cGIS Pro is enhanced with an AI agent that allows any enterprise employee — engineer, dispatcher, designer, or contractor — to work with the company's entire array of spatial and engineering data using natural language: ask a question, receive an accurate answer, see the result on a map or in a 3D model, and find the desired section in the company's regulations.
Who needs it.
Industrial enterprises, infrastructure companies, developers, and government customers have hundreds of layers of spatial data, thousands of engineering infrastructure units, and regular tasks that require coordination between people and systems.
It is important to understand this right away.
The AI agent is an integral part of the cGIS Pro platform. We deliver it in two configurations: “platform + agent”for new customers and “agent as an extension”for those who already have cGIS Pro running.“Under the hood” is the open standard MCP (Model Context Protocol, developed by Anthropic, published in November 2024). cGIS Pro is one of the Russian spatial data platforms where the MCP server is already in production operation on real projects.
“Artificial intelligence” in 2024-2026 has become a word that requires clarification: a chatbot in technical support, the generation of marketing pictures, an assistant in an office package — everything is called AI.The head of production, responsible for information systems, has legitimate questions about these AI solutions: what they do with his data, how it is integrated with his infrastructure, and what measurable benefit it will bring.
This document is about a different class of solutions.It's about AI that works with your data and within your platform.Instead of describing a table with words, it builds a query for a layer, draws the result on a map, opens a 3D model, finds the necessary fragment in a corporate regulation, and generates a PDF resolution based on the dialogue.
Let's explore the fundamentals, how it works in a specific industrial scenario, the safety requirements met, and the limitations of the solution today.
A large language model (LLM) is a very well-educated conversationalist who doesn't know anything about your infrastructure. In order for it to work with your data, it needs to understand what you have and how to use it.
Before MCP was introduced, this problem was solved in three ways:
1. Prompt engineering - description of the API inside the system prompt.This approach is unstable: token consumption is high, the model is wrong, and the integration breaks when the API changes.
2. Custom connector for each model - separate integration for GPT, Claude, GigaChat, and YandexGPT. Expensive to develop and maintain.
3. Function calling in a vendor - specific formatis an improvement over the previous two, but without compatibility between models.
MCP is a standardized protocol for interaction between LLMs and tools.The MCP server declares its capabilities (a list of tools with parameter schemes), and the client, which can be any LLM that supports MCP, can use them.One server can work with Claude, GigaChat, and local Qwen on a corporate GPU.Replacing the model does not affect the server side.
The second key feature: the MCP server sets the recommended order of actions:not "choose randomly", but for the question about objects in a layer, "first find the layer, then get its structure, then request data".This increases reliability: the model does not guess the sequence, but follows a documented scenario.
As a result, MCP can be described as an operating system for AI agents: a unified standard on which all models and tools are based.
GIS, BIM, and the digital twin are large arrays of spatial layers with specialized attributes.In a large industrial enterprise, an engineer may have hundreds of layers in their platform: pipelines, cable networks, foundations, transformer substations, roads, railway tracks, security zones, repair schedules, video cameras, weather stations, and history of inspections.This is supplemented by an archive of project documentation, corporate regulations, GOSTs, and SNIPs.
When an engineer needs to answer a request like "which valves on section B-2 have an expired service interval and whether they overlap with scheduled repairs of the heating system," they have three options for solving the problem:
The third option works under one condition:the agent must have structural access to layers, attribute schemes, BIM models, and documents, not through a proxy API that must be manually described in the model, but through the standardized MCP protocol.
This is especially important in the case of cGIS Pro. The platform was originally designed as a single web application for all types of spatial data: 2D maps, 3D models, BIM, point clouds, orthophotos, video streams, and IoT sensors.By the time the AI agent was introduced, the architecture already contained all the necessary context.The AI was not added on top; it was built on a pre-existing structured stack.
The agent sees the platform through the eyes of an experienced user and determines the order in which to perform actions. When asked to "show everything on this site," it sequentially accesses the project, the list of layers, and the 3D models. When asked to "give me all the valves with an inter-service interval of more than five years," it accesses the attribute data rather than the geometry. When asked to "find the regulations for security zones," it switches to searching through corporate documents.
These transitions are not performed by trial and error.The cGIS Pro server tells the agent the rules at the time of connection:which scenario applies to working with layers, which applies to searching for documents, and which applies to BIM models.This ensures repeatability and predictability.
The agent covers five typical scenarios based on a set of tasks.
"Show all expired technological connection objects."The agent finds the required layer, filters the data by attributes, and displays the result on the map without copying coordinates or manually exporting.
"How many of them are at the connection stage, and what is the total capacity? "Theagent builds aggregates based on the same data and returns a table ready for export.
"Which of them intersect with the 110 kV security zones?" Theagent matches the two layers by geometry and returns a list of conflicting objects with highlighting on the map.
"Show the model of workshop No. 3 in section B-2. "The agent finds the required model in the registry, opens it in the viewer, and proceeds to the requested element or comments in BCF format. cGIS Pro supports the international IFC standard and its own optimized format for fast web viewing.
"What does our regulation say about work in the heating system area?" Theagent searches for meaning in the documents that you have uploaded: internal regulations, GOSTs, SNiPs, and design documentation. It returnsa fragment with a quote and a link to the source.
These five scenarios work within the same dialogue. The contractor asks "where will the work be done," "what is in the work area," and "find the appropriate regulations," and the agent selects the correct tool for each step.
In total, the platform provides two AI modes that are rarely seen together:
1. Data Analytics - operations on your GIS and BIM data: "show", "calculate", "draw", "compare".
2. Navigating Corporate Knowledge - Semantic search of your documents, returning quotes and links to the source.
The combination of these two modes in a single agent, which works on top of the open MCP protocol, does not have any public analogues in the open releases of the world's leading spatial data vendors at the time of this document's preparation.
To be more specific, I will tell you about a specific scenario that we are implementing for one of the major Russian manufacturing companies. This is a typical business process that any industrial enterprise with an extensive engineering infrastructure faces.
Situation..
The contractor needs to lay a new pipe from workshop No. 3 to the pump station. To obtain permission for excavation work, the following is required:
1. Specify the coordinates and parameters of the work.
2. Obtain an extract of all existing engineering communications in the work area (electricity supply, water supply, heating system, communications, and gas).
3. Notify the owners of each communication.
4. Get approvals.
5. Obtain a permit and a map showing the work area and existing networks.
In the traditional scheme, the process takes from three business days to two weeks. The contractor contacts the dispatchers, the chief engineer collects data from several systems, the draftsman prepares the drawing, and the documents are approved and signed. Each step depends on the availability of the specific employee.
This is a live project. A local LLM is deployed on the customer's side (Qwen and DeepSeek on their own GPU infrastructure): data about the industrial site and regulations do not leave the enterprise perimeter. cGIS Pro provides an MCP server, platform operations, and search for corporate documents.
An important point. The agent does not require the engineer to learn new interfaces, and it does not oblige the contractor to understand GIS. The interaction is conducted in natural language by the user. The entry threshold is reduced to the level of regular business correspondence.
The second typical scenario is the operational work of a unit responsible for the uninterrupted operation of infrastructure. In this role, the AI agent acts as a "co-pilot" for the on-duty shift and simultaneously as a repository of experience for predictive analytics.
Situation.
An industrial enterprise has an extensive engineering infrastructure: substations, pump groups, distribution cabinets, and technological lines — a total of hundreds and thousands of pieces of equipment. Some of them are equipped with IoT sensors. The on-duty shift must respond to deviations from the normal parameters, taking into account the following:
1. The location of the event is on the industrial site or inside a specific building room (BIM model).
2. The history of equipment maintenance and repairs (maintenance and repair system).
3. Data from video surveillance systems and personnel and vehicle monitoring.
In the traditional scheme, response is based on the heuristic knowledge of specific employees. This model does not scale well in a large and complex territory and is vulnerable in the absence of key specialists. Predictive analytics are typically conducted in separate systems without reference to the spatial context of objects.
This is a live project. A local LLM (Qwen and DeepSeek on the customer's own GPU infrastructure) is deployed on the customer's side, as data about the production site and operational regulations should not leave the enterprise perimeter. cGIS Pro provides an MCP server, platform operations, search for corporate documents, and integration with external systems such as maintenance, video surveillance, and personnel monitoring.
An important point. The value of this scenario is not only in the speed of response. The main thing is that each resolved incident adds to the corporate experience base. After a few months of operation, the agent begins to offer the dispatcher more accurate hypotheses about the causes of failures and better recommendations for regulations based on the accumulated statistics. The combination of "response + predictive analytics based on spatial context" is a significant difference from classic SCADA and automated process control systems, which only work in the moment without reference to the 3D model of the workshop and without integration with operational regulations.
When evaluating an AI solution with access to corporate data, the CIO has three key questions:
1. Who has access to the data?
2. What actions can the agent perform, and what are the limits of his authority?
3. What evidence of compliance with regulatory requirements and internal policies can we provide?
Below is a sequential answer to each one.
The decision is made by the customer. cGIS Pro supports three options for deploying LLM. The agent works the same in all three configurations. Changing the model does not affect the business logic or application code.
Data does not leave the perimeter. Qwen, DeepSeek, Llama, Mistral, Saiga, and similar products are supported. Applicable for critical information infrastructure under 187-FZ, the banking industry, the military-industrial complex, and the oil and gas industry with regulatory requirements.
GigaChat (Sber), YandexGPT (Yandex). Data remains in the Russian Federation, and the providers are certified. Applicable for most government and municipal scenarios.
Claude, GPT, Gemini, and others. Applicable for commercial tasks where model quality is a priority and there are no restrictions on cross-border data transfer.
All AI operations are subject to the cGIS Pro platform's role model. If a user does not have the right to edit the engineering networks layer, the agent will also not be able to do so, regardless of the query's wording or attempts to bypass it through prompt injection. Read-only roles retain their read-only behavior. AI security aligns with platform security, simplifying auditing and certification. Additionally, there are additional mechanisms in place:
Prompt sanitization is the filtering of incoming requests for common injection patterns.
Action log - every request to the agent is recorded: author, wording, tools used, data returned. This is a requirement for internal and external audits.
Isolation of the corporate document index - each customer works only with its own case. Cross-client access is architecturally excluded.
Deployment in a secure environment, role model, activity log, and local LLM engine comply with the requirements of the Federal Service for Technical and Export Control for systems processing data of CII subjects. Certification procedures are being performed.
At meetings with potential customers, one of the most common questions is what preparatory steps the company needs to take. Below is a practical answer. The minimum requirement for a pilot launch is three components.
You need data for a single object or site. This is not a migration of the entire corporate GIS or an archive from twenty years ago, but a current set of layers for a single industrial site, district, or workshop. We accept vector layers in any common format, such as SHP, GeoJSON, GeoPackage, MapInfo, and DXF, as well as orthophotos, 3D models in IFC format, point clouds, and CAD drawings. Most companies with a functioning GIS or BIM practice already have this data.
Documents that the agent must answer questions about: regulations, instructions, GOSTs and SNiPs, project documentation, corporate standards, contracts, and methodological materials in PDF, DOCX, and XLSX formats. No special preparation is required; documents are uploaded in their original form. The corpus can range from dozens of documents to several thousand, and the quality of semantic search improves with increasing corpus size.
There are three possible configurations: within the customer's corporate environment (virtual machines or a Linux-based container environment, support for Astra Linux and RED OS), in our cloud, or a hybrid setup. If there are 187-FZ requirements or other restrictions on data placement, only the customer's corporate environment is used.
For critical infrastructure, a local model on your own GPU is usually mandatory. The minimum sufficient configuration is one machine with an A100/H100-class GPU for the scale of one enterprise. In the absence of your own GPU infrastructure, a Russian cloud-based LLM service (GigaChat, YandexGPT) is used at the start, with a possible transition to a local model later.
A separate AI team: deployment, training of administrators and product managers, and documentation are provided by our team. Reconfiguration of business processes for the platform: the agent adapts to the customer's existing processes. Additional user training: natural language as an interface does not require training.
The AI agent does not compensate for the low quality of data and the inconsistency of the regulatory framework. Before the pilot launch, our team conducts an express data and regulatory audit with the customer. This takes one to three business days and significantly improves the predictability of the outcome.
A practical tip: for a pilot project, we recommend choosing a scenario with measurable business metrics. For example, reducing the time it takes to approve standard permits, reducing the workload on the dispatch service, and transitioning contractors to self-service based on standard invoices. This allows you to make decisions based on actual numbers rather than general impressions.
We offer three levels of engagement.
Demo. The platform with the AI agent enabled is ready for demonstration. During the demo, we run typical scenarios on our reference data, answer technical questions, and discuss the applicability to your context. The format is remote or in your office, and it lasts about an hour.
Pilot project. Deployment of cGIS Pro with an AI agent within the customer's perimeter at a single site, such as an industrial site, workshop, or district. Uploading spatial data and corporate documents, customizing the agent for your scenarios, and training administrators. The duration is 4-8 weeks, depending on the amount of data and the complexity of the scenarios. The scope of work and commercial terms are discussed individually based on a short discovery call.
Full-fledged implementation. The cGIS Pro platform is available in three versions: on-premise, SaaS, or hybrid. The AI agent is an extension of the platform and is charged based on the company's scale and usage intensity. Specific parameters are discussed after reviewing the customer's requirements.
The cGIS Pro implementation team is compact, consisting of 1-2 people (a product manager and a technical architect). We do not build a multi-level structure with lengthy workshops. If the customer has a specific task and prepared data, we focus on the result.
Svenska 
Русский 
English