AI in claims handling is useful in three areas: it recognizes and classifies inquiries from all channels, extracts order data and the essence of the claim, calculates the legally required response deadline, and prepares a draft resolution according to company rules. The decision to pay, refuse, or send for inspection remains with the employee, while deadlines and amounts are calculated by rules, not by the language model. This setup reduces manual review and lowers the risk of missing a deadline that triggers statutory penalty interest.
Prepared by AIrassvet and verified against public sources September 30, 2026. It is intended for e-commerce managers, retail chains, service and distribution companies, customer service teams, and legal professionals. The rules are based on the versions of the laws in ConsultantPlus as of the verification date; this is not legal advice, and decision templates for disputed cases should be reviewed with a lawyer. The AIrassvet section at the end describes the publisher's services.
Contents
- Why claims are a good use case for AI
- What deadlines the law sets
- The claims workflow: from inquiry to decision
- What AI does at each stage
- Inquiry classification
- Photos and documents: where computer vision helps
- B2B claims: supply and contract disputes
- Decision rules and automation boundaries
- Return reason analytics
- Step-by-step implementation
- Claims handling metrics
- Common mistakes
- Frequently asked questions
- Source documents and verification method
Why claims are a good use case for AI
Claims and returns combine three characteristics that make them well suited for automation:
- Lots of unstructured text. Customers write by email, chat, messenger apps, on marketplaces, and by phone. The same issue can be described in dozens of different ways.
- Strict deadlines. For many consumer claims, the law sets a specific deadline, and missing it leads to penalty interest.
- Repeated decisions. Most inquiries fit a limited set of scenarios: defective product, wrong item shipped, damage in transit, item doesn't fit, delay, incomplete order.
At the same time, a claim is not just a support ticket. A mistake in the response can cost money, and an improper refusal can lead to litigation. That is why AI here acts as an accelerator for triage and decision drafting, not as an autonomous judge. The general logic of AI in customer service is covered in the article AI for Customer Support.
What deadlines the law sets
For retail trade, the main document is the Russian Federation Law "On Consumer Protection." Below are the provisions that most often determine the claims workflow.
| Situation | What the law says | Rule |
|---|---|---|
| Request to refund money, reduce the price, or reimburse expenses because of a defect | Satisfy within 10 days from the date the claim is submitted | Article 22 |
| Missing deadlines under Articles 20-22 | Penalty interest 1% of the product price for each day of delay | Article 23 |
| The claim was not satisfied voluntarily and was awarded by a court | Fine 50% of the amount awarded by the court to the consumer | Article 13, Section 6 |
| Exchange of a non-food item of proper quality | Within 14 days, not counting the day of purchase, if the item was not used | Article 25 |
| Withdrawal from goods purchased remotely | Before delivery — at any time; after delivery — 7 days; 3 months, if the return procedure was not provided in writing upon delivery; refund the money no later than 10 days | Article 26.1, Section 4 |
| Goods with defects | Consumer rights and the claims procedure | Article 18 |
This table points to the key requirement for the system: for every inquiry, you need to record the date the claim was submitted, the type of claim, and the product price, and then automatically calculate the deadline and potential penalty interest. This is a rules task, not a model task: the model extracts fields from text, and the rules calculate dates.
The law has exceptions and special cases — for example, technically complex products, goods with individually defined characteristics, and separate product lists. That is why scenarios and templates should be developed together with a lawyer for your specific assortment.
The claims workflow: from inquiry to decision
A typical process consists of six stages:
- Receipt. The inquiry comes through any channel.
- Registration. A case is created with the date, customer, order, and channel.
- Review. The problem type, the customer’s request, and the evidence needed are identified.
- Review. The order, delivery, photos, and warranty terms are checked, and if needed, a quality inspection or expert review is scheduled.
- Decision. Refund, exchange, repair, discount, rejection with justification, or a request for additional information.
- Execution and closure. Refunds, product logistics, customer response, and reason tracking.
The biggest time losses usually occur in stages 2–4: the employee reads the correspondence, looks up the order in the records system, requests missing photos, and enters data manually. This is where AI delivers the greatest impact.
What AI does at each stage
| Stage | AI task | Tool | Control |
|---|---|---|---|
| Intake | Collect inquiries from email, chats, messengers, marketplaces, and transcribe calls | Integrations, speech recognition | Channel coverage |
| Registration | Extract the order number, product, purchase date, request, and contact details | LLM with validation against the order database | Match to the order in the records system |
| Analysis | Classify the problem type and request, assess sentiment and urgency | LLM or classifier | Sample category review |
| Review | Request missing photos and documents, analyze damage photos | LLM + computer vision | Employee decision on disputed photos |
| Decision | Select a scenario based on the rules and prepare a draft response | Rules + LLM | Employee approval |
| Execution | Create tasks for refunds, product pickup, and system updates | Integrations with 1C, CRM, and the payment system | Transaction confirmation |
For more on handling inquiries from email and analyzing them, see the article automating email requests with AI; for CRM integration, see AI agent for Bitrix24 and amoCRM.
Inquiry classification
Good classification is the foundation of the entire automation. It must answer two separate questions: what happened and what the customer wants.
Type of problem: defect or flaw, shipping damage, wrong item or incomplete order, product that doesn’t fit despite being of proper quality, delay or non-delivery, pricing or payment error, complaint about service or an employee.
Customer request: refund, exchange, repair, price reduction, reimbursement of expenses, delivery of missing items, complaint only with no request.
This separation matters because deadlines and the process depend on the combination. For example, “wrong size” in an online purchase and “defect found” lead to different rules and different checks.
Practical recommendations:
- start with 8–15 categories that cover the main flow, plus an “other” category for manual review;
- build a labeled sample of real inquiries and test classification quality on it before launch;
- if the model is not confident about a category, route the inquiry to an employee rather than to the closest category;
- review the categories when a recurring type starts accumulating in “other.”
Photos and documents: where computer vision helps
Many claims come with photos: damaged packaging, fabric defect, broken screen, item mismatch. Computer vision and multimodal models help:
- check whether the photo actually relates to the claimed product: is the label, SKU, or serial number visible;
- highlight the damaged area and describe it for the employee;
- compare the photo with product photos taken at shipment, if the company captures them;
- detect repeated use of the same images across different claims.
Limitations are important: a photo alone cannot reliably establish the cause of a defect — whether it is a manufacturing defect or caused by use. That is a quality control or expert review issue. The model only prepares the material for a decision. For more on approaches, see the article computer vision in retail, and for signs of abuse, see fraud and anomaly detection with AI.
B2B claims: supply and contracts
Claims handling between companies works differently. Deadlines and procedures are determined first and foremost by the contract, and for monetary disputes the general rule in part 5 of Article 4 of the Russian Commercial Procedure Code applies: a dispute over recovery of funds under a contract may be taken to arbitration court after pretrial settlement measures thirty calendar days after the claim is sent, unless another deadline or procedure is established by law or contract.
AI in B2B claims helps:
- extract contract terms on acceptance, deadlines for reporting defects, the claims procedure, and penalties;
- compare the claim with delivery notes, acceptance certificates, and correspondence;
- calculate the date after which a lawsuit may be filed, based on the specific contract terms;
- prepare a draft response or counterclaim for the lawyer.
This requires access to contracts and source documents, often in the accounting system. Legal review and signing remain with the lawyer; see also the article AI for the Legal Department.
Decision Rules and the Limits of Automation
The higher the cost of an error, the less autonomy there should be. A practical approach is three decision modes that a company configures for its own amounts and risks.
| Mode | When to Use | What the System Does |
|---|---|---|
| Automated decision | Standard cases within policy: timely return of a product of acceptable quality, low amount, clean customer history | Approves according to the rules; the employee sees it in the log |
| Draft for approval | Defect, damage during delivery, amounts above the threshold, incomplete data | Prepares the response and decision; the employee approves |
| Manual review only | Dispute over the cause of the defect, signs of abuse, threat of litigation, claims for compensation for damages | Collects materials and passes them to a lawyer or senior specialist |
Two rules are mandatory in any mode:
- A refusal is never sent without a human. An unjustified refusal increases the risk of liquidated damages and a penalty under Article 13 of the law.
- Deadlines are calculated by rules, not by the model. The model may misread a date in the text, so the date the claim was submitted is recorded by the system at registration and verified by an employee.
Return Reasons Analytics
Claims are a source of data on product quality, suppliers, and logistics. When each request is classified and linked to a product, you can regularly answer questions such as:
- which products and suppliers have the highest defect rates;
- where damage occurs during delivery and with which carrier;
- which product listings drive “doesn’t fit” returns because of inaccurate descriptions or sizing charts;
- how the volume of claims changes after assortment or packaging changes.
A language model helps group free-form customer comments into topics, but conclusions about root causes should be checked against operational data. For marketplace sellers, related scenarios are covered in the article AI for Wildberries and Ozon Sellers.
Step-by-Step Implementation
Step 1. Map the current process and rules
Create a map of channels, stages, and owners. Together with a lawyer, document the scenarios and applicable deadlines for your assortment.
Step 2. Collect and label a sample
Take several hundred real requests from past periods, label the problem type, claim, and final decision. This is the foundation for quality evaluation.
Step 3. Set up extraction and classification
Connect the model to the sample, compare its labels with human labeling, refine the categories and instructions. Route low-confidence requests to people.
Step 4. Implement deadline and mode rules
Deadline calculations, penalty calculations, and routing by decision mode must be deterministic and test-verified.
Step 5. Integrate with accounting systems
Connecting to the orders database, delivery systems, 1C, or CRM eliminates manual searching. For integration details, see the article how to connect AI agents to 1C.
Step 6. Pilot in draft mode
At the first stage, the system only prepares case files and drafts, while employees approve all decisions. Measure processing time and draft quality.
Step 7. Gradually expand automated decisions
Move only those scenarios into automatic mode where drafts are consistently approved without edits, and keep a log for spot checks.
Claims Handling Metrics
| Metric | How to Calculate | Why It Matters |
|---|---|---|
| Share of responses within the legal deadline | Claims resolved on time / all claims with a legal deadline | Direct penalty risk |
| Time to first response | From registration to the first substantive response | Customer experience |
| Time to resolution | From registration to closure | Productivity |
| Classification accuracy | Match with employee review on the sample | AI quality |
| Share of drafts without edits | Approved without changes / all drafts | Readiness for automation |
| Share of disputes that reach court | Lawsuits / claims | Decision quality |
| Return reasons by product and supplier | Category distribution | Quality management |
Compare the metrics with the pre-pilot baseline. General principles for measuring impact are described in the article AI Agent Performance Metrics.
Common Mistakes
- A bot that argues with the customer. Automated refusals trigger escalations and increase legal risk.
- Deadlines estimated by eye. If the date the claim was submitted is not recorded by the system, missing the deadline cannot be prevented, nor can the opposite be proven.
- One category: “complaint.” Without separating the issue from the request, it is impossible to apply the rules correctly and get useful analytics.
- Closed data in an open model. Claims contain customers’ personal data; the processing framework must take Federal Law No. 152-FZ into account. Read more in the article can customer data be sent to AI.
- Conclusions based on a photo without verification. The cause of the defect is determined by quality inspection or expert review, not by the model.
Frequently asked questions
Can AI make claim decisions on its own?
For standard cases within the approved policy, yes, according to the rules and with logging. Disputed cases, refusals, and large amounts should be approved by employees. A refusal without human involvement increases the risk of a penalty and court-ordered fine.
What is the response deadline for a consumer claim?
It depends on the request. Under Article 22 of the Consumer Protection Law, requests for a refund, a price reduction, or reimbursement of expenses caused by a defect must be satisfied within ten days. Other requests and special products are governed by their own rules, so the scenarios should be agreed with a lawyer.
How many days does a buyer have to return a product purchased online?
Under Article 26.1 of the law, a buyer may refuse a product of proper quality before delivery at any time, and after delivery within seven days. If the return procedure and deadlines were not provided in writing upon delivery, the period extends to three months. The money must be refunded no later than ten days after the claim.
Can AI tell a manufacturing defect from damage caused by the buyer?
AI can prepare the materials: check whether the photos are relevant, highlight the damage, and compare it with the shipping photos. It is not possible to reliably determine the cause of a defect from a photo alone — that is the job of quality inspection or expert review.
What data is needed to get started?
A history of inquiries with final decisions, access to the orders and delivery database, the return policy, and a list of scenarios approved by a lawyer. For photo analysis, product images from shipment are useful.
Is this kind of automation suitable for B2B?
Yes, but contracts and primary documents are the foundation. AI extracts the acceptance and claims-handling terms, matches the claim against delivery notes, and prepares a draft for the lawyer, while deadlines are calculated according to the contract terms and the Commercial Procedure Code of the Russian Federation.
Source materials and verification method
The material was prepared based on sources verified on September 30, 2026:
- The Russian Federation Law “On the Protection of Consumer Rights” in the ConsultantPlus version: Article 13, Article 18, Article 22, Article 23, Article 25, Article 26.1.
- Commercial Procedure Code of the Russian Federation, Article 4 — pretrial procedure for monetary claims.
Methodology: deadlines and penalties are given according to the text of the articles; process stages, task allocation, decision modes, categories, and metrics are editorial analysis by AIrassvet, not legal requirements. The recommendations on sample size mentioned in the text are planning benchmarks, not measured values.
The main limitation: the article does not include its own measurements of the speed or accuracy of automated claim processing. The effect should be tested on your own inquiries and compared with the initial baseline.
How AIrassvet helps automate claims and returns
If claims are handled manually and deadlines are tracked from memory, the process can be rebuilt around a single inquiry record and rules. On the AIrassvet website, which also uses the name “AI Studio by Vladimir Lomtev,” the following areas applicable to this task are described:
- An AI agent that receives inquiries from messengers, email, and telephony, extracts order data, and creates a CRM record.
- Analysis of photos and claim documents using computer vision to help an employee prepare the decision.
- Automation of actions in 1C and other systems: order lookup, return tasks, and reason tracking.
- Process audit, integration, testing on historical inquiries, and employee training.
The first step is to choose one inquiry flow, for example e-commerce returns, describe the current handling, data sources, constraints, and the criterion by which the pilot will be considered successful.
Bottom line: where to start with claims automation
AI-powered claims and returns automation works when the model speeds up reading and preparation, the rules handle deadlines and amounts, and people handle disputed decisions and refusals. This division both reduces manual work and lowers the risk of a late-payment penalty for missed deadlines.
Start with a process map and scenarios approved by a lawyer, label historical inquiries, set up classification and deadline calculation, then launch a pilot in draft mode. Add automatic decisions only where the drafts are consistently accepted without edits.