AI-driven analytics in the mining industry

How data analytics can use AI to improve worker health and safety in the mining industry

Mining remains a high-risk industry with significant health and safety challenges for both site personnel and corporate leadership.

In 2024, the Australian mining industry recorded 10 worker fatalities, resulting in a worker fatality rate of 3.4 per 100,000 workers. Criminal WHS prosecutions in 2024 continued to focus on heavy machinery, underground stability, and tailings management.  

That same year, safety regulators finalised severe enforcement actions against several operators, such as the $400,000 fine against Big Bell Gold Operations, specifically citing failure to provide adequate machinery guarding and proper isolation training.

None of this is new, but the cost and consequences of failure continue to rise. Regulatory scrutiny is increasing, and the gap between what operators know happened and the insights they extract from those events remains stubbornly wide. Generative AI can help mining operators do what they have long struggled with: systematically learn from their own history before the next incident occurs. 

Making sense of unstructured safety data

IDC estimates that 90% of today’s enterprise data globally is unstructured. Think incident reports, inspection records, near-miss logs, environmental monitoring documents. Much of it lives in physical archives, scanned PDFs, or legacy systems that don't talk to each other. Some of it may even be handwritten.  

Most mining operators are sitting on decades of data they have been required to collect but have never been able to use.

Near-miss events, in particular, are leading indicators that offer valuable intelligence. However, they tend to be recorded with the least detail, because at the time, nothing appears to have gone wrong.  

In our experience, operators typically hold 20 or more years of records, but only a small fraction exists in a format suitable for analysis. The rest is locked away, and to date, has not been practical to unlock.

The practical pipeline that changes this looks like:

 Across multiple client engagements, we analysed more than 2,000 workplace incident reports over a 15-month period. While the datasets varied in quality and consistency, generative AI revealed patterns and contributing factors that conventional analysis methods would likely have missed. For example:

• Falls from height remained a primary fatal risk, with well-understood control measures.
• Poorly labelled incident data (typically classified as “other”) tended to relate to health and medical episodes.
• Incidents frequently occurred at transition points between tasks (handover, shift change, task interruption) rather than during steady-state work.

Understanding these underlying patterns can help organisations move beyond reactive incident management towards more targeted and proactive risk reduction. It also provides a stronger evidence base for safety interventions, resource allocation and operational decision-making. 

How does generative AI help with data analysis? 

Generative AI works across structured and unstructured data simultaneously, reading a decade of incident narratives and asking qualitative questions that no analyst would have the bandwidth to pursue manually. In a mining sector WHS context, those questions might include:

• What did people notice before things went wrong?  
• What language appears consistently in the reports that preceded serious events?
• How are these issues connected to policies?

The output isn't just a number, and it can iterate. Each pass refines the question.

We have applied this approach in our own forensic analytics work, building structured risk taxonomies that define classification logic upfront, then letting language models apply that logic consistently at scale across large bodies of unstructured investigative material. The architecture is directly transferable to safety data because the problem is the same. You have years of narrative records, and you need to know what those records are telling you.

The technology is just a tool. The value is the speed at which your organisation can move from having records to understanding what those records are telling you. 

Getting the data house in order first

None of the above works if your underlying data is a mess, or if no one oversees it.

The mining industry has long understood the theoretical value of data governance. What it has struggled to do is put that understanding into practice.  

The reasons may be familiar: it's difficult, it asks people to take on responsibilities they weren't hired for, and it competes with operational priorities that feel more urgent.  

Data stewardship, in most organisations, is the thing everyone agrees is important and almost no one owns.

AI has shifted this dynamic in a way that years of governance frameworks couldn't. Operators can now see a direct line between getting their data in order and unlocking something that can be actioned. That's a different conversation than "you should do this because it's best practice."

If your AI strategy doesn't consider data governance, it's not a strategy, it's a demo. Effective governance in this context means:

• knowing what safety data you hold and where it actually lives
• assigning ownership to business leaders who understand what the data means
• defining how records are classified, retained, and accessed
• treating unstructured data, documents, audio, video, with the same rigour as structured databases.

This foundation work takes longer than anyone expects and generates less to show than anyone wants. Do it anyway.

Responsible AI use

In late 2025, the National AI Centre (NAIC) released six essential practices for responsible AI adoption that translate the Safety Standard into something operationally actionable:  

Decide accountability: Assign a single executive owner for AI systems. Shadow AI adoption is already a live risk in most organisations, often a direct symptom of unclear accountability. 
 

Understand impact and plan: Conduct stakeholder impact assessments before deployment. Scoping impact purely through a commercial efficiency lens misses secondary consequences. 

Measure and manage risks: Tier AI applications by risk level and integrate them into your existing enterprise risk framework. 

Share essential information: Maintain a centralised register of all AI models in use and mandate disclosure from third-party vendors embedding AI in their services. 
 

Test and monitor: Move beyond happy-path testing. Like employee performance, AI outputs can drift and testing and monitoring need to be a continuous endeavour. 

Maintain human control: Design meaningful friction into AI-assisted workflows. Circuit breakers, override authority, and explicit procedural permission to challenge outputs. Australia's own Robodebt programme demonstrated what happens when automated decisions execute at speed without intervention, $1.8bn in settlements and 400,000+ individuals impacted.  

 

Incident reports contain personal information. Feeding them into consumer facing AI platforms almost certainly breaches the Australian Privacy Principles. Enterprise deployment, keeping data within your organisation's security boundary, with appropriate de-identification controls, is the minimum viable approach.

The choice of deployment model also matters. Public cloud, private cloud, on-premise, and local workstation LLMs each carry different trade-offs across data sovereignty, performance, and operational cost.