 2024 as one standard way to connect AI to data and services.

Rights follow the rule “read everything, write only drafts”. The agent creates a document as a draft, and the accountant checks and posts it. What the agent does well in 1C and what you must not trust it with is covered in the article on AI in 1C.

We launch an agent connected to 1C as a pilot in 14 days, from $1,499, with the metric in the contract. For example, how many incoming documents a day the accountant receives already checked.

Who is responsible for the books

The accounting law names the responsible person directly: the CEO organizes the accounting and the keeping of documents. The CEO assigns the accounting to the chief accountant, another official or a firm under contract.

An agent is not on this list. So every transaction it prepares goes through a person, and the agent’s log shows who approved it and when.

In practice the boundary is this: the agent reads, checks, prepares drafts and answers questions; the accountant posts, signs, files reports and writes to the tax office. If a contractor offers an agent that posts documents by itself, ask who is responsible for an error in the reports.

If an outsourced accounting firm keeps your books, the agent works on your side: it collects documents and answers staff, and the firm receives a complete set with no gaps.

Questions and answers

Will an AI agent replace an accountant?

No. Under 402-FZ, Russia’s accounting law, the books are kept by the chief accountant, another official or a firm under contract, and an agent is not on that list. The agent takes over reading and checking documents and answering standard questions, while entries and reports stay with a person.

Can you build an accounting agent without a programmer?

Yes, if the agent works with files: it answers from the accounting policy or compares an act with an invoice in uploaded documents. A link to 1C and the bank needs a developer, because access rights, drafts and a log must be set up. Do not upload payroll sheets or employees’ personal data to public chatbots.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. ConsultantPlus, 402-FZ, Article 7, 6 December 2011 (consultant.ru)
  2. ConsultantPlus, 402-FZ, Article 9, 6 December 2011 (consultant.ru)
  3. ConsultantPlus, 402-FZ, Article 29, 6 December 2011 (consultant.ru)
  4. arXiv 2005.11401, 22 May 2020 (arxiv.org)
  5. Anthropic, 25 November 2024 (anthropic.com)
  6. Brace Law Firm, 9 July 2023 (brace-lf.com)

Found a mistake? Tell us

Publications

Contact us