In a previous post, I suggested a basic position on the role of AI in the world of Health and Safety. AI brings real strengths, such as pattern recognition at speed and scale, while also carrying some important limitations, especially when applied to environments where mistakes can lead to serious injury or worse. The stakes around dangerous machinery can be high.
Yet it is equally true that, when used wisely, AI tools can improve efficiency in material ways. And in the context of our industry, efficiency is not simply about cutting paperwork. It’s about helping people go home safer and enabling our businesses to be more productive over for the long-haul.
In this post, I want to focus on a particular application of AI: using AI to build, maintain and query an internal database of risk evaluations, hazard descriptions, and control measures.
The use of AI in this scenario is not about replacing human judgement.
Instead, the aim is to support our businesses by improving consistency, reducing repetitive rewriting, and helping teams to carry institutional knowledge forward even as personnel and processes change.
Consistency in Large, Changing Workplaces
Consider a large factory with dozens or even hundreds of machines. Over the past 10 or 15 years, various people will have conducted machinery compliance assessments and health and safety evaluations on those machines. They will have used similar (but not identical) templates. Similar hazards have been described many times. They’ve assessed similar risks again and again. But the results are surprisingly inconsistent:
- Different assessors phrase things differently.
- A person assessing Machine A on Monday may word things differently when assessing Machine B on Friday.
- Over time, corporate memory fades. Things written ten years ago become harder to find or interpret.
- Attempts to enforce consistency can lead to bulky documents that no one has time to reference.
- Without such attempts, assessments drift apart and internal coherence can be lost.
And what is the result? :
- People waste time rewriting content they’ve already written dozens of times, trying to recreate what they think they wrote last time.
- Consistency evaporates, because it is too time-consuming to manually maintain it.
- Institutional knowledge is lost.
Both outcomes have real implications for safety. Inconsistent or unclear hazard descriptions can lead to uneven control measures. Important lessons learned from older assessments can be forgotten. And the cumulative inefficiency drains time that could be spent addressing real, on-the-ground safety concerns.
A Targeted AI Solution: Building and Querying a Risk Evaluation Database
This is where a carefully designed AI-supported system can meaningfully improve both safety and workflow efficiency.
The core idea is simple: Use AI to rapidly search, match, and re-express your own historical risk information, while keeping strict constraints to prevent AI from inventing or speculating.
Step 1: Build a Structured Internal Database
The first step would be to gather your existing assessments into a structured format. This might include:
- Hazard descriptions
- Risk evaluations
- Suggested control measures
- Relevant ratings etc. from your organisation’s risk matrix
These documents already exist; they simply need to be collated into a form the AI tool can query.
Quality assurance remains entirely in human hands: the organisation decides which documents are “authoritative”.
Step 2: Calibrate the AI to Perform Similarity Lookups
Once the data is captured, an AI tool can be calibrated with a very narrow task:
Given new input from an assessor, search the internal database for risk evaluations most similar to the new scenario and summarise or re-express them using the organisation’s established terminology.
Imagine an assessor walking the factory floor and identifying a hazard that requires evaluation. Instead of starting from scratch or trying to remember how a colleague phrased something two years ago, they input a brief description into the tool.
The AI then:
- Searches all existing assessments using fuzzy logic, semantic matching, and pattern-recognition techniques.
- Finds the closest matches from years of accumulated organisational knowledge.
- Returns hazard descriptions and control measures that reflect the wording, risk categorisation, and reasoning already used elsewhere.
- Adapts the phrasing only where necessary to align with the details of the new situation.
The assessor remains in control. They review the output, adjust as needed, and ensure the final evaluation reflects the real situation. But the time-consuming work of searching through old documents or reinventing standard phrasing is eliminated.
Step 3: Enforce Strict Guardrails Against AI Guessing
Many concerns about AI in high-risk domains relate to “hallucination”, the generation of invented facts. That risk can be drastically reduced by implementing strict prompting rules, such as:
- The AI may only draw from the internal database.
- If no good match exists, the AI must explicitly return: “No good match found.”
- The AI must not guess or generate new hazards, evaluations, or control measures.
- The AI output must be formatted into a clearly structured set of fields and not free prose.
In short, the AI behaves not as a creative assistant but as a librarian: finding relevant historical material, organising it, and presenting it consistently.
The Benefits: Efficiency, Consistency, and Safety
1. Dramatically Reduced Time Spent on Rewriting
Risk assessors spend an enormous amount of time rephrasing content they already produced in previous assessments. AI-assisted lookup almost entirely removes this repetitive labour. Assessors can spend their time on thinking, not on typing.
2. Stronger Internal Consistency Across All Assessments
Because the AI uses the organisation’s own historic wording as its primary reference, the language of risk becomes more stable. Terms are used consistently. Similar hazards receive similar descriptions. Control measures are framed uniformly. Over years, this builds a coherent internal “safety dialect” that becomes part of the organisational culture.
3. Reduced Loss of Institutional Knowledge
When experienced assessors retire or move on, their way of describing hazards often goes with them. An AI-supported database preserves their phrasing and reasoning for future teams.
This is especially valuable in organisations with high turnover.
4. Improved Quality of Assessments Through Better Memory
If a rare but serious risk was identified ten years ago on one machine, similar risks found today can be flagged instantly when the tool matches the pattern. This strengthens hazard recognition across the entire site.
5. Better Safety Outcomes Through Efficiency
Faster, clearer, more consistent assessments mean more time for implementing control measures, training operators, and addressing real-world hazards. Efficiency in this context is both productive and protective.
Addressing AI Limitations Through Good Design
The effectiveness of this approach depends on respecting the limitations of AI and designing the system accordingly.
- Narrow, well-defined tasks avoid overreach. The AI is not making safety decisions. It is retrieving relevant past assessments.
- Data limited dataset removes uncertainty. The system is trained solely on internal, vetted material. You know what’s in the dataset.
- Strict rules about generation prevent hallucination. The AI cannot “improvise” hazards or controls.
- Human reviewers remain responsible. AI supports, but humans decide.
When used with these constraints, AI can become a powerful tool for consistency and speed.
Conclusion: A Practical, Safe Path for AI in Machinery Compliance
AI does not need to be sweeping or transformative to be valuable. Small, targeted applications can yield significant benefits without overstepping the boundaries of safety or common sense.
By treating AI as a structured lookup tool rather than a creative problem-solver, organisations can:
- Preserve institutional knowledge
- Reduce inefficiency
- Improve consistency
- Strengthen assessments
- And ultimately help ensure that everyone goes home safe at the end of the day
This is the kind of practical, grounded AI adoption that respects both the opportunities and the limits of the technology.
