AI in 1C:Accounting: what the assistant takes over and where you can't trust it

AI in 1C:Accounting is an assistant that works with the data in your accounting database (1C being the accounting and ERP platform most Russian companies run on). It reads documents, matches them against reference lists, looks for discrepancies, answers questions in plain language and hands the result to an accountant for review. Today it can be given five tasks: entering primary documents, reconciliations with counterparties, finding errors and duplicates in postings, answering questions about the database, and drafting management reports. The assistant should not post documents or sign financial statements. By law, the head of the company and the chief accountant are responsible for the accounting, and the model makes confident mistakes precisely where a company has its own reference lists, customizations and exceptions.

Below, each of the five tasks is broken down into what the assistant does on its own and where its output needs checking. Separately, we show what 1C's built-in services can do and why they stop at the standard configuration. Then we cover where database data goes when you call an external model and who is responsible for an error the AI introduced into the books. At the end: what to prepare before starting and why an off-the-shelf assistant covers only part of the job. The article is written for business owners and CFOs; it does not cover code generation for 1C developers.

Which accounting tasks does AI handle inside 1C?

AI handles tasks with many repetitive operations, a reference example of the correct answer, and errors that can be caught before posting. In accounting there are five such tasks, and in each the assistant prepares a draft: a completed document, a list of discrepancies, a detected anomaly, an answer with a figure, or the text of a report. The decision about what goes into the books stays with the accountant.

Accountants themselves are still cautious. According to a survey by Tochka Bank's Finlead service and the Buhexpert platform (September 28, 2026, 1,248 accountants, 94% with more than eight years of experience), 3% of respondents are actively bringing AI into their work, 37% use it occasionally, 17% tried it without noticeable benefit, and 43% don't use it at all. Companies, meanwhile, are moving faster than their accounting departments. According to a study by SberAnalytics and Sber Business Soft (summarized by ComNews, January 22, 2026), 39% of organizations use AI agents and assistants. Among the processes they hand over to AI, accounting and financial reporting rank second: 55%, versus 70% for document workflow and request processing.

Primary documents: what does AI already do on its own, and what not?

Entering primary documents is the most mature of the five tasks, and it has an official 1C service. According to the description of 1C:Primary Document Recognition, the service accepts scans, photos and digital files, including PDF, Word and Excel. It recognizes invoices, TORG-12 delivery notes, certificates of completion, VAT invoices, UPDs and UKDs (Russian universal transfer and adjustment documents) and cash receipts. The final step of the process is spelled out plainly: check the result and create the documents in the accounting database.

The difficulty starts after the fields are recognized, when the document lines have to be matched to your item catalog. The supplier writes “50×50 angle steel”, the catalog has three similar items, and the model can pick the right one only by rules someone has written down. The accuracy ceiling, the document volume at which it pays off, and the cases where an off-the-shelf service isn't enough are covered in detail in the article on primary document recognition in 1C. We won't repeat that here.

How does AI help with reconciliations with counterparties?

1C handles matching documents by their details with its own tools, and AI takes on what a standard reconciliation doesn't cover: a counterparty's statement in its own format, and an explanation of the discrepancies. The assistant parses the PDF or scan that was sent, finds lines with no match in the database and suggests a reason. A document may not have been recorded, may have landed in a different period, or may have gone through with a different amount after an adjustment.

The deterministic part is already covered by the 1C:Reconciliation 2.0 service: it reconciles VAT invoices, primary documents (UPDs, UKDs, certificates of completion, delivery notes) and bank documents. According to 1C's description, no agreement with the counterparty about document exchange is required. It works rule by rule, without AI, and for counterparties that are in the service this kind of reconciliation is usually enough.

AI is useful for what's left over. That's a counterparty who sent the reconciliation statement as a scan, a discrepancy of a hundred lines over a year, or a supplier who keeps their books differently and closes a delivery with a single certificate covering several shipments. Here the assistant saves the hours an accountant spends searching line by line. The decision stays with a person: which side to correct, whether to write to the counterparty and how to record the adjustment. The model sees only your database and the other party's statement, and doesn't know what the sales manager agreed with the supplier over the phone.

Can AI find errors and duplicates in postings?

It can, provided each finding is backed by a link to a specific document. Standard errors are caught by 1C:Accounting's built-in checks, such as the express accounting check, which runs on predefined rules. AI adds a search for what doesn't fit a rule: a posting that's atypical for this counterparty, a single document entered twice, a payment that has been sitting without closing documents for six months.

Typical findings from such a search:

• Double entry. The same receipt was entered from a scan and also arrived via EDI (legally binding electronic document exchange between companies in Russia); the numbers differ by one character, the amounts and dates match.
• Atypical account pairing. A telecom supplier always went to one expense account, but in March its invoice went to a different one.
• Stuck settlements. An advance to a supplier with no matching delivery for longer than usual for this counterparty.
• Duplicate reference entries. One counterparty was created three times with different spellings, and settlements with it are scattered across three records.
• Mismatch with the source document. The amount in the document doesn't match the amount in the attached file.

There is one rule for checking: a finding without a link to a document doesn't count as a finding. A language model can plausibly describe an error that doesn't exist, so the assistant produces a list with documents, and the accountant opens each one and decides. The threshold for “atypical” is also set by the company. For a wholesaler with a thousand suppliers and for a service company with twenty contracts, normal looks different.

Can you ask a 1C database questions in plain language?

You can, and this area moved forward noticeably in 2026. The assistant turns a question like “how much do customers owe us that's more than sixty days overdue” into a database query and answers with a figure. The risk here is subtler than in the other tasks: the figure looks plausible even when it was calculated using a different definition than the CFO's.

Ready-made solutions already exist. Yandex launched Alice AI for Business on September 8, 2026: the assistant is available to more than 185,000 organizations using Yandex 360, and 1C connects through plugins alongside Bitrix24 (a Russian CRM and collaboration suite), amoCRM and SPARK. Access to external sources requires the user's permission, and rights are configured by an administrator. The company described a Yandex Cloud version hosted inside the customer's perimeter as a plan. Another announcement came on October 7, 2026: K2Tech announced an AI agent inside 1C:Enterprise solutions. The agent acts only within the rights of a specific user, asks for confirmation on sensitive operations and writes its actions to an audit log.

Those last three properties are worth requiring from any such assistant. One more requirement is written by the company itself: a glossary of metrics. Is “revenue” with or without VAT, with or without returns, across all legal entities in the database or for one, by shipment or by payment? Without a glossary, two identical phrasings of a question produce two different numbers, and neither matches the accounting department's report. A good assistant shows, next to the answer, which report or query produced the figure, so it can be recalculated.

Which management reports can AI assemble as a draft?

Those whose figures are already in the accounting registers, where the manual work goes into assembly and commentary. The assistant takes the trial balance, account analysis and counterparty data, assembles them into the familiar format and writes the text: what changed compared with last month and why. The CFO gets a draft that only needs checking.

The boundary is drawn by where the figures come from. Figures in the report are pulled from the registers by query; the model writes only the text around them. If you let it “finish calculating” a metric in the text, the error ends up in the owner's report and looks convincing.

The second boundary is management breakdowns that don't exist in the financial accounts. Allocating overhead to business lines, margin by project, splitting fixed and variable costs: every company does these its own way. If the rules aren't written down, the model will make them up, and differently every month. How to set up management reporting before connecting an assistant is covered in the article on management reporting, and which figures an owner should see on a single screen is in the article “The owner's dashboard”.

What can 1C's built-in services do, and where do they hit their limits?

1C's built-in services handle data entry and standard reconciliations well in a standard configuration. They hit their limits with customizations, custom reference lists and tasks that go beyond a single document. Some of the services mentioned in connection with AI in 1C have nothing to do with accounting at all.

The first three rows come from 1C's official solutions page for 1C:Accounting 8 and the service descriptions on the 1C website. 1C:Naparnik is described on its own page as a developer assistant in 1C:EDT. It requires an ITS account (1C's support subscription), is free until February 1, 2027 and, according to 1C, doesn't let data leave 1C's protected perimeter. It's the tool most often meant by the articles that rank for “AI in 1C”, but it does nothing for accounting.

Where the built-in services stop working:

• Customized configuration. A standard service doesn't know your custom fields, documents and registers. It will fill in the standard fields and leave your additional ones empty or fill them with the wrong values.
• Custom reference lists. Expense items, item groups and contract types set up for your analytics have to be explained to the model separately.
• Several legal entities in one database. The question “how much did we earn” without specifying the entity returns a total nobody asked for.
• Links to other systems. If orders live in the CRM, inventory in the WMS, and accounting receives them through data exchange, the answer to “why doesn't it add up” lies at the junction of systems, not in a single database.

Where does database data go when you call an external model?

Wherever the model runs. With a public foreign chatbot, the data goes to someone else's servers outside Russia; with a Russian vendor's cloud, to the vendor's servers under contract; with a model inside the company's perimeter, it stays on your servers. For a 1C:Accounting database this is a legal question: it holds personal data and, as a rule, trade secrets.

There is more personal data in an accounting database than it seems. It includes employees with passport details and salaries, if payroll is run in the same database, sole-proprietor counterparties, individual contractors and bank details. Sending such data to a foreign server is a cross-border transfer, and under Part 3 of Article 12 of the Personal Data Law (Federal Law 152-FZ, Russia's personal data protection law), the operator must notify Roskomnadzor, Russia's data protection regulator, about it separately and in advance. Purchase prices, margins and contract terms are often trade secrets. Under the Trade Secrets Law, their protection rests on measures taken by the holder itself, including agreements with those the information is shared with.

In practice, data leaves without any intent. According to a Solar Group study (February 2026, traffic analysis of 150 companies), the volume of data employees send to publicly available neural networks grew thirtyfold in a year. That data includes financial reports, legal documents and customer databases, and about 60% of organizations have no formal rules for working with AI.

The choice between these rows is made before choosing a model, because it rules out some solutions entirely. More on keeping AI inside the company's perimeter in the article “AI without data leaks”.

Who is responsible for an error AI introduced into the books?

The same people who were responsible for the accounting before AI. Under Part 1 of Article 7 of the Accounting Law, the head of the organization organizes the accounting and the storage of documents. Under Part 3 of the same article, they assign the accounting to the chief accountant or another officer, or sign a contract with an accounting firm. The model has no place in this chain. A document posted on its suggestion counts as posted by whoever clicked the button.

What follows from this is an architecture, not a ban. The assistant proposes, a person posts, and that split is enforced through access rights:

1. A separate user account for the assistant. Rights to read only and to create documents with the status “not posted”.
2. A period lock date. Closed periods are off-limits to the assistant, as they are to any employee without the right role.
3. An action log. What the assistant read, what it proposed, who confirmed it. Without a log, six months later there's no way to reconstruct where a posting in the books came from.
4. Confirmation of sensitive operations. Changes to counterparty details, adjustments to prior periods and bulk operations go through a person only.
5. A source link in every answer. A figure without a report, a finding without a document, a line without the original file are not accepted.

The contractor who built the assistant is responsible under the contract for how the system works: which rights are granted, what goes into the log, how data is processed. The company is responsible for the accounting data and the financial statements. This split is worth putting in the contract before launch rather than sorting out after the first error.

Where to start?

With one task out of the five and a description of what a correct result looks like. It's easiest to start with a task that has high volume and a clear reference example: reconciliations with major suppliers, double entry of primary documents, the monthly package of management reports. That description immediately shows what an off-the-shelf service will cover and what will have to be built for your database.

1. A task with a reference answer. What result counts as correct: a completed document, a list of discrepancies with reasons, a report in the approved format. A few examples of correct results produced manually by an accountant.
2. Clean reference lists. Counterparties without duplicates and with taxpayer IDs, an item catalog without duplicates, expense items following one rule. The model relies on what's in the database, and with dirty reference lists it will make confident mistakes.
3. A decision on data. Which sections of the database the assistant sees, which it never sees (payroll, purchase prices), where it runs: a Russian vendor's cloud or the company's perimeter. The business makes the decision; the contractor translates it into access rights.
4. Rights and a log. A separate user account without the right to post, a period lock date, an action log. Describing roles before development is easier than fixing them afterward.
5. A glossary of metrics. What counts as revenue, receivables, margin, and by what rule each is calculated. Without a glossary, answers to questions about the database diverge from the accounting department's reports.
6. A test copy of the database and integrations. A copy on which the assistant is tested before launch, and a list of systems the database exchanges data with: the bank, EDI, CRM, warehouse. Errors are cheaper to catch on a copy than in the live database on the day reports are due.
7. An owner after launch and ongoing support. An employee who looks after the reference lists, rules and glossary, and a support contract. Accounting policy changes, the configuration gets updated, and without an owner, within a quarter the assistant is answering by the old rules.

The company handles the first two items on its own, and they alone strongly affect the quality of answers. The other five show what makes up the project cost: how many customizations the database has, which data is restricted, which systems need to be connected. The general framework of stages and budgets is in the article on the stages and cost of AI adoption.

Off-the-shelf or a system built around your process?

An off-the-shelf service covers the standard part: recognizing a document in a standard format, reconciling VAT invoices with a counterparty, answering a question from a standard report. Companies' accounting diverges below that level, and that's where the assistant has to be built for your database. Every company has its own procedures, reference lists, access rights and exceptions.

At the start, companies are almost identical: primary documents need entering, settlements need reconciling, reports need filing. It makes sense to take this part ready-made, and a small accounting department on a standard configuration often finds 1C's built-in services sufficient for a long time.

The differences begin with customizations and breakdowns. One company has three legal entities in its database and its own expense items, another has a customized receipt document with fields for production, a third gets orders from its CRM through data exchange and has to reconcile two systems. One employee sees payroll, another only their own area. An off-the-shelf assistant knows the standard configuration, but not your customizations or your glossary of metrics. We covered the choice between a ready-made solution and a system built around your process separately: custom system or off-the-shelf.

INCUBE AI works under contract, keeps data in Russia and builds the assistant around your accounting. It works with your reference lists and customizations, receives rights according to your role matrix, keeps an action log and is connected to the CRM, warehouse and bank. After launch, the system remains under ongoing support. You can see what such a setup looks like on a live project in the project for a wholesale and manufacturing company: there, the item catalog, stock levels, shipments and payments sync with 1C without manual transfer. If your accounting department is drowning in reconciliations and manual discrepancy checks, tell us about the task: we'll take one area of your accounting, work through it on your database and tell you plainly where a 1C service is enough and where the assistant will have to be built for you.

Sources and links

• Klerk, September 28, 2026: survey by Tochka Bank's Finlead service and the Buhexpert platform — 1,248 accountants, 94% with more than eight years of experience: 3% actively bringing in AI, 37% using it occasionally, 17% tried it without noticeable benefit, 43% not using it; 29% consider replacing junior accountants possible, 57% expect qualification requirements to rise
• ComNews, January 22, 2026: study by SberAnalytics and Sber Business Soft — 39% of organizations use AI agents and assistants; document workflow and request processing 70%, accounting and financial reporting 55%
• 1C: Primary Document Recognition — file formats, document types, the step of checking the result before creating documents in the database
• 1C: Reconciliation 2.0 — reconciliation of VAT invoices, primary documents (UPDs, UKDs, certificates of completion, delivery notes) and bank documents with counterparties
• 1C: solutions for 1C:Accounting 8 — 1C:Primary Document Recognition, 1C:Receipt Scanner, 1C:Document Scanner in the “AI and predictive analytics” category
• 1C:Naparnik — developer assistant in 1C:EDT, access with an ITS account, free until February 1, 2027
• Rusbase, September 8, 2026: Yandex launched Alice AI for Business — available to more than 185,000 Yandex 360 organizations, plugins for 1C, Bitrix24, amoCRM and SPARK, a Yandex Cloud version hosted inside the customer's perimeter announced as a plan
• CNews, October 7, 2026: K2Tech launched an AI agent inside 1C — operation within user rights, confirmation of sensitive operations, audit log
• Habr, Solar Group, February 2026: study of data leaks through neural networks — traffic analysis of 150 companies: the volume of data in publicly available neural networks grew thirtyfold in a year, about 60% of organizations have no formal rules for working with AI
• Federal Law No. 152-FZ “On Personal Data” of July 27, 2006, Article 12 — cross-border transfer of personal data and notification of Roskomnadzor
• Federal Law No. 402-FZ “On Accounting” of December 6, 2011, Article 7 — organization of accounting by the head of the organization and its assignment to the chief accountant, another officer or an outside firm under contract