There is a lot of valuable information buried in regulatory reports such as Canada’s National Instrument 43-101 (NI 43-101), the United States’ SEC S-K 1300 and Australia’s JORC Code, but digging it out can be a long and tedious job. However, a new artificial intelligence (AI) platform, MineGPT, is designed to help companies and investors improve their decision making by surfacing the insights they need more quickly.

Francisca Lombard, a trained chemical and metallurgical engineer with a background in project and study management, founded MineGPT in April 2025 after more than a decade working in the mining industry. MineGPT is a large language model (LLM)-agnostic platform designed to transform technical reports into an AI-queryable format, allowing advanced models to focus on the most relevant information within the documents. It then applies structured workflows and mining-specific logic to interpret the data, converting them into comparable metrics like capital costs, grades and resource sizes.

MineGPT was officially launched in November 2025 with the first of its planned four tiers: the Explorer plan, which allows users to browse and download reports directly from the site using filters such as project name, commodity and project phase, or dive into in-depth analysis and review of the report through MineGPT-chat. The other tiers—called Strategist, Analyst and Specialist— that will support different levels of analysis will follow later this year.

CIM: What are the major potential benefits of MineGPT for the mining industry?

Lombard: The most significant enabler is that it allows people to reallocate their efforts towards decision making and technically significant higher-level conversations rather than being bogged down by repetitive, time-consuming tasks such as manual review of full technical reports.

It lets users query the technical reports in natural language, by asking MineGPT-chat complex questions such as, “Show me all feasibility-stage copper projects filed after 2022 with capital costs over US$500 million,” or “What strip ratios are typical for feasibility-stage copper projects in Chile?”

Where we would forget significant information from page 45 by the time we hit page 364, AI is able to identify interconnectivity and relevance between multiple disciplines. For example, it can spot limited reconciliation between a mine plan and processing constraints.

It can’t be used for NI 43-101 qualified person sign-off though; humans still need to be the final authority taking accountability.

CIM: What prompted you to develop MineGPT?

Lombard: I spent more than a decade on the front-end engineering studies portion of the project management cycle. I managed the studies departments, overseeing the whole spectrum of disciplines working together to shape NI 43-101, S-K 1300 and JORC reports.

I was frustrated with the inefficiencies that repetitive tasks place on project budgets and wanted to shift the focus to what really matters.

After several months of trying to push the limits of what can be extracted from NI 43-101 reports using foundational AI models like ChatGPT, I realized that there was a gap in the mining industry when it comes to both AI education and a real high-value, mining-specific AI agentic reasoning agent. Foundational models are generalists and context size limitations restrict the quality of their outputs. MineGPT was purpose-built to maximize the accuracy of responses, reduce hallucinations and provide mining industry contextualized reasoning—since, for example, whether a project is at the exploration phase or a prefeasibility study matters when you ask for the process flow.

During project development, mining, processing and tailings designs are usually developed in isolation. Each of them have reasonable assumptions being made within the context of their silo, but nobody consistently steps back and asks whether those assumptions still hold up when everything is brought together.

That is where AI—especially AI that is able to do “thinking” on so many different levels with so many nodes—can give you the benefit. MineGPT can independently assess the interactivity across silos and identify if these assumptions become high risk when viewed as a full system, which you wouldn’t have been able to see otherwise.

CIM: How long did it take to develop MineGPT, and what were the greatest challenges?

Lombard: It was eight months from deciding to jump in full steam to the release of the first-tier product range. The biggest challenge has been that there are no resources, benchmarks, use cases or applications available to compare or spar with within the mining field. MineGPT is completely first of its kind in every way.

One of the biggest technical challenges was converting highly unstructured NI 43-101, S-K 1300 and JORC reports—including tables, appendices and inconsistent formatting—into structured data that an AI system can reliably reason over. I had to build a lot of the infrastructure myself, since there were no resources available for taking a flat text document and converting it into an artificial “brain.”

Another challenge was enabling cross-disciplinary reasoning, where geological, mining and processing information needed to be interpreted together rather than in isolation. Building something that can “think” for itself—whether it’s providing a process-specific response or commenting on the geological characteristics—took a lot of trial and error, and continues to improve.

CIM: How are you managing AI technology issues such as hallucinations?

Lombard: By far the biggest driver is that the MineGPT system is limited to “thinking” within its “brain.” If you were to ask a foundation model [such as ChatGPT or Google Gemini] a question directly, you are competing with general knowledge. By restricting the queried content to only the most relevant, you reduce the opportunity for the AI agent to get confused. MineGPT’s responses are grounded in retrieved evidence across multiple chunks before synthesis.

The system prompts also explicitly include rules to not generate any information, and to state that information is not available if the question can’t be answered from the technical report content. In addition, each report is pre-processed, with some basic context metadata getting extracted to help aid the MineGPT agent in understanding questions. For example, if you request the process flowsheet from a report that is still at the exploration level, the system will explicitly give you the response that [the flowsheet] does not exist yet.

MineGPT can proudly state that no user prompts are used to train the model. For users in a highly competitive industry such as mining, ensuring that prompts, queries and proprietary analysis are not retained or used in model training is critical. Security was one of the most important design considerations and was prioritized on every level.

CIM: You refer to MineGPT as an agentic intelligence system, not a chatbot. What is the difference?

Lombard: A chatbot retrieves answers; an agent “reasons” through problems. A chatbot is usually tied to simple retrieval-augmented generation, meaning there’s a database of information and all you do is query it. Think of it as getting a “human-like” conversational response for a question you could type into a Google search.

An agentic intelligence system is different in that it has multi-tiered reasoning; it’s called “agentic” since it is effectively multiple AI agents working together to generate a single response. The system first dissects the user prompt into micro-prompts, retrieves the relevant information for each of the micro steps and performs reasoning over the full context. One MineGPT prompt is the equivalent of between 10 and 25 prompts to a regular LLM.

The entire textual content of these reports have also been vectorized into the “brain,” which allows the agents to retrieve the most relevant information for each micro step—versus the first relevant information.

This combined MineGPT functionality enables reasoned responses to complicated questions such as, “If throughput is increased by 15 per cent without plant expansion, which unit operations become bottlenecks, and what secondary risks emerge downstream?” or “Identify environmental and sustainability risks implied by the project design, processing route and infrastructure requirements, even where these risks are not explicitly framed as environmental, social and governance issues.”

CIM: What are the limits or drawbacks of the platform, and how are you addressing them?

Lombard: MineGPT has been designed to maximize transparency, which means sources and page numbers are cited as in-text references for users to look up on their own. The current limitations are that it is a text-focused service and does not analyze images yet.

Simply popping a report into your local LLM, which is sycophantic, will give you a response that seems reasonable but that could contain hallucinations and is based on the first relevant information it found—as opposed to, in MineGPT’s case, the most relevant information.

CIM: What has been the response so far to MineGPT, and what are the plans for the future?

Lombard: My favourite response so far has been “I am not impressed, I am scared,” from a user with more than 30 years of experience [who was] seeing the quality and speed of analysis for a resource estimate report.

The feedback so far has been extremely positive, especially from the consulting and investor streams. But to a large degree, MineGPT has also left a lot of people speechless, especially when I demonstrate the capabilities or advise on prompting.

The upcoming agents will still be agentic under the hood, but will be discipline-specific and enable users to interact with specialist-level reasoning, whether that is to assist in due diligence research or with flowsheet design. The imminent focus is to release the Strategist and Analyst tiers, since both of these will add immense benefit to various stakeholders.

The Strategist tier, to be released in the second quarter, will enable global queries where users gain access to the entire industry’s knowledge from the past decade’s reports condensed into an “AI brain,” while the Analyst tier will be focused on benchmarking data and tools.

Finally, the Specialist tier will enable multi-disciplinary review of reports in a structured review report output, but also will allow you to interact with, for example, a process engineering agent and have it formulate design concepts for you based upon your ore characteristics.