Artificial intelligence in mineral resource governance: A framework for consistency, traceability and technical assurance of resource estimates

Mineral resources are core assets of a mining company, and their governance requires a high degree of assurance. Audit processes are key activities in developing this assurance. The audit of mineral resources is a detailed and often tedious process that connects the estimation, evaluation, documentation, oversight and public disclosure of a mining company’s principal mineral assets. Such audits often require reviewers to reconstruct workflows from native project files, parameter sets, batch macros, or scripts and incomplete reports. This article presents a human-directed approach in which specialized artificial-intelligence (AI) agents operate incumbent geological, geostatistical and mine-planning software, collect evidence and produce a reproducible audit record. Four anonymized applications illustrate the method: reconstruction of a legacy estimate during transaction due diligence; parameter review across 46 estimation domains; forensic recovery of settings from batch journal files; and reconciliation of medium- and short-term block models over common stope volumes. In one transaction, a reconstruction estimated to require about five weeks by conventional review was completed in two days. The more consequential result was evidence traceability, not speed alone. A three-level evidence-sufficiency framework classified audit questions as answerable, partially answerable or not answerable, and identified the additional information needed for unresolved items. The cases show that agentic systems can improve coverage and repeatability without replacing professional judgment. Competent Persons (CPs) and Qualified Persons (QPs) remain responsible for materiality, classification, estimation choices and public-reporting sign-off. The MineMind system architecture. Introduction Mineral resource governance for a project is a management framework that integrates the estimation process, the resulting estimates, documentation, oversight and public disclosure of a mining company’s principal mineral assets. An efficient framework makes the resource estimation process auditable and transparent while keeping it compliant with the applicable internal technical standards and regulatory requirements of the operating jurisdiction. The mineral resource estimation (MRE) process in any mining project evolves through many cycles of data collection over time, geological interpretation, definition of estimation domains (domaining), exploratory data analysis (EDA), geostatistical modeling (variography), grade estimation, resource classification and validation of the estimates against the input data to confirm their quality. The technical data and records are commonly distributed across various databases and across resource modeling projects conducted over time. Each resource modeling project has its own data files, estimation parameter files, software-specific commands and workflow scripts for data manipulation and repetitive processes (macros or scripts) and reports. When a past project changes ownership or personnel, many details are lost because they were held only by the person or team that owned the project. Reviewers trying to understand the process and results of such a project must determine whether the reported estimate appears reasonable and whether the executed workflow agrees with its documentation (scope of work, standard operating procedures or a technical report). Reviews depend on experienced professionals navigating multiple software interfaces and recording their findings manually. Repetitive checks across many domains can be slow and vulnerable to omission or fatigue. There is an opportunity for an AI-assisted system to perform these tasks systematically, accelerate information gathering and evidence retrieval, perform consistency checks and retain provenance while preparing reports for judgment by the CP or QP. This article describes an AI-assisted architecture and evaluates its practical application through four anonymized engagements. The architecture The approach complements established technical software through an integrated orchestration layer. This layer uses AI agents to translate plain-language instructions into coordinated workflows across existing technical tools, reducing manual handoffs between disciplines. MineMind (https://www.minemindai.com/#plataforma) converts user-defined methods into reusable workflows called skills. These skills can integrate different data sets and software tools, execute workflows repeatedly, maintain version histories, and document each step without relying on fixed software menus or difficult-to-maintain scripts. As shown in Fig. 1, a user defines what is needed as a bounded technical question. Specialized agents plan the required checks, call geological, geostatistical, database or mine-planning tools through available programming interfaces, and return results with references to the source files, parameters and executed steps attached. The generated report is reviewed by an expert geoscientist or engineer for validation and acceptance. Evidence-sufficiency control Each review project retains at least the project-file identity and version, the query or instruction, the software operation, the parameters extracted or applied, intermediate validation results, exceptions and the reviewer disposition, which together provide a complete, auditable and evidence-rich package. Each result has to be linked to a data source and a process. A result that cannot be linked to its source is treated as an unsupported assertion.Deterministic computations, such as drill-hole interval validation, block-model aggregation or domain assignment, are traceable and reproducible. When a user asks the MineMind system a question, the outcome falls into one of three categories. Accordingly, MineMind classifies questions in three ways: An answerable question has direct, sufficient evidence for a bounded conclusion. A partially answerable question has relevant evidence but requires explicit qualifications. A question is not answerable when the available record cannot support the requested conclusion. In the last two categories, the final results name the missing evidence or analysis. This is important because the explicit listing of missing items allows the user to go back, address the gaps and repeat the process. Case studies Four anonymized case studies, selected to represent different settings, are presented here. The case studies are observational engagements rather than controlled experiments. Case A: Transaction due diligence. A copper-gold asset in Latin America was in an active sale process, and a third-party review was warranted. Because of legacy data issues, the review was taking much longer than planned. MineMind AI agents were deployed to complete the due-diligence process. Starting from simple plain-language questions, workflows were created to synchronize the data analyses, results and processes, and the original estimation processes were reconstructed. The numbers shown in the technical report were validated using several skills within MineMind. The reconstruction and validation were completed in two days, compared with an indicative conventional-review quotation of approximately five weeks. The stronger claim is not speed alone: each recovered parameter was linked to a project file and journal entry. Fast data processing combined with evidence-based reporting resulted in an efficient review with reliable, traceable results. Figure 2: A structured process for mineral resource model review. Case B: Review of a complex resource model for assurance. A resource model with complex modeling parameters was examined for consistency and assurance. The objective was to audit the configuration, internal consistency and traceability of the published model, not to generate or endorse an alternative mineral resource estimate. The model had 46 estimation domains, each with its own combination of estimation parameters and geological controls. It was an extremely complex review process. The AI agents in MineMind systematically extracted, standardized and compared all parameters applied within every domain, such as variogram models, search-neighborhood geometry and orientation, minimum and maximum sample counts and per-drillhole sample constraints (Fig. 2). Workflows were created for this detailed domain-by-domain comparison to check for inconsistencies, unexplained parameter changes, potential transcription errors, and departures from the documented estimation strategy. Once a workflow for extracting information and comparing it with the report was constructed for one domain and validated on another, the process could be repeated across all 46 domains, reducing the risk of oversight and maintaining a consistent level of review rigor throughout a repetitive process. Case C: Reconstruction of models from internal scripts. In this case, a resource model for a gold-silver deposit was presented with computer-generated batch macros or scripts as the main source of information linking the data sets and parameters. From these files, the audit recovered all modeling parameters, including search ellipses, sector and octant settings, nested expanded searches, sample limits, gold and silver variogram structures, multiple-indicator-kriging cutoffs and post-processing steps. It also found evidence that a stored orientation variable locally rotated search neighborhoods. That observation was reported as indicative of dynamic anisotropy, not as formal confirmation, because the original method documentation was unavailable. The models could be reconstructed from the derived parameters (Fig. 3). Figure 3: Model reconstruction from the parameters derived from the batch scripts and macros. The MineMind system can automate the process so that further sensitivity analyses can be conducted. The record could not quantify the resource effect of analytical bias, establish that a representative population free of high-grade bias had been analyzed, or reconstruct missing sample and assay-certificate status. The next evidence required included sensitivity reruns, declustering and capping studies, and access to sampling and certificate databases. A key audit principle is to report evidence limitations as clearly as the findings. Of seven audit questions, two were classified as answerable, two as partially answerable and three as not answerable (Fig. 4). Figure 4: Three-level evidence classification applied to seven audit questions (Case C). Case D: Block-model reconciliation. At an underground gold operation, medium-term and short-term planning block models were compared over common stope volumes. Agents extracted block populations; calculated tonnage, average grade and contained metal by geological domain and classification; quantified absolute and percentage differences; and produced tables, swath plots and domain statistics. Approximately 8,000 blocks in one model were compared with approximately 5,000 in the other. Density and grade-field consistency were checked during extraction. The application extends the approach from one-time due diligence to a recurring assurance process (Fig. 5). Summary of case studies. The four anonymized case studies discussed above demonstrate how AI-assisted workflows can support different types of mineral resource projects. In the first case, AI agents reconstructed a due-diligence analysis workflow from native project files and scripts and executed the process within two days, while also checking consistency between the executed workflow and the technical report. Case B reviewed resource modeling parameters such as search, variogram and sample-selection parameters uniformly across 46 estimation domains and delivered a consistent review within a few days. Case C reconstructed the forensic audit trail of a gold-silver estimate from batch scripts, reran the models and distinguished questions that could be answered from those that were only partially supported or remained unresolved. Case D compared medium-term and short-term mine planning models over common stope volumes at an underground mining operation, quantifying differences in tonnage, grade and contained metal by geological domain and resource classification. Figure 5: Reconciliation of medium- and short-term block models over common stope volumes (schematic). Implications for compliance: Reporting codes and professional responsibility of the QP/CP It is important to note that the methods discussed support, but do not themselves establish, compliance with CRIRSCO-family reporting codes. For instance, data-integrity checks provide evidence relevant to drilling, sampling and data-verification disclosures, and audits of resource modeling parameters support disclosures on estimation and modeling methods, including domains, variography, search strategy and sample selection. The AI system keeps the reference to the source to support transparency and assists the QP or CP in making judgments on materiality. The system is jurisdiction-neutral and can be used to support compliance with major reporting codes and guidelines such as the JORC Code (https://www.jorc.org/docs/JORC_code_2012.pdf), NI 43-101 (https://mrmr.cim.org/media/1017/national-instrument-43-101.pdf) and S-K 1300 (https://www.ecfr.gov/current/title-17/chapter-II/part-229/subpart-229.1300). While the AI may identify inconsistent parameters, missing records or unexplained differences, it cannot accept professional accountability. Tasks such as geological interpretation, treatment of outliers, estimation method selection, classification, reasonable-prospects considerations and public-reporting sign-off remain the responsibility of the QP or CP. Discussion and conclusions The four case studies show that AI-assisted auditing can improve the speed and consistency of mineral resource reviews and audits. Automated workflows can apply the same checks across every geological domain or mining area in a mine plan, reconstruct technical records from project files and scripts, and compare successive models more efficiently. Importantly, regular model reconciliation can also become a continuous feedback process, which is vital for improving productivity in any mining project. However, there are system limitations that need improvement over time. The case-study results do not provide statistical proof of performance. Legacy files may be incomplete, damaged or unclear, and scripts may show what was executed without explaining why particular decisions were made. An AI system may also misinterpret technical terms or overstate conclusions unless its outputs remain linked to evidence and subject to professional review. Organizations should therefore establish clear requirements for data access, version control, validation, exception thresholds and record retention. A risk-based approach provides a practical way to adopt these methods. Deterministic database and calculation checks can be tested against known results, while extracted parameters can be compared with samples reviewed manually. Interpretive conclusions should require approval by an experienced reviewer, and material findings should receive independent confirmation. This approach uses automation for repeatable evidence-based tasks while preserving professional judgment under applicable reporting codes. The main contribution of this article is an evidence-sufficiency framework that separates audit questions into answerable, partially answerable and not answerable categories and identifies the additional work needed. Combined with reliable calculations, traceable evidence and mandatory professional review, this framework is more important than speed alone. AI should therefore serve as an audit accelerator and assistant, not as a substitute for a CP or QP. (The authors acknowledge the technical teams and clients whose anonymized engagements informed this paper. Client and project identities have been withheld. Quantitative case information is reported with the anonymized descriptors supplied for publication.)

Artificial intelligence in mineral resource governance: A framework for consistency, traceability and technical assurance of resource estimates

Authored by: A.R. Samal and G. Usero

To view the past issues of the UCA magazine you must be a member of SME/UCA.

Lula’s election win dashes US hopes of quick access to Brazil’s rare earths

Critical minerals, power infrastructure and the AI energy surge: Mining’s expanding role

Mining Engineering’s Weekly Rundown: Episode 77

Lula’s election win dashes US hopes of quick access to Brazil’s rare earths

Critical minerals, power infrastructure and the AI energy surge: Mining’s expanding role

Mining Engineering’s Weekly Rundown: Episode 77

To view the past issues of the UCA magazine you must be a member of SME/UCA.

Comprehensive utilization of iron ore tailings: A review of sustainable practices and technologies

Florence Copper and in situ copper recovery for sustainable copper mining

Related Articles

Critical minerals, power infrastructure and the AI energy surge: Mining’s expanding role

Mining Water Tech Pilot Pathway Program: Connecting the industry to startup innovation

Get your SME Foundation VIP Access Conference Pass to MINEXCHANGE 2027

“Check the box” and help support the SME Foundation

Lula’s election win dashes US hopes of quick access to Brazil’s rare earths

Critical minerals, power infrastructure and the AI energy surge: Mining’s expanding role

Mining Engineering’s Weekly Rundown: Episode 77

Lula’s election win dashes US hopes of quick access to Brazil’s rare earths

Critical minerals, power infrastructure and the AI energy surge: Mining’s expanding role

Mining Engineering’s Weekly Rundown: Episode 77

Use Single Sign On