Counterparty due diligence can be automated with AI like this: by Tax ID, an agent gathers information from official registries and commercial databases, rules generate risk flags, and a language model explains court cases and document discrepancies in plain English, while the responsible employee makes the final deal decision. The model should not assign a “reliability score” on its own: its job is to reduce manual data collection and reading, while the evidentiary value of the review for the Federal Tax Service and the court depends on preserved sources and documented logic for selecting the counterparty.
Prepared by AIrassvet using public sources, verified September 30, 2026. It is intended for CFOs, lawyers, security teams, procurement, and business owners who manually vet suppliers and customers and want to speed up the work without losing quality. This is not tax or legal advice: the review process for large and unusual transactions should be aligned with a lawyer or tax advisor. The AIrassvet section at the end describes the publisher’s services.
Contents
- What due diligence means
- What the Federal Tax Service checks when selecting a counterparty
- Why manual review does not scale
- Which sources to connect
- Where AI helps and where rules are needed
- Counterparty AI review architecture
- Review levels by transaction risk
- Dossier: what should remain after the review
- Step-by-step implementation
- Mistakes and limitations
- How to measure results
- Frequently asked questions
- Primary sources and review methodology
What due diligence means
Definition. Due diligence is the standard of conduct for a reasonable market participant who, before entering into a deal, checks the counterparty’s business reputation, ability to perform, and solvency, and can show why that specific counterparty was chosen.
In tax law, the key rule is Article 54.1 of the Tax Code of the Russian Federation. It allows a tax base to be reduced for a transaction only if two conditions are met at the same time: the main purpose of the transaction is not tax nonpayment, and the obligation is performed by the contract party or by a person to whom performance was assigned under the contract or by law. Paragraph 3 of the same article clarifies that signing primary documents by an unauthorized person or the tax violations of the counterparty itself are not an independent basis for denying deductions. Therefore, the tax authority assesses the totality of circumstances, including how the company selected the partner.
The review is needed not only for taxes. It reduces the risk of prepaying a fly-by-night company, shipment failure, account freezes due to transactions with suspicious companies, and claims against management for losses.
What the Federal Tax Service checks when selecting a counterparty
The tax authority’s approach is explained in the most detail in Federal Tax Service of Russia letter dated 03/10/2021 No. BV-4-7/3060@ “On the practice of applying Article 54.1 of the Tax Code of the Russian Federation”. In paragraph 14 of the letter, the due diligence standard is described as checking the business reputation, ability to perform, and solvency of the counterparty. Paragraph 15 lists circumstances that tax authorities must examine together. These include:
- the counterparty’s actual location and its production, warehouse, or retail space are unknown;
- there is no evidence of a real ability to perform the contract: capacity, qualified staff, assets, licenses, or SRO membership;
- the responsible employees do not have information about how the counterparty was selected and how the transaction was performed;
- there is no contact with the counterparty’s executive or authorized persons, or documentary proof of their authority;
- no analysis was conducted of information about the counterparty, including data posted on the official Federal Tax Service website as open data;
- there is no publicly available information about the counterparty’s real economic activity;
- internal approval procedures were violated and the contract terms are atypical for ordinary business practice.
This leads to an important conclusion for automation: the review must leave a trail. It is not enough to “check a service” — you need to preserve what information was obtained, on what date, who reviewed it, and why the transaction was considered acceptable. AI is especially good at collecting and formatting this trail.
Why manual review does not scale
In a typical company, counterparty review looks like this: a manager or lawyer opens several services, copies an extract, scans a list of arbitration cases, reviews financial statements, and writes a short conclusion in a memo or in the contract comments. The problems with this approach are:
- Different depth. One employee reads the substance of the disputes, another looks only at the number of cases.
- No re-check. The counterparty was checked for the first contract, and a year later it is already in bankruptcy proceedings.
- Loss of evidence. Screenshots and extracts are stored in email and are hard to find during a tax audit.
- Bottleneck. With a large flow of new suppliers or customers, review slows down sales and procurement.
- Documents are not cross-checked. The charter, power of attorney, licenses, and registry data are rarely matched against each other.
Automation solves these tasks not through a “smart score,” but through a single procedure that the system applies the same way to every counterparty.
Which sources to connect
Most of the information is available through official public services. Commercial aggregators combine these sources and add analytics, but it helps to know the primary source so you understand exactly what is being checked.
| Source | What it provides | How to use it in the review |
|---|---|---|
| Federal Tax Service Unified State Register of Legal Entities / Individual Entrepreneurs (EGRUL/EGRIP) | Registration data, executive, founders, address, OKVED, change dates | Compare against the counterparty’s documents; check recent changes of executive and address |
| Federal Tax Service Transparent Business | Consolidated company information from Federal Tax Service data | Quick initial profile |
| Federal Tax Service open data | Data sets on tax debt, tax offenses, special tax regimes, headcount, income and expenses, “mass” executives and founders | Automatic flags based on data sets |
| Register of Disqualified Persons | Persons prohibited from managing an organization | Check the executive |
| GIR BO | Financial statements | Revenue, assets, trends, and whether the deal is proportionate to the company’s scale |
| Arbitration case file | Court Disputes Involving the Company | Role (plaintiff/defendant), subject matter and dispute amounts, bankruptcy |
| Fedresurs | Bankruptcy records and material facts | Insolvency proceedings, liquidation notices |
| FSSP Enforcement Proceedings Database | Enforcement proceedings | Outstanding debts |
| Bank of Russia’s Know Your Customer platform | Level of risk of involvement in suspicious transactions | Check the “red” level by Tax ID |
| Registry of Unscrupulous Suppliers | Suppliers that evaded performance of government contracts | Critical for procurement and contracting |
About the Bank of Russia’s “Know Your Customer” platform reportsthat it classifies legal entities and sole proprietors into low-, medium-, and high-risk groups (“green,” “yellow,” and “red”), and anyone can check whether a company or sole proprietor is in the “red” groupusing the Tax ID lookup service. For businesses, this is an important signal: transactions with a “red” counterparty may lead to questions from your bank.
Beyond registry checks, it is also useful to review documents requested directly from the counterparty: articles of incorporation, the resolution appointing the manager, signing authority power of attorney, licenses and SRO extracts, and information about premises and personnel. This is exactly the type of cross-checking with registries that AI is most often tasked with.
Where AI Helps and Where Rules Are Needed
The most common mistake is asking a language model to “assess a company’s reliability.” A model can get facts wrong, give different answers for the same data, and fail to explain how it reached its conclusion. It is more reliable to separate the work.
| Task | Who performs it | Why |
|---|---|---|
| Collecting data by Tax ID from registries and databases | Integrations and scripts (API, exports) | Deterministic, with date and source |
| Flags: disqualification, bankruptcy, “red” KYC level, debt, recent change of manager | Rules | Same result for everyone, easy to explain to an auditor |
| Scoring the likelihood of problems | Statistical model, if you have your own transaction history | Requires data and quality checks |
| Reviewing court cases: subject matter, role, amounts, outcome | LLM | The model quickly reads texts and creates a summary |
| Reconciling counterparty documents with registries | OCR + LLM + comparison rules | The model extracts fields, and the rules compare them |
| Searching for mentions in the media and on the company website | LLM with search | Initial screening of relevant materials |
| Final conclusion and decision on the transaction | Responsible employee | Legal responsibility and consideration of transaction context |
Principle: facts and flags come from sources and rules, paraphrasing and text comparison come from the model, and the decision is made by a person. Every model conclusion should link to the original document so the employee can verify it. For more on the reliability of extracting fields from scans, see the article OCR and document recognition for business.
Architecture of AI-Based Counterparty Checks
A working solution usually consists of five blocks:
- Trigger point. The check starts when a counterparty is created in 1C or a CRM, when a contract request is submitted, or on a schedule for rechecking active partners. How to connect agents to an accounting system is covered in the article How to connect AI agents to 1C.
- Data collector. Uses the Tax ID to request information from selected sources—official ones or through a commercial aggregator with an API—and stores the responses with the receipt date.
- Rules engine. Applies flags based on a preapproved list, for example: the manager is in the disqualified persons registry, bankruptcy proceedings are underway, the KYC level is “red,” the company was registered recently, and the transaction amount is disproportionate to its financial statements.
- AI analyst. The language model reads court decisions, counterparty documents, and found publications, produces a summary, and notes discrepancies. For sensitive documents, the model can be deployed within the company’s environment.
- Dossier and approval workflow. The system generates a report, attaches it to the counterparty record, and routes it for approval: “green” — automatically approved for small amounts, “yellow” — to legal, “red” — blocked or escalated to management.
It is important to decide in advance where the dossier will be stored. In a tax audit, what matters is not the model’s conclusions, but the saved information as of the transaction date and the signature of the responsible employee.
Review Levels by Transaction Risk
It makes no sense to review every stationery supplier and every general contractor with the same level of depth. The FNS letter explicitly links the expected depth of due diligence to the size and frequency of transactions, the nature of the goods and work, and the commercial terms. A practical approach is three levels, with thresholds set by the company itself.
| Level | When to use | What is checked | Who approves |
|---|---|---|---|
| Basic | Small one-time purchases from known suppliers | Registration, active status, disqualification, bankruptcy, KYC level | Automatically by rules |
| Standard | Recurring supplies and services, advance payment | Basic + financial statements, lawsuits, enforcement proceedings, signer authority | Responsible employee |
| Enhanced | Large transactions, licensed work, subcontracting, atypical terms | Standard + documents on capacity, personnel, licenses, SRO membership, company history, publications, meeting with the executive | Attorney and executive |
No specific threshold amounts are given here: they depend on the company’s revenue, industry, and internal risk policy.
Case file: what should remain after the review
The result of an automated review should be understandable to someone who opens it two years later. Recommended case file structure:
- Company details and review date. Tax ID, OGRN, legal name, date and time the information was received.
- Sources. List of requested sources with a note indicating a response was received or an error occurred.
- Rule flags. Which rules triggered and on what data.
- Summary of court cases. The company’s role, subject matter, amounts, and outcome with links to the case records.
- Document reconciliation. Whether the executive, address, details, and authority match the registries.
- Basis for the selection. Why this counterparty was chosen: price, experience, recommendations, available resources.
- Decision and responsible person. Who approved the deal and under what terms, for example without an advance payment.
The last two items can only be drafted by AI — an employee records them. It is the basis for the selection that answers the questions in item 15 of the Federal Tax Service letter about how the company learned about the counterparty and why it considered the counterparty capable of performing the contract.
Step-by-step implementation
Step 1. Describe the current process
Document how the review is handled today in procurement, sales, and by legal: which sources are checked, which signs are considered red flags, and where the result is stored. It often turns out there is no single standard process.
Step 2. Approve the rules and levels
Together with legal and finance, create a list of flags and review-level thresholds. This is a management decision and cannot be delegated to the model.
Step 3. Select data sources
Decide what comes from official services and what comes from a commercial aggregator with an API. Take data usage terms and request costs into account.
Step 4. Build a test sample
Take several dozen real counterparties with known outcomes: reliable partners, problematic deliveries, bankruptcies. Check which flags would have triggered and assess the quality of the model’s case summaries.
Step 5. Integrate with the accounting system
Launching the review when a counterparty profile is created in 1C or CRM and saving the case file in the same profile eliminates manual copying.
Step 6. Run a pilot in one department
For example, in procurement. Compare the time needed to prepare the review and the completeness of the case file before and after, and collect legal’s feedback on the model’s summaries.
Step 7. Add rechecks
Regularly rechecking active counterparties and notifying on changes — bankruptcy, a change in executive, new major lawsuits — is often more useful than the initial review.
Errors and limitations
- The model as an evaluator. An LLM-generated final “reliability score” without rules and sources cannot be explained to either the reviewer or the court.
- No data receipt date. Registries change; without a date, it is impossible to prove what was known at the time of the transaction.
- Blind trust in the aggregator. The indexes and ratings of commercial services are a useful signal, but their methodology belongs to the provider, and the decision remains with the company.
- Hallucinations in the summary. The model may incorrectly state the company’s role in a dispute or the amount. Links to the case record and spot checks are needed.
- Personal data. Executives and founders are individuals. Storing and transferring such data to external models requires compliance with Federal Law No. 152-FZ; see the article can client data be sent to AI.
- Formality instead of substance. A perfect case file does not replace real contact with the counterparty and verification of its resources for large transactions.
How to measure results
| Metric | How to calculate |
|---|---|
| Review preparation time | Measure before and after for the same types of counterparties |
| Case file completeness | Share of reviews where all required sections and sources are completed |
| Accuracy of model summaries | Share of court-case summaries without errors based on a legal spot check |
| Timeliness of rechecks | Share of active counterparties checked within the required period |
| Triggered alerts | Number of deals stopped or changed after flags, with a cause analysis |
The effect should be compared with the baseline measurement in your company. General approaches to AI solution metrics are in the article AI agent performance metrics, and the related task of contract analysis is covered in the article AI for the legal department.
Frequently asked questions
Can counterparty verification be fully entrusted to AI?
No. AI speeds up data collection, court case reading, and document reconciliation, but the decision on the deal and the responsibility remain with the employee. For tax purposes, the saved information as of the transaction date and the basis for selecting the counterparty are important.
What sources are needed to verify a counterparty by Tax ID?
The basic set is the Unified State Register of Legal Entities and Individual Entrepreneurs (EGRUL/EGRIP), open Federal Tax Service data, the register of disqualified persons, financial statements in the State Information Resource for Accounting Reporting (GIR BO), the arbitration case database, Fedresurs, the bank of enforcement proceedings, and a risk-level check on the Bank of Russia’s “Know Your Customer” platform. For procurement, the register of unscrupulous suppliers is added.
Does automated verification protect against Federal Tax Service claims?
There is no guarantee. Article 54.1 of the Russian Tax Code and Federal Tax Service Letter No. BV-4-7/3060@ assess the totality of circumstances, including the actual performance of the transaction. Automation helps collect and preserve evidence of due diligence, but it does not replace a real check of the counterparty’s capabilities.
How often should active counterparties be rechecked?
The company sets the frequency itself based on transaction volume and risks. Automation makes it possible to recheck active partners regularly and notify users of changes, such as bankruptcy or a change in executive.
Can ChatGPT be used to check a counterparty?
A general-purpose chatbot without access to registries can produce outdated or fabricated information. A reliable setup is to pull data from sources via API and use the model to present and compare that data with links to the original source.
What should you do if the counterparty is in the “red” zone of the Bank of Russia’s KYC platform?
This is a high-risk signal for possible involvement in suspicious transactions, according to the Bank of Russia’s assessment. Such a deal should be sent for deeper review and a manager’s decision, taking into account possible questions from the servicing bank.
Primary sources and verification method
The material was prepared using sources verified on September 30, 2026:
- Article 54.1 of the Russian Tax Code in the ConsultantPlus version as of the verification date.
- Letter of the Federal Tax Service of Russia dated 03/10/2021 No. BV-4-7/3060@ — paragraphs 13–15 on assessing counterparty selection; also available in ConsultantPlus.
- Description of the Bank of Russia’s Know Your Customer platform and TIN verification service.
- FNS Open Data Catalog — a list of datasets on tax debt, violations, special tax regimes, headcount, and reporting.
- Official services: the Unified State Register of Legal Entities (EGRUL), Transparent Business, the Government Accounting Information Resource (GIR BO), the register of disqualified persons, the arbitration case database, Fedresurs, the Federal Bailiff Service, and the register of unfair suppliers.
Methodology: the rules and the Federal Tax Service position are presented based on the text of the documents; the tables of sources, task distribution, and review levels are an editorial AIrassvet analysis, not legal requirements. Commercial verification services are not evaluated or recommended; the publisher does not claim any paid relationship with them.
The main limitation: the article has no original measurements of the speed or accuracy of automated verification. These indicators need to be determined on a sample of your counterparties.
How AIrassvet helps automate counterparty verification
If counterparty checks are slowing down procurement or sales, you can turn them into a single standard process with automated data collection and a clear dossier. On the AIrassvet website, which also uses the name “Vladimir Lomtev’s AI Studio,” the following areas are described as applicable to this task:
- Reviewing the current verification process and preparing the data: sources, flag rules, and levels.
- An AI agent that collects information by TIN and builds a dossier in the counterparty profile in 1C or CRM.
- Recognizing counterparty documents and matching extracted fields against registry data.
- Integration, testing on a sample of your counterparties, and training employees to work with the dossier.
The first step is to choose one verification flow, for example new suppliers, describe the current process, data sources, constraints, and the criterion you will use to consider automation successful.
Bottom line: how to start automating verification
AI-powered counterparty verification delivers the biggest impact when the model does not replace the process but carries it out: data comes from sources with timestamps, flags are assigned by approved rules, the model summarizes court cases and checks documents, and the responsible employee makes the decision and records the rationale.
Start by documenting current practices and approving review levels, connect TIN-based data sources, check quality on a sample of known counterparties, and launch a pilot in one department. That way, verification will become faster, consistent across all deals, and defensible during tax audits.