Gains:
- Ability to classify near miss, accident and observation data with AI and extract recurring risk patterns
- Ability to draft risk assessment, JSA and emergency scenarios with AI support
- Ability to verify with OHS expert and legislation, knowing that AI output does not replace safety-critical decisions
Mining is one of the riskiest sectors in terms of occupational health and safety (OHS): collapse, blasting, heavy equipment, gas, dust, heights, electricity and vehicular traffic coexist. Therefore, mine safety is an engineering field that is as important as production, and often comes before it. At the heart of safety management is data: near-miss records, accident reports, field observations, equipment inspections, training records. Most of these records are free text and take a lot of time to read by hand. AI is a powerful aid in classifying this mass of text, extracting recurring risk patterns and drafting risk assessment/JSA. But it should be noted: AI output is not a substitute for safety-critical decisions; It is the responsibility of the OHS specialist and the legislation to stop the work, evacuate the area and determine the root cause.
Types of security data
- Near miss records: Incidents that did not turn into accidents; It is the most valuable early warning source because it is abundant and inexpensive.
- Accident/incident reports: Incidents that occurred, type of injury, root cause.
- Field observations and inspection: Reports of unsafe condition/behavior.
- Risk assessments and JSA: Hazard-risk-precaution analysis for a job (JSA: Job Safety Analysis).
A common principle in OSH is the “safety pyramid”: the large number of near misses and unsafe situations underlie a small number of minor accidents and the rare serious accident. Taking near misses seriously and unraveling their patterns is the most effective way to prevent serious accidents. AI is invaluable in sorting hundreds of free-text near-miss recordings into themes in minutes and answering the question “which condition, which area, which shift repeats the most?”
Step by step: Security data analysis with AI
- Anonymize records. Remove staff names and identification information; OHS data is personal.
- Determine classification scheme. Hazard type, area, shift, equipment, root cause categories. AI: suggests schema, you finalize.
- Auto tag. AI categorizes records; marks the uncertain.
- A pattern emerges. Which theme is recurring, where is it concentrated, does it increase over time? AI: summary and trend.
- Turn it into action. The OSH specialist derives precautions from patterns; AI drafts measures.
- Monitor and verify. The impact of the measure is monitored with subsequent data.
Attention: There is a risk of personal data and accusations in OHS records. The purpose of the analysis is not to punish the individual, but to see systemic risk. Anonymize the records and use the output as "condition improvement" rather than "person search".
Risk assessment and JSA draft
A job safety analysis breaks a job down into steps; It identifies the hazard, risk and precaution at each step. AI can quickly produce a draft from past JSAs and job description; This saves a lot of time compared to writing from scratch. But this draft is a start: site-specific hazards, local conditions and regulatory requirements are added and approved by the OHS expert and the team doing the work. A hazard that AI misses may hide a real risk; That's why the draft never goes directly into implementation.
three mini cases
Case 1 — Repetitive pattern. 800 near-miss records from a furnace are waiting in the archives. The OSH specialist has the AI anonymise and classify them; It turns out that 22% of the records are related to vehicle visibility at a particular ramp intersection during the night shift. This pattern was not visible with individual readings. Lighting and speed regulation are added to the intersection; In the next three months, near misses in that area decrease significantly. AI found the pattern; The precautions and decisions were made by the OHS team.
Case 2 — JSA acceleration. A new maintenance job requires JSA. The OHS specialist gives the work steps to the AI and has it produce a draft JSA; The draft will be ready in 15 minutes. The expert and team review the draft, add and approve two site-specific hazards (nearby live line and confined space). AI saved time; Man has completed the missing dangers. If the draft had been used as is, two critical dangers would have been avoided.
Case 3 — False confidence (warning). An intern asks the AI “is this job safe, can we do it?” AI gives a general "safe with appropriate precautions" answer. The intern thinks this is confirmation. However, AI has neither seen the field nor is competent; Safe-working decision is made only with site evaluation and authorized approval. Lesson: AI is not asked to decide whether it is "safe"; A danger list is made by AI, and the decision is made by the authorized person.
Copiable prompt templates
NEAR MISS CLASSIFICATION "Role: You are an OHS data analyst assistant. Below are anonymised near miss records. Label each record into the following categories: hazard type, area, shift, equipment involved, possible root cause. Flag ambiguous records separately. Then summarize the 5 most recurring patterns and the concentrated condition. Person/culprit SEARCH; focus on systemic conditions. Data: [paste]."
DRAFT JSA"Prepare a draft JSA for the following job: divide the job into logical steps; write down the hazard, risk and recommended precaution for each step. Write at the top that this is a DRAFT and that site-specific hazards and regulatory requirements need to be added and approved by the OHS Expert and the team. Job: [describe]."
RISK ASSESSMENT CHECKLIST "Create a risk assessment checklist for the following activity: [activity]. Scan the possible hazard headings (mechanical, chemical, fall, gas, electrical, traffic, ergonomics, environmental) and write down the control questions that should be asked for each. Do not give the final risk score; state that this requires expert evaluation."
STRUCTURE INCIDENT REPORT"Translate the following free-text incident narrative into a structured incident report: what happened, where, when, who was affected, direct agent, contributing conditions, immediate action taken. ADD information NOT in the text; write 'unspecified' in missing fields. Don't pinpoint the root cause; suggest investigative questions. Text: [paste]."
Weak prompt / Strong prompt
WEAK PROMPT: "Is this safe, should we do it?"
STRONG PROMPT: "Role: You are the OSH assistant. Create a step-by-step list of hazards for the following job and possible precautions for each hazard. Emphasize that this is a DRAFT and that a safe-to-work decision requires site assessment and authorized OHS approval. Also list what additional site-specific hazards need to be expertly controlled. Job: [describe]."
Comparison table: security mission and AI limit
Quest
AI role
whose decision
note
Near miss classification
Powerful browser
OHS specialist
Anonymize
Pattern/trend extraction
Summary
OHS specialist
Systemic focus
JSA draft
quick draft
Expert + team approval
Risk of missing danger
risk score
Not used (decision)
expert
The decision is not up to AI
Work permit/release
Not used
authorized
Security-critical
Common mistakes
- Having AI decide "is it safe?" AI cannot make decisions; He draws up a list of dangers and the authorized person makes the decision.
- Implementing the draft JSA as is. The outline is incomplete without adding site-specific hazards.
- Analyzing personal data without anonymizing it. OSH records are personal.
- Turning analysis into a tool of blame. The goal is systemic improvement, not individual punishment.
- Not taking near misses seriously. It is the cheapest and richest early warning source.
Tip: A good safety culture encourages near-miss reporting. AI quickly turns these notifications into value; But this can only be sustained in an environment where the lessons of analysis translate into healing rather than punishment.
In summary
Mine safety is a data-driven field of engineering, and near misses are the most valuable early warning. AI is powerful in classifying free text security records, pattern extraction, JSA and risk checklist drafting. But safe-working, evacuation and root cause decisions are the responsibility of the OHS specialist and the legislation; AI output is input, not decision. Anonymize records and use analysis for systemic improvement.
Application task
Feed your anonymized (or sample) near miss records to the AI with the “Near miss classification” template and extract the 5 most recurring patterns and note a possible action for each. Then generate a draft for a job with the "JSA draft" template and manually add at least two site-specific hazards. Finally, configure a free text event narrative with the “Event report configuration” template and check for added/missing information.
checklist
- [ ] I anonymized the security recordings before analysis.
- [ ] I had the AI produce a list of hazards/patterns, not decisions.
- [ ] I manually added site-specific hazards to the JSA draft and had it approved by the expert.
- [ ] I used analysis for systemic improvement, not individual blame.
- [ ] I left the safe-work/evacuation decision to the authorized OHS approval.
- [ ] I have verified that there is no AI added/fabricated information in the incident report.