AI helps government agencies and public institutions handle the flow of citizen inquiries at every routine stage: it extracts key details and the subject, classifies and routes the inquiry to the responsible party, identifies duplicate and collective inquiries, prepares a draft response based on the regulatory framework and prior responses, and monitors deadlines. The substantive decision and the signature remain with the official: the law requires the inquiry to be reviewed, not just answered. The task is becoming more urgent because citizens themselves have started writing inquiries with AI — and there are more of them.
Prepared by AIrassvet using open sources verified October 6, 2026. It is intended for heads and staff of citizen-request departments in regional and municipal government bodies and subordinate institutions, as well as IT teams selecting solutions for this work. This article is not legal advice: the procedure for handling inquiries is governed by Federal Law No. 59-FZ and the regulations of the specific agency. AIrassvet is not affiliated with the organizations or products mentioned, and the section at the end describes the publisher’s services.
Contents
- Why the number of inquiries is increasing
- What the law requires and where AI can help
- In brief: inquiry workflow stages and AI’s role
- Intake and classification
- Routing, duplicates, and collective inquiries
- Draft response: assistant to the reviewer, not an auto-reply
- Deadline tracking and issue analytics
- Risks and how to reduce them
- Where the model should run
- Implementation plan
- Which metrics to track
- FAQ
- Primary sources and verification method
Why the number of inquiries is increasing
On September 10, 2026, 3DNews summarized research by Chris Schmitz on a phenomenon the author called "agentic flooding": AI tools have made it easier to fill out forms and file complaints, and government services in different countries have faced a sharp rise in inquiries (3DNews, September 10, 2026). According to the publication:
- in the UK, complaints to the housing ombudsman have doubled since ChatGPT appeared;
- the U.S. Consumer Financial Protection Bureau recorded a fivefold increase in inquiries;
- the author analyzed 84 potential inquiry-surge cases across 11 jurisdictions; in most of them, the number of inquiries had barely changed before 2022 and then began rising faster and faster.
Schmitz does not directly claim that AI is the cause of the increase, and he emphasizes that a significant share of such inquiries are submitted by legitimate complainants who previously would not have filed a claim. There is no data on Russia in the publication, but the mechanism is universal: if it has become easier for a citizen to submit an inquiry, the agency needs to process it faster.
At the same time, Russia is developing centralized feedback collection. According to TAdviser, on July 2, 2025, Deputy Prime Minister Dmitry Grigorenko instructed the Ministry of Digital Development and the Analytical Center under the Government to create, on the Gosuslugi platform, a tool for collecting, analyzing, and monitoring digital feedback from citizens and businesses; agencies with their own systems must transfer data to the unified platform (TAdviser, "Feedback Platform (POS)"). The path to this was long: according to the same TAdviser, as early as April 2021 it became known that 7.46 billion ₽ would be allocated from the federal budget through 2024 inclusive to create a Unified Digital Feedback Window for receiving inquiries and complaints via the government services portal. Inquiries are increasingly arriving in digital form — and they are becoming easier to process automatically.
What the law requires and where AI can help
The basic rules are set by Federal Law No. 59-FZ "On the Procedure for Considering Appeals from Citizens of the Russian Federation." According to its brief description in the GARANT system, written inquiries must be registered within three days of receipt and reviewed within 30 days; the law also governs what agencies must do when an inquiry falls outside their jurisdiction (GARANT, 59-FZ). Exact wording, exceptions, and deadline extensions should be checked against the current version of the law and the agency’s regulations.
These requirements make it clear where AI is useful and where it is not:
- useful where the work is mechanical and high-volume: read, extract key details, determine the topic, find the responsible party and similar inquiries, and flag deadlines;
- useful as an assistant when preparing a response: find the rule, prior response, and data from the department;
- does not replace the official in substantive review, action-taking, and signing the response.
In brief: inquiry workflow stages and AI’s role
| Stage | What AI does | What remains with a person | Risk of error |
|---|---|---|---|
| Intake | Recognizes letters, scans, handwritten text, and messages from different channels | Review of illegible and disputed documents | Medium |
| Registration | Extracts full name, address, contacts, subject, and issue summary | Card verification | Low |
| Classification | Assigns the inquiry to the organization’s classifier topics and identifies multiple issues in one text | Disputed topics | Medium |
| Routing | Suggests the responsible employee or agency based on jurisdiction | Decision to forward | Medium |
| Duplicate detection | Finds repeat, collective, and large-scale similar inquiries | Decision on how to handle them | Medium |
| Draft response | Prepares a draft based on regulations, prior responses, and data from the responsible department | Substantive review, editing, signing | High |
| Deadline tracking | Sends deadline reminders, flags overdue and high-risk items | Management Decisions | Low |
| Analytics | Shows topic spikes and recurring issues by area | Findings and Actions | Medium |
Table — AIrassvet editorial classification; risk is assessed based on the impact of an error on the citizen.
Registration and Classification
Registration takes time on every request: you need to read the text, note the details, determine the topic and the core issue. An AI request classifier is a model that uses the request text to fill out the registration card and suggest topics from the organization’s classifier. The employee checks and confirms it instead of entering everything from scratch.
Practical details to build in right away:
- multiple questions in one request. A complaint about the yard, lighting, and trash pickup is three topics and possibly three assignees;
- scans and handwritten requests. Paper submissions are still very much part of the workload, so text recognition is needed; details on capabilities and limitations are in the article on OCR and document recognition;
- different channels. Mail, online intake, a public feedback platform, in-person visits — the card should look the same no matter where the request comes from. A similar business use case is covered in the article on AI automation for requests from email.
Routing, Duplicates, and Joint Requests
Routing errors cost days: the request goes to the wrong assignee, gets sent back, and the deadline keeps running. AI suggests an assignee based on the topic, address, and history of similar requests; if the issue is outside the agency’s authority, it flags that for a staff member to resolve.
The surge in AI-written requests adds a new task — clustering similar texts. The model compares requests by meaning, not just matching words, and shows:
- repeat requests from the same applicant on the same issue;
- joint requests and large volumes of similar texts;
- a spike in requests on one topic from one district.
Important: text similarity is not a basis for leaving a request unanswered. A person decides how to handle repeated and similar requests under the law and internal rules. AI only groups them so the assignee can see the full picture and prepare coordinated responses.
Draft Response: Helping the Assignee, Not Auto-Replying
The most valuable and most risky scenario is drafting a response. A well-designed system works on the RAG principle: the model looks for an answer not “in its memory,” but in an approved knowledge base — regulations, internal rules, past responses to similar questions, and data from departments — and shows which documents it relies on. More on the approach in the article on a RAG system, and on errors in the article on RAG accuracy and hallucinations.
To keep the assistant from turning into a boilerplate generator, it’s worth agreeing on rules:
- One answer per question. The draft is checked to make sure every question in the request has been answered.
- Only verifiable citations to rules and regulations. Every citation to an act must point to a document in the database; a made-up legal article is grounds to reject the draft.
- Facts come from the assignee. Repair timelines, inspection results, and actions taken are taken from department data, not filled in by the model.
- Plain language. The model is good at rewriting bureaucratic jargon in simple language, and that should be used.
- Human signature. The response is reviewed and signed by an authorized official.
Deadline Tracking and Issue Analytics
The review deadline is the most formal and most easily verifiable metric. Here AI works together with the electronic document management system: it counts days, reminds the assignee, and flags requests at risk of being overdue — for example, those where the assignee has not yet requested data. How AI is integrated into EDMS is covered in the article on AI for document management.
The second level is analytics. Requests are a map of local problems: where people complain more often about roads, where about heating, where the issue repeats year after year. AI turns free-form text into structured data that reveals spikes and trends. This follows the same logic built into the federal feedback analytics dashboard project: turning requests into a working tool for response. An overview of international and Russian experience with AI in public administration is in a separate article on AI in public administration.
Risks and How to Reduce Them
| Risk | How It Shows Up | How to Reduce It |
|---|---|---|
| Formulaic response | The draft answers “in general” rather than addressing the applicant’s questions | Check that all questions are covered; assignee accountability |
| Made-up rules | A citation to a nonexistent article or an outdated version | Responses only from the current document database; citation checks |
| Classification error | The request goes to the wrong assignee | Employee confirmation; monitoring the reassignment rate |
| Ignoring similar requests | Similar texts are left without review | Clustering for grouping only; decision made by a human |
| Personal data | Texts with full names, addresses, or health information are sent to an external service | Model hosted within the agency’s environment, access controls, audit logs |
| Bias | The model systematically downplays the urgency of certain topics | Regular sample checks, error analysis by topic |
Where the model should run
Incoming requests contain personal data and often health, family, and dispute-related information. That is why government agencies and public institutions usually consider a model deployed within their own secure environment or in a certified infrastructure, rather than a public cloud service. The choice depends on the specific agency’s information security requirements. Local deployment options and their costs are covered in the article on on-premises LLMs for Russian business.
The tradeoff is clear: a local model is usually weaker than the biggest cloud models, but for classification, extracting key details, and finding similar requests, its capabilities are generally enough. The hardest tasks—drafting responses—should be tested on the agency’s real requests before launch.
Implementation plan
- Flow audit. How many requests per month, from which channels, on which topics, how many reassignments and overdue items. Without baseline numbers, the impact cannot be proven.
- Choose one stage. Agencies usually start with intake and classification: the risk is low, and the effect is easy to measure.
- Data preparation. Topic classifier, assignee directory, and an anonymized sample of past requests for testing.
- Pilot in suggestion mode. AI suggests the record, topic, and assignee; the employee confirms or corrects it. All corrections become training material for improvement.
- Scale-up. Duplicate detection and deadline monitoring, then draft responses for a limited set of standard topics.
- Regular review. Monthly sample review of records and responses, error analysis, and updates to the document base.
What metrics to track
| Metric | How to calculate | Why it matters |
|---|---|---|
| Intake time | Median minutes from receipt to a registered record | First-stage impact |
| Classification accuracy | Share of records where the employee did not change the topic / checked records | Model quality |
| Reassignment rate | Requests transferred to another assignee / all requests | Routing quality |
| On-time response rate | Responses sent within the deadline / all responses | Main formal KPI |
| Edits to draft responses | Share of drafts accepted without major edits | How useful the assistant is to the employee |
| Repeat requests | Share of repeat requests on the same issue | Indirect measure of response quality |
Frequently asked questions
How does AI help process citizen requests?
AI identifies requests from different channels, extracts key details and the substance of the issue, classifies the request by topic and suggests the assignee, finds duplicate and group requests, drafts a response based on the regulatory base and past answers, tracks deadlines, and shows analytics by topic and territory.
Can AI respond to citizen requests on its own?
It should not. Under Federal Law No. 59-FZ, the request must be reviewed on its merits, and the response is signed by an authorized official. AI can prepare a draft response based on an approved document base, but review, decision-making, and signature remain with a person.
What response deadlines are set by 59-FZ?
According to the short summary of the law in the GARANT system, written requests are registered within three days of receipt and reviewed within 30 days. Exceptions and deadline extensions should be checked against the current version of the law and the agency’s internal rules.
Why are citizen requests increasing?
One reason is AI tools that make it easier to draft complaints. According to a study by Chris Schmitz, as summarized by 3DNews in September 2026, the number of requests to government services in a number of countries rose sharply after ChatGPT appeared; a significant share of these requests are submitted by legitimate applicants.
Can citizen requests be processed in a cloud neural network?
Requests contain personal data and sometimes health information. For government agencies and public institutions, the usual choice is a model running in their own environment or in infrastructure that meets information protection requirements. The specific solution depends on the agency’s requirements.
Where should you start when implementing AI for citizen requests?
Start with a flow audit and a pilot at one stage—usually intake and classification—in suggestion mode: AI proposes the record and assignee, and the employee confirms it. Then add duplicate detection, deadline monitoring, and draft responses for standard topics.
Primary sources and verification method
Verified on October 6, 2026:
- GARANT, Federal Law No. 59-FZ — short summary of the deadlines for registering and reviewing written requests.
- 3DNews, September 10, 2026 — summary of Chris Schmitz’s study on “agentic flooding.”
- TAdviser, “Feedback Platform (FOS)” — directive dated July 2, 2025, to create a feedback analysis tool based on Gosuslugi and funding for the Unified Digital Feedback Window (2021).
The data on request growth refers to foreign jurisdictions and is presented based on a summary of the study; the source does not provide Russian statistics. The tables of stages, risks, and metrics, as well as the implementation plan, are AIrassvet’s editorial analysis; the article does not include its own testing or project results.
How AIrassvet helps government agencies handle requests
If the citizen requests team is drowning in intake, losing time on reassignments, or failing to prepare responses on time, the task can be broken down step by step. On the AIrassvet website, which also uses the name “Vladimir Lomtev’s AI Studio,” the following applicable areas are described:
- Recognition of scanned and handwritten requests with extraction of key details into the intake record.
- A RAG system based on the agency’s regulatory base and past responses to prepare draft replies with document references.
- Deployment and fine-tuning of a language model within the agency’s environment for classifying requests containing personal data.
- Agentic automation of routine actions in the document management system interface: filling out records, assigning responsible staff, and sending deadline reminders.
The first step is to choose one stage, for example intake and classification, and define its current metrics, data sources, constraints, and the criterion by which you will consider the pilot successful.
Bottom line: process faster, not just reply faster
Requests are increasing, in part because it has become easier for citizens to write them with AI. The agency’s response is to use the same technologies for the routine work: intake, classification, routing, duplicate detection, deadline monitoring, and drafting responses.
The line is drawn at responsibility: substantive review, decision-making, and signature remain with the official, while AI works within the agency’s environment with verifiable document references. It makes sense to start with intake and classification in suggestion mode—this delivers measurable results and poses almost no risk to citizens.