S2S2 DIGITALSMEKH · SOLONENKO
Back to blog

Internal AI Assistant for Construction Company: 1C and Bitrix24 Integration

Published: July 29, 2026·19 min read

ai-automationrag-search1c-integrationdocument-automationdata-security

An internal AI assistant is not a cloud ChatGPT or Copilot. It is an agent that runs on your server, understands your company's internal documents, knows the structure of 1C and Bitrix24, and helps employees find information and assign tasks directly from a messenger. For construction and logistics companies, such an assistant is especially valuable: it speeds up searches for contractor data, helps prepare documents, and does not send confidential information overseas.

Why not just ChatGPT?

Customs declarations, contracts and employee information must not leave your company and end up in OpenAI or Anthropic's cloud. The answer is your own agent running on your server, with direct access to internal data and an LLM hosted on Russian Yandex AI Studio servers.

In this article we break down the architecture, what such an assistant can actually do and cannot do, how to set it up and integrate it with your systems, and when AI helps versus when human review is still essential.

What is an AI assistant for a construction company and why you need one

An AI assistant for a company is a self-hosted application that lives within your organization and integrates with your core business systems: accounting software (1C), business collaboration hub (Bitrix24), and document repositories. An employee asks questions in a messenger or web interface using natural language, and the assistant processes the query through a language model. It searches your accumulated corporate data and either provides direct answers with citations or drafts a document for review. Unlike a general-purpose chatbot such as ChatGPT, this is a specialized assistant trained to work specifically with your company's knowledge base, data formats, and business workflows.

In construction, logistics, and supply chain operations, employees spend significant time searching for information across multiple systems and documents. They need to check whether a specific supplier exists in the accounting database, find technical alternatives to a material, verify payment history for a client, or confirm inventory levels at a warehouse. Each inquiry requires switching between applications, scrolling through email archives, or calling colleagues. An AI assistant answers all such questions in seconds and delivers the information directly within the chat environment.

What the Assistant Can Do

  1. Search across internal documents. The assistant indexes your accumulated manuals, policies, price lists, and example documents, then retrieves the specific information you need in seconds.
  2. Retrieve data from accounting and business systems. Answer questions in natural language about suppliers, orders, shipments, inventory levels, payment terms, and interaction history.
  3. Draft standard documents. Generate templates for customs declarations, sales proposals, and acceptance certificates based on your existing knowledge base and document formats.
  4. Automate task creation and tracking. Add tasks to your team tracker by simply telling the assistant what needs to be completed and who should handle it.
  5. Integrate with your existing channels. Work within Telegram, company messengers, or a web interface without requiring staff to adopt new tools or change their workflows.
Data Protection and Compliance

The fundamental difference from public LLMs like ChatGPT is that this assistant operates entirely within your company's infrastructure and does not connect to external clouds. Supplier information, contract numbers, customs declarations, and employee personal data are never transmitted to OpenAI or Anthropic servers outside Russia. Instead, the assistant uses Yandex AI Studio, a language model hosted in Russia, and processes all data on your own servers. This is essential for compliance with Russian data protection regulations and ensures that business confidential information remains secure and private within your organization.

Language models can produce inaccurate or even fabricated information, a phenomenon known as hallucination. For this reason, any draft documents related to contracts, legal matters, or customs procedures must be reviewed by a qualified person before submission. The assistant handles approximately 80 percent of the mechanical work, but humans retain final decision authority and responsibility. Answer quality also depends directly on how well your company's knowledge base is organized and kept up to date. If required information is not documented, the assistant simply cannot retrieve it.

In a recent project for a logistics client, we deployed an AI assistant that helps employees quickly search a large internal document library, retrieve supplier and shipment data from accounting systems in plain language, draft customs declarations based on existing templates and examples, and automatically create and track task tickets in the team management system. This demonstrated that without expensive, complex rule-based automation, an intelligent assistant can meaningfully reduce time spent on routine operations and significantly ease the workload on individual staff members.

RAG search: how the assistant finds answers in the company's document library

RAG, or Retrieval-Augmented Generation, solves the core problem of standard AI models: they only know what they were trained on. If you need to find an answer in a large library of internal company documents, a standard model cannot help. RAG works differently: the assistant quickly finds the relevant documents from your database, passes their context to the AI model, and the model provides an answer based on your specific data, not general internet knowledge.

In practice, it looks like this. An employee writes in chat: 'What were the contract terms with LLC Logistics-Nord for shipments in March?' The assistant searches the document repository for contracts, invoices, acts, and letters related to that counterparty during the requested period, extracts relevant excerpts, and formulates a precise answer within seconds. Without RAG, such a search would take minutes of manual review of emails and documents.

How search works in the assistant

  1. The employee's query is converted into a semantic vector, a numerical representation of the question's meaning
  2. The assistant searches the index for all documents and fragments that are semantically similar to this vector
  3. Found fragments are ranked by relevance, with only the most important ones sent as context
  4. The AI model analyzes these fragments and provides an answer, referencing specific documents
  5. The result appears in chat with a source citation, so the employee can immediately verify the original
Why this works for your knowledge base

RAG does not require retraining the model. It is enough to load documents into the index once, and search begins working immediately. As your database grows, the index updates automatically. This saves time and money that would otherwise be spent on fine-tuning or building your own model.

Types of documents that work well

CategoryExamplesRAG readiness
Contracts and agreementsCounterparty contracts, NDAs, terms of supplyHigh, if structured and up-to-date
Operational proceduresInstructions for customs declarations, checklists, regulationsHigh, if documents are current
Communication historyClient correspondence, approvals, lettersMedium, requires spam and duplicate filtering
Reference dataCounterparty directories, template proposals, templatesHigh, especially if well-classified
Customs and legal infoRate clarifications, previously completed declarations, training materialsMedium, requires periodic verification of accuracy

Search quality depends on document quality. If the database is full of unstructured text, duplicates, and outdated versions, the assistant will often make mistakes. Therefore, preparation is required before implementation: removing archived files, standardizing formats, adding tags and categories. This is labor-intensive but pays off with fast and accurate search throughout the system's lifetime.

Real limitation: hallucinations

The AI model sometimes 'invents' details, even when based on real documents. Therefore, when searching for legal terms, inventory calculations, or customs declaration data, the employee must always verify the result. This is not full automation but rather accelerated search and draft preparation. The human makes final decisions.

Data security in RF jurisdiction

Customs declarations, counterparty credentials, and contract amounts contain commercial and personal information. Sending such data to a cloud AI service outside Russia creates a risk of data breach and violation of Russia's personal data protection law. Therefore, in our architecture, RAG operates entirely within Russian territory: the document index is stored on the client's local server, the AI model works through Yandex AI Studio (Russian hosting), and communication stays within Russian borders. This requires more infrastructure management but guarantees data control.

Honest assessment: where RAG helps and where it does not

  • Helps: fast search across large document libraries when manual searching would take hours
  • Helps: pulling context when drafting standard documents, such as a proposal template
  • Helps: training new employees when the document database acts as a living reference guide
  • Does not help: when documents are few or poorly structured, search will be guesswork
  • Does not help: when the document database has not been updated in a long time, the assistant returns outdated information
  • Does not help: for tasks requiring calculations or analysis of new, previously unknown situations

In one of our projects in the logistics sector, we deployed RAG for searching contracts and customs declaration instructions. Instead of hunting through folders for agreements with a specific counterparty or rereading 50 pages of regulations, an employee simply asks the assistant. This delivered a noticeable speed gain but required two weeks of database cleanup and indexing. That is a fair price for a tool that then saves hours every week.

RAG is not magic but a practical tool to make large volumes of internal information accessible in just a few clicks. The question is not whether you need search, but whether your database is ready for this work.

Integration with 1C and Bitrix24: where the assistant gets data on contractors, shipments and balances

Integration with 1C and Bitrix24 is not a replacement for your systems, but an addition of a natural language layer on top. Instead of opening multiple forms in 1C or copying information between systems, your employee simply asks the assistant in plain words. The assistant calls the APIs of these systems, retrieves the needed information, and displays it in a familiar interface such as a messenger or web portal.

What the assistant can retrieve based on your queries:

  • Contractor information: details, INN, KPP, payment history, current balance
  • Data on specific shipments or orders: number, date created, delivery status, cost
  • Stock balances at all warehouses or by specific product categories
  • Payment and transaction history for each contractor
  • Information from Bitrix24: responsible person for the deal, last contact date, negotiation status

This information comes not from the Internet or a cloud-hosted LLM, but directly from your systems. The assistant connects to 1C and Bitrix24 through secure channels on your server, retrieves the response, processes it in the context of the query, and delivers the result to the employee in a clear format. Technically, the assistant framework uses an integration layer (MCP or function calling) to connect to the 1C and Bitrix24 APIs. You define which fields and directories are available for queries, and which data can be shown depending on the employee's role.

Example of how it works: an employee asks "Show me contractors we worked with in the new construction on Kosmonavtov Street over the last 3 months". The assistant translates this into the correct query to 1C, retrieves the result, filters it by dates, and displays it in a table. All data stays in the RF contour. The LLM runs on Yandex AI Studio (server in Russia), the framework is deployed on your server, endpoints are fixed in RF. Personal data and contractor information does not leave the country to cloud services abroad, which is critical for compliance with 152-FZ.

In practice, in one of our logistics projects, the assistant handles dozens of queries daily: warehouse balance checks, contractor information for document preparation, searching for needed acts or invoices in the database. Work that previously required switching between systems and manual data entry now happens in a single messenger window. This cuts time spent and reduces the risk of errors when staff transfers data between applications.

Honest limitations

If data in 1C or Bitrix24 is not structured, the assistant will not find it. For example, if a contractor is listed only as plain text in a comment instead of linked to the directory, or if a name has a typo, the search may fail. Integration requires preparation: directories must be in order, fields filled in, data logic structured. If information does not exist or is scattered across free-text fields, the assistant cannot help.

Help with document preparation: from a draft customs declaration to an invoice and quote

Task automation: assigning tasks to a tracker with a single phrase in a messenger

When dozens of tasks are created manually every day, employees spend time not on work, but on administration. Send a task to the tracker, pick an assignee, write a description, set priority. What if you asked the assistant to do it with one phrase in a chat?

The internal AI assistant integrates with Bitrix24 (and other trackers via MCP) and catches commands like 'Create a task for Sergey to check the invoice from Rooga and Kopyta LLC' or 'Send the quote for the Rostov client to approval queue'. The assistant parses the text, extracts the assignee, identifies the work type, composes a description from context, and creates the task in the tracker via API.

Typical automation scenarios

  1. Task from a new order. A manager tells the assistant about a deal, it automatically creates tasks for document prep, approval, and client delivery.
  2. Fixes over the phone. A client calls and finds a typo in a customs declaration. An employee says in chat: 'Redo the customs declaration for Moscow-Kazan cargo, fix the SKU'. The task goes straight to the tracker for the owner.
  3. Urgent work. Without intermediate forms, rush jobs land in work immediately and show in the tracker for the whole team.
  4. Smart assignment. The assistant learns from the message, chat history, and employee role. If legal approval is needed, it can propose or assign the right specialist automatically.

This table shows how the process changes:

What needs to be doneWithout assistantWith assistant
Create task in trackerOpen Bitrix24, fill form, pick assignee, saveWrite a phrase in chat
Account for order contextCheck 1C manually, type detailsAssistant pulls data from 1C automatically
Notify the teamCopy link manually, send to chatAssistant creates task, it shows right away
Confirm priorityTalk it over, change in trackerPhrase includes urgency, assistant sets priority

Under the hood, it works like this. The assistant gets the phrase, recognizes the intent (create, reassign, or defer task), extracts entities (assignee name, work type, deadlines), enriches context from databases (1C, Bitrix24, chat history), and calls Bitrix24 API with tasks.task.add. Everything goes into the audit log and shows in the activity feed.

Caveats and best practice

The AI may misidentify an assignee if several employees have similar names, or misread a deadline. In critical cases, review the task before it runs, or use a 'suggest and confirm' flow where the assistant proposes and a human approves. Quality works better when the team communicates clearly and uses standard terms.

Integration with the tracker gives the team real time savings on routine admin work and lets everyone focus on actual work.

Data security and compliance: why LLM operations stay within Russia

Customs declarations, counterparty data, and Russian residents' personal information fall under Russia's Federal Data Protection Law (Zakon 152-FZ). If an AI assistant sends this data to OpenAI or Anthropic servers overseas, you violate the law. Many teams start with open-source agent frameworks, then discover that the default configuration routes queries to foreign LLM providers. That is an unintended breach of confidentiality that spreads across your entire workflow.

Our solution runs the framework on the client's own server. The LLM operates on Yandex AI Studio, keeping all traffic within Russia's data perimeter. The endpoint is fixed and known. No stray requests to the internet happen by accident. Client documents stay local and are never transmitted externally. The RAG index lives on the server. Data fetched from 1C or Bitrix24 is processed in memory and not cached by any third-party service provider.

If you configure the assistant through a public service like OpenRouter, even if you select a Yandex model, the prompt and context still pass through intermediary servers. That is a hidden leak, hard to spot without a detailed security audit. Insist on seeing the data flow diagram and confirm that the LLM endpoint is hardcoded to a Russian provider.

MeasureHow it worksWhy it matters
Self-hosted frameworkRuns on client's serverData never touches the internet
LLM in Russia's perimeterYandex AI StudioCompliant with 152-FZ
Fixed endpointNo configurable routes to the internetImpossible to accidentally call OpenAI
Local RAG storageIndex lives on client's serverDocuments remain on-premises
Local context processingData from 1C/Bitrix24 processed in memoryMinimal transmission footprint
Internet isolationOptional air-gap or dedicated VPN tunnelCritical for highly regulated sectors

LLMs hallucinate. The assistant generates a draft of a customs declaration or quotation. A lawyer or accountant reviews it before submission. Answer quality depends on the knowledge base. If your documents lack the needed information, the assistant will produce wrong results. RAG demands corpus preparation. You need to chunk documents, clean noise, and keep the index current. The architecture itself is secure, but it does not prevent human oversight failures.

In one of our logistics projects, the assistant indexes a large library of typical templates and past communications with suppliers. Staff ask in Russian about cargo status, inventory levels, or terms with a specific counterparty, and get instant answers. Draft customs declarations are built from the accumulated knowledge base, cutting paperwork time and reducing entry errors. Crucially, no data leaves the company's server. Compliance becomes a built-in property of the workflow, not a separate checklist.

Leak check

If you see the assistant can call OpenRouter, OpenAI API, or other public endpoints, that is a red flag. Demand that your provider disclose the architecture and confirm that the LLM only operates within Russia's data perimeter. Ask to audit the dependency graph and configuration to ensure no fallback routes exist.

Honest limitations: where AI helps and where human verification is needed

We believe in honest conversation. An AI assistant is not a silver bullet that eliminates checks and responsibility. On the contrary, a realistic picture of where it truly saves time and where human control is necessary helps prevent problems and integrate the tool correctly into your workflow.

Where AI helps most effectively

  1. Search across large document libraries. Instead of browsing archives, an employee asks the assistant in plain language, and RAG retrieves relevant fragments in seconds.
  2. Quick information retrieval from systems. Need contractor details, inventory balances, or shipment history for a quarter. The assistant pulls data from 1C or Bitrix24 and presents it clearly.
  3. Drafting routine documents. A template for a customs declaration, quotation, or act based on accumulated knowledge. This saves hours before final review.
  4. Automating routine work through messengers. Create a task in Bitrix24, assign it to a colleague, capture important information. One sentence in Telegram or WebUI does it all.
  5. Accessibility. The interface is familiar, no new systems to learn. Ask it like you would ask a colleague.
ScenarioAI strengthWhat requires verification
Document searchSpeed and relevanceSearch completeness, source freshness
Contractor and shipment dataNatural language vs queriesData currency in the source system
Document draftEmployee time savingsAll legal and financial details
Task creationProcess speed and simplificationTask correctness and real-world accuracy

Where human verification is not a detail, it is the foundation

Customs declarations, proposals, contracts, financial reports. LLMs hallucinate, inventing details absent from documents or that should not be there. Before sending to clients or authorities, a competent person must review thoroughly. This is not criticism of the assistant, it is the reality of working with neural networks.

LLM hallucinations are normal, not exceptions

Hallucinations are not bugs, they are normal. A neural network can provide a seemingly credible answer about information absent from documents or contradicting them. This is why on critical stages (sending a contract, financial report, customs declaration), a human verifies, not the assistant.

The same applies to critical decisions. If a task involves deadlines, amounts, responsibility, contractor exclusion, or logistics changes, a responsible employee verifies it. The assistant only speeds up information gathering.

Assistant output quality depends entirely on your knowledge base quality. If documents are unstructured, outdated, written in different styles, or contain errors, the assistant reflects that. Preparing your document corpus is separate work: cleaning, annotation, periodic updates. Without it, RAG loses accuracy.

Implementation is phased. The first stage is often document search, the second data retrieval from systems, the third document drafting assistance. Not all companies will reach the third stage, and that is fine. For small teams where people already hold knowledge in their heads, an assistant may be redundant.

Case study: how a client in logistics automated document preparation and database search

The Challenge

One of our logistics clients faced a common problem: a growing library of internal documents, certificates of conformity, quotes, standard templates, sample customs declarations, scattered across different systems. Whenever an employee needed to draft a quote or review a counterparty's history, it meant toggling between 1C, Bitrix24, and local folders, spending 5-10 minutes just to locate the right document or information.

The Solution

We deployed an internal AI assistant built on a self-hosted agent framework integrated with Yandex AI Studio, 1C, and Bitrix24. The assistant indexes the client's document library using RAG and responds to natural-language questions via messenger or web UI. All data stays within Russian-hosted infrastructure, ensuring compliance with local regulations.

Capabilities in Action

TaskHow the Assistant Works
Search the full document libraryEmployee asks 'Show me a quote template for contractors.' The assistant searches the RAG index and returns 2-3 relevant samples in seconds.
Look up counterparty detailsQuery 'What's the VAT ID for Company X?' triggers a function call to 1C. The assistant displays current payment status, credentials, and transaction history.
Draft a customs declarationUsing the stored library of previous declarations and cargo data from 1C, the assistant generates a first version for human review and submission.
Create and assign tasksPhrase 'Remind the dev team about the SLA by Friday' creates a task in Bitrix24, assigns it to specific people, and sets a deadline.
Conversational memoryThe assistant maintains context within a session, asks clarifying questions, and tracks task status without jumping between systems.

Real-World Impact

Document retrieval now takes 20-30 seconds instead of 5-10 minutes. Drafts are prepared faster, yet every legally significant document (customs forms, contracts, invoices) requires human review before submission. The shift freed up time for higher-value work instead of system hopping.

How It Actually Works

The employee opens Telegram or the web interface and asks a question in plain language. Behind the scenes: the query goes to the self-hosted agent framework, which routes it to Yandex AI Studio for natural-language understanding. If it's a document search, RAG retrieves relevant items from the indexed library. If it's a data lookup, function-calling triggers API calls to 1C or Bitrix24. The response comes back in seconds, complete with source context.

Important Caveats

LLMs hallucinate. For any legally binding document, customs declarations, contracts, invoices, the employee must review and edit the draft before submission. Answer quality depends entirely on how well the knowledge base is organized and kept current. If your company's documents are outdated or poorly tagged, the assistant will reflect that. Initial setup requires time to prepare and clean the document corpus, and ongoing maintenance is essential.

Is It for Everyone?

No. The tool shines in logistics, construction, consulting, and other fields where employees spend significant time cross-referencing internal documents and systems. For teams with a small, simple document library or minimal system integration, a general search tool may be enough. For those managing hundreds of documents, multiple integrations, and routine data lookups, an internal AI assistant can be a substantial time saver.

The real win is not flashy automation. It is removing friction from the daily routine so skilled people can focus on judgment calls and strategy rather than navigation and copying data.

Questions

Why an internal assistant if I can just use ChatGPT Plus?

ChatGPT Plus stores queries on OpenAI servers, some information may be used for training. For sensitive data this is risky. An internal assistant means full control: data stays on your server, the LLM runs on Russian Yandex servers, traffic never leaves the country.

How does the assistant know about contractors and warehouse balances?

Through integration with 1C and Bitrix24. An employee asks about contractor payment history, the agent calls the 1C API, retrieves all interaction data, and responds in natural language. Technically this is function-calling or MCP tools that the assistant invokes as needed.

What if the assistant gives wrong information?

AI can hallucinate. On financial and legal documents, an error in a draft requires human review before submission. You need to expect and plan for this. A good assistant can provide accurate answers on straightforward requests, but complex scenarios require verification.

How much does it cost to implement?

It depends on the scope of integration and data volume. An exact estimate requires an audit of your company's architecture. Contact the S2 team for a quote.

Can I deploy the assistant on my own servers?

Yes, that is exactly the point of this solution. The agent framework deploys on your server or corporate cloud. The LLM connects via Yandex AI Studio API. Your data never leaves your infrastructure.

What programming language do I need?

Agent frameworks are available in Python, TypeScript, Go. The choice depends on your current architecture. For most projects Python or TypeScript is sufficient, integration with 1C goes through REST API or built-in 1C language.

What about personal data? Does this violate Russian data protection laws?

Russian law requires protection of personal data. If the assistant runs on your server in Russia, data is not sent overseas and stays within your infrastructure, there is no violation. But you must document processing, conduct a data protection audit and appoint a data protection officer. This requires alignment with your company's legal and information security teams.

Read next
Back to blog