Automation of accounting with AI: maintaining control over data
Every fifth data extraction from an invoice may fail a simple arithmetic check.:Blind automation
For the chief accountant and the CFO, the question is not: "is it possible to do accounting work through AI?" The correct question is different: which areas can be removed from manual input, where to put control, how to maintain responsibility and not send the primary to an unknown cloud. AI helps well where it is necessary to recognize, classify, find, compare and highlight risk. But he does not become a responsible person, does not sign the primary document instead of the person and does not exempt the company from the requirements of the law.
remove the routine, speed up the processing of documents and leave the final decision to the person.
Main frame:
What AI does well in accounting
The second understandable scenario is classification and diversity. The AI can suggest a cost item, a division, a project, a document type, an associated contract, or an order. This is especially noticeable in companies where accounting departments sort through hundreds of similar documents every month from service providers, rental, logistics, marketing, communications, and travel expenses.
The third zone is data search and answers. "Instead""""""" Instead of a long manual review of folders and cards, the employee asks the question: "is there an act on this contract", "what is the debt owed to the counterparty", "when was the last shipment", "what documents are missing by the end of the period". Here, AI is valuable not as an author of transactions, but as an interface to accumulated data and documents.
The fourth scenario is anomaly control. The system highlights duplicate payments, amounts outside the usual range, atypical cost items, new counterparties, discrepancies in reconciliation, documents without closing pairs. Such checks are not a substitute for professional judgment, but they help to avoid missing the obvious in the flow of routine.
The fifth scenario is preparation for reconciliation and closing of the period. The AI collects a set of documents, shows the missing acts, offers drafts of explanations, and groups exceptions by reason. In this area, the effect is particularly noticeable not in the "magic automation", but in reducing the time for collection and navigation.
Vendor market figures look impressive, but they are important to read with attribution and methodology. Vic.ai claims 97-99% accuracy on invoices and 85% no-touch by the sixth month. When the Ramp Accounting Agent was launched, it reported 98% accuracy on transactions marked as "ready for synchronization" and a 70% reduction in corrections in the first month; at the same time, the methodology was disclosed: accuracy was considered as the proportion of marked transactions synchronized without subsequent edits. AppZen talks about agents trained to spend over $50 billion, and 87% automation cases for AP, that is, accounts payable — accounts payable and bill processing.
no-touchaccounting automation
What's really going on with accuracy
The main difference between good and dangerous automation is an honest conversation about accuracy. The header fields of the documents have already become a solved task in many ways: the supplier, number, date and total amount of most mature tools are extracted with an accuracy above 97% on standard documents. But accounting life rarely ends with a bill header. The real difficulty begins in the lines: nomenclature, quantity, price, VAT rate, discounts, line breaks, multipage tables, nested tables and ambiguous formulations.
every fifth extraction violated the arithmeticarXiv 2510.15727
In practice, the quality of speech recognition is highly dependent on repeatability. If a company receives similar documents from the same suppliers every month, the system can confidently keep the level above 90% after configuration and training. If a new supplier appears, a non-standard layout, a dense multipage table, or a row moved through a page break, accuracy decreases. Therefore, mature implementation realistically sets the range of 85-95%, rather than promising universal 99%.
Conclusion:
Therefore, the correct architecture is not a self—study, but a draft with a provable origin of each field. The accountant should see not only the recognition result, but also the fragment of the document from which the system took the value. Then the check does not turn into re-reading the entire file, but into a quick check of disputed places.
Why full automation is legally impossible
The legal framework is no less important here than the technological one. Federal Law No. 402-FZ "On Accounting" requires that the primary accounting document contain mandatory details, including signatures of responsible persons. Article 9 specifies that the primary document may be written on paper or in the form of an electronic document signed with an electronic signature. But an electronic signature is not a "model solution." This is the action of an authorized person or a representative with the appropriate authority.
A separate layer is machine—readable power of attorney, or MCHD. This is an electronic power of attorney in a machine-readable format that confirms the representative's authority when signing documents. Since 02/01/2026, the Ministry of Finance's order No. 1001 dated 11/05/2025 has been in effect: the structure of the details and the method of describing the powers have been changed, EDI operators are adopting a new format, including the universal format 003 and the reporting format of the Federal Tax Service 5.03. Previously issued MFDs are not being reissued, but new processes must take into account current requirements.
For automation, this means a rule of thumb.: AI can prepare a document, extract data, suggest wiring, highlight an anomaly, and collect an explanation. But the decision to conduct, the signature, and the responsibility remain in the controlled human circuit. This is not a reinsurance, but a normal design of the accounting process.
The Federal Tax Service also uses AI — what does this mean for you
The topic of AI in tax control causes a lot of noise, so it should be formulated carefully. The official line of the Federal Tax Service is that technologies are used to search for tax risks: the system learns from known risks and proven violations, analyzes VAT, income tax, personal income tax, STS, and helps inspectors see potential violators by the beginning of the declaration campaign. In professional publications in 2026, such statements were associated with desk control and a risk-based approach.
There is also a more skeptical position, including from former employees of tax authorities: what is called AI, in many cases is automatic data comparison, scoring, control ratios and analytics in the "Tax-3" AIS and VAT-2 ASC systems. For the taxpayer, this distinction is important more terminologically than practically. If your data systematically diverges, the primary is poorly designed, and the document chains are incomplete, the probability of getting into the sample increases regardless of whether the tool is called a neural network or scoring.
The legally essential boundary remains the same: a discrepancy in the system is a signal for verification, not a ready conclusion about a violation. In order to turn the signal into an additional charge, the inspectorate must prove the unreality of the operation and the taxpayer's awareness, including taking into account clauses 3.1 of art. 100 and 2 of art. 101 of the Tax Code of the Russian Federation.
The practical conclusion for the business is straightforward: the accuracy of the primary now has another price. It affects not only the internal closing speed of the period, but also how clean the company looks in automated checks. This is an argument for introducing AI into accounting, but only under the condition of control. A quick "random" entry does not create savings, but deferred tax risk.
Data control: where your documents are physically located
where does the document physically go?
the Russian cloudthe external APIpersonal data
When choosing a contractor or service, you should ask some mandatory questions.:
- Where is the document being sent to:
- What is logged:
- Who has access:
- Is the data used for training:
- What is the shelf life?:
Special attention should be paid to personal data. Primary documents often contain full names, signatures, passport details, addresses, and other information of Russian citizens. From 07/01/2025, an updated version of Part 5 of Article 18 152-FZ is in effect, related to restrictions on the use of foreign databases for recording, storing, clarifying and extracting personal data of citizens of the Russian Federation. In addition, turnover fines under Article 13.11 of the Administrative Code have increased since 05/30/2025: in case of repeated leakage, we can talk about 1-3% of revenue, but not less than 20 million rubles.
The conclusion for the project is simple: if the AI processes the primary, the architecture should be described as carefully as the architecture of the payment or personnel contour. Where files are stored, where recognition is performed, where logs are stored, who sees the results, how data is deleted, how access is restricted — these issues should be closed before the pilot, and not after the first leak.
How to build a process so as not to lose control
Secure integration of 1C accounting and AI is built as a chain of checks. The model should not be the only judge of quality. Its task is to extract data, propose a structure, and transfer the result to the process being tested. The higher the risk of an operation, the more control there should be before it is performed.
- Machine arithmetic verification.
- Checking with reference books.
- Rules for exceptions.
- Confirmation and implementation by a person.
An end—to-end requirement for such processes is the source for each value. If the system has extracted the INN, VAT amount, or contract number, the examiner should see a fragment of the document where the figure came from. This reduces fatigue and speeds up control: the accountant does not check the entire document from scratch, but only the risk fields.
To accept a project, it is useful to specify the criteria in advance: the permissible error rate in the header fields, a separate metric by line, the percentage of documents sent for manual verification, the processing time of one document, and the number of returns after completion. Without these criteria, the project easily turns into an argument about impressions.: ""it seems to have become faster" versus "it seems that the system is making mistakes."
A practical example: a company processes 3,000 incoming documents per month. After a pilot on 300 anonymized documents, it turns out that header fields are recognized with 98% accuracy, tabular lines — 89%, and 18% of documents are excluded. This is not a failure. This is a normal basis for calculation: documents pass faster without exceptions, complex ones remain with the person, and the system does not get the right to carry out questionable operations automatically.
How much does it save and how to count
It is better to consider the economics of automation not from the promises of the vendor, but from your own process. The basic formula is simple: the number of documents per month is multiplied by the average time it takes to manually enter one document. If an accountant spends 6 minutes on a document, and 3,000 documents are processed per month, manual entry alone takes about 300 hours. But automation does not remove all this burden. First of all, it reduces the input and initial posting, but the verification remains.
The realistic goal is not to remove the person, but to reduce the processing time of the document. For example, if the system prepares a draft, pulls up a counterparty, suggests a cost item and shows the source of each field, the verification may take 2-3 minutes instead of 6 minutes. For large volumes, this is already tens and hundreds of hours per month, especially during the closing period.
But you also need to consider the cost of the error. It's quick to fix an incorrect value in a draft. It is much more expensive to find an error after the audit, reassemble the register, correct the reporting, explain the discrepancy and restore the chain of documents. Therefore, savings from AI appear only when errors are caught before they are carried out, and not transferred to the accounting system with a beautiful speed.
There are also hidden cost items. The knowledge base needs to be maintained, reference books need to be cleaned, new suppliers need to be trained, and quality needs to be monitored. According to market benchmarks, support, customization, and monitoring can account for 20-50% of the cost of implementation per year. This is the normal price of mature automation, if it is budgeted in advance.
A good calculation model includes four indicators: how many hours of manual entry are removed, how many documents remain in manual verification, how many errors are caught before the procedure, and how much maintenance costs. If the project shows savings only in the first point and is silent about the other three, the finance director will rightly not believe the calculation.
What to do next
Automation of accounting with the help of AI should start from one site, rather than from the entire accounting department at once. Select a repeatable document type: incoming invoices, acts, UPD, receipts, lease documents, logistics, or regular services. Prepare 20-30 anonymized documents of the same type and run them through the pilot loop.
The purpose of such a pilot is not to get 99% advertising, but to see the real accuracy on your documents: separately on the header, separately on tabular lines, separately on VAT, amounts, nomenclature and comparison with reference books. After that, you can honestly say what percentage of documents will be processed quickly, how much will be spent on manual verification, and what exception rules are needed before launch.
It is this approach that removes the main objection of a cautious audience. An accountant and a financial director do not need to believe benchmarks or presentations. They need to see the quality of the extraction on their own documents, understand the boundaries of risk and make sure that control remains within the company.
Working formulation of the next step:
Автор
Антон Амосов
