Companies in the metalworking sector and heavy manufacturing increasingly deliver complex infrastructure, energy, and export contracts that require working with thousands of pages of technical documentation, quality standards, material specifications, technical procedures, and contractual clauses. At the same time, requirements around protecting know-how, process documentation, and contract data are rising, as these assets are a core source of competitive advantage. In this environment, using public AI tools becomes a risk that is difficult to accept from a management perspective.
On one hand, the potential of large language models (LLMs) for analyzing technical documentation, RFQs, and historical projects is significant. On the other, cloud-based solutions mean losing full control over data, increasing the risk of intellectual property leakage, and creating misalignment with contractual requirements and internal security policies. A complete ban on AI tools, however, leads to lower efficiency of engineering and bid teams, longer offer preparation cycles, and reduced competitiveness in international markets.
In industrial practice, the lack of secure tools that support knowledge analysis results in engineers, technologists, and bid teams manually processing documentation. Compared to companies operating in less restrictive industries, these processes are slower, more error-prone, and harder to scale. The way out of this deadlock is to build a dedicated, isolated AI environment that enables modern capabilities while preserving full data sovereignty.
Our client is a manufacturer of steel structures and industrial components, delivering projects for construction, energy, and heavy industry. The organization managed an extensive documentation repository covering execution designs, process instructions, material certificates, contract records, and archived offers. Access to this data had to be strictly controlled, and its processing was governed by internal security procedures as well as contractual clauses agreed with key customers.
Previous attempts to use AI tools remained at the conceptual stage due to a lack of confidentiality guarantees in cloud solutions. The company was looking for a way to support technical and bid teams in working with documentation without violating strict information security rules.
The main barrier was security, which prevented the use of public AI tools. Engineers and bid teams spent a significant portion of their time manually searching for relevant clauses in project documentation, standards, and archived contracts. For large tenders and infrastructure contracts, this process became a bottleneck, slowing down response times to RFQs.
An additional issue was the lack of a central, intelligent knowledge search system that could quickly surface specific technical and formal requirements. The organization also feared the risk of imprecise or unverified answers generated by AI models, which in a technical documentation context could lead to serious contractual consequences.
We designed and implemented a Proof of Concept for a local AI assistant based on an on-premises architecture. The first step was to launch dedicated server infrastructure equipped with GPU accelerators, enabling the language model to run without any connection to external networks. Next, we implemented a vector search mechanism (RAG) that fed the model exclusively with documents stored in the client’s internal, secured repository.
As part of the pilot, we built an interface that allowed users to ask natural-language questions against a selected document base. The system was configured so that every answer was supported by a source citation, with precise pointers to the file and the relevant document section. This enabled verification of answer accuracy and eliminated the risk of model hallucinations. The project concluded with AI Governance guidelines defining data access rules and query monitoring principles.
The pilot of the local AI assistant showed that the time required to analyze bid and technical documentation decreased by approximately 45%. Engineering teams gained the ability to search thousands of pages of documents almost instantly, enabling earlier identification of project risks. In user tests, answer accuracy exceeded 85%, while maintaining full control over data.
Over a three-year horizon, the ROI for a full rollout was estimated at approximately 200%. Savings came primarily from reducing labor hours spent on manual document analysis and lowering the risk of contractual errors.
Deploying an on-premises LLM became the foundation of a new knowledge management culture within the manufacturing organization. The company protected its technological know-how and built a solid base for further automation of engineering and bid processes, without compromising data security.
Jakub Smeda
Jakub Smeda
Karol Zielinski
