Big language models (ChatGPT, LlaMa, Mistral, Claude, DeepSeek, QWEN, Grok) are neural networks trained on huge amounts of textual data from the Internet, books, articles and other sources. Their main task is to understand the context of human speech and generate meaningful responses.
Modern LLMis based on the Transformer architecture proposed by Google researchers in 2017(“Transformer (deep learning architecture),”). The key component of this architecture is the attention mechanism, which allows the model to consider relationships between words regardless of their position in the text.
The learning process of LLM is remotely similar to the way humans learn a language, only millions of times larger. The model analyzes billions of examples of words and expressions, identifying patterns in the structure of language and in the logic of semantic transitions. The entire text is divided into tokens – minimal semantic units (words or their parts), which are then transformed into vectors in a multidimensional space (Fig. 8.2-2). These vector representations allow the machine to “understand” the hidden relationships between concepts, rather than simply operating the text as a sequence of symbols.
Big Language Models are not just tools for generating text. They are able to recognize meaning, find connections between concepts, and work with data, even if it is presented in different formats. The main thing is that information should be broken down into understandable models and represented as tokens that the LLM can work with.
The same approach can be applied to construction projects. If we think of a project as a kind of text, where each building, element or construction is a token, we can start to process such information in a similar way. Construction projects can be compared to books that are organized into categories, chapters, and groups of paragraphs consisting of minimal tokens – elements of a construction project (Fig. 3.3-1). By translating data models into a structured format, we can also translate structured data into vector bases (Fig. 8.2-2), which are an ideal source for machine learning and technologies such as LLM.
If a construction project is digitized and its elements are represented as tokens or vectors, it becomes possible to access them not through rigid formal queries, but in natural language. This is where one of the key advantages of LLM comes into play – the ability to understand the meaning of a query and link it to the relevant data.
The engineer no longer has to write SQL -query or Python code to get the required data – he can simply, understanding the LLM and data structure, formulate the task in the usual way: “Find all reinforced concrete structures with concrete class higher than B30 and calculate their total volume“. The model will recognize the meaning of the query, turn it into a machine-readable form, find the data (group and transform) and return the final result.
Documents, tables, project models are converted into vector representations (embedding) and stored in the database. When a user asks a question, the query is also converted into a vector, and the system finds the closest meaningful data. This allows the LLM to rely not only on its trained knowledge, but also on actual corporate data, even if it has already appeared after the model has been trained.
One of the most important advantages of LLM in construction is the ability to generate program code. Instead of passing the technical task to a programmer, specialists can describe the task in natural language, and the model will create the necessary code, which can be used (by copying it from the chat) in the creation of process automation code. LLM -models allow specialists without deep programming knowledge to contribute to the automation and improvement of the company’s business processes.
According to a study conducted by Wakefield Research and sponsored by SAP in 2024 (SAP, “New Research Finds That Nearly Half of Executives Trust AI Over Themselves,” 12 Mar. 2025), which surveyed 300 senior executives at companies with annual revenues of at least $1 billion in the US: 52% of senior executives trust AI to analyze data and provide recommendations for decision making. Another 48% use AI to identify previously unaccounted-for risks, and 47% use AI to suggest alternative plans. In addition, 40% use AI for new product development, budget planning and market research. The study also showed the positive impact of AI on personal life: 39% of respondents reported improved work-life balance, 38% reported improved mental health, and 31% reported lower stress levels.
However, for all their power, LLMs remain a tool that is important to use consciously. Like any technology, they have limitations. One of the most well-known problems is so-called “hallucinations” – cases where the model confidently produces a plausible but factually incorrect answer. Therefore, it is critical to understand how the model works: what data and data models it can interpret without errors, how it interprets queries, and where it gets its information from. It is also worth remembering that the LLM’s knowledge is limited to the date of its training, and without a connection to external data, the model may not take into account current norms, standards, prices, or technologies.
The solution to these problems is to regularly update vector databases, connect to relevant sources, and develop autonomous AI -agents that do not just answer questions, but proactively use data for training, manage tasks, identify risks, offer optimization options, and monitor project performance.
The transition to LLM -interfaces in construction is not just a technological novelty. It’s a paradigm shift, removing barriers between people and data. It’s the ability to work with information as easily as we talk to each other – and still get accurate, verified and actionable results.
Those companies that start using such tools earlier than others will gain a significant competitive advantage. This includes speeding up work, reducing costs, and improving the quality of design decisions due to quick access to data analysis and the ability to quickly find answers to complex questions. But there are also security issues to consider. The use of cloud-based LLM -services can be associated with risks of data leakage. Therefore, organizations are increasingly looking for alternative solutions that allow them to deploy LLM tools in their own infrastructure – locally, with full protection and control over information.
