In the first posts in this series, I proposed a broad, practical, position on how we can approach AI in the machinery compliance and health and safety space: AI is neither a gimmick nor a silver bullet.
In other words, AI isn’t something that should fill us with terror, nor is it something to adopt unthinkingly, ignoring its limitations. Instead, we should recognise AI for what it can be when used in a disciplined, well-scoped way: a tool that can materially improve efficiency, consistency, and ultimately safety.
In the second post in this series, I explored how AI can help to support the creation of internally consistent risk assessments by sifting through an organisation’s historic library of hazard descriptions, control measures, and evaluation notes. The goal in that example was not to let AI think on our behalf, but to use it to preserve institutional knowledge, cut out repetitive rewriting, and strengthen consistency across assessments.
In this post, I want to build on these previous posts by looking at another practical, high-value application of AI:
AI-powered natural-language querying of health and safety datasets.
Data Matters, but Querying it is Difficult
One of the features of the health and safety world is that we accumulate large amounts of data. Risk assessments and evaluations, maintenance records, near-miss and accident reports, inspection notes, audit results, control measures, timelines, and more all form a growing reservoir of institutional knowledge.
Though it can be overwhelming, if managed well, this volume of information is not a problem, instead it is a strength. When properly used, this data:
- helps to track recurring hazards,
- can highlight emerging problems,
- help to identify which of our controls are effective,
- raises urgent issues requiring swift intervention, and
- helps us to remember the lessons our organisation has learnt in the past (sometimes at great monetary or human cost).
However, data only becomes valuable when we can query it. That is, it is only useful when we can ask it meaningful questions and get meaningful answers back.
That is the sticking point. Most of us know what kinds of questions we want to ask our data:
- Which hazards on Site X have the highest risk ratings?
- What control measures have historically been used for guarding failures on our machines?
- Which assessments from the past 12 months identified urgent corrective actions that are still outstanding?
- Where have we seen repeat near-misses involving conveyor systems?
These are important questions, with real-world implications. But getting answers from a traditional database is not straightforward unless the person issuing the query is comfortable with SQL or similar. And in most organisations, those who need the answers most frequently are not specialist data engineers. They’re health and safety managers, site leads, engineers, and assessors..
The Limits of Traditional Querying Tools
SQL databases are extremely powerful for storing and linking data, but they require precise, structured queries. Unless you’re trained in SQL, or have the time to become fluent, your querying options are generally limited to:
- exact word matches,
- strict filtering rules,
- shallow search criteria.
In traditional querying languages, there is little room for nuance or semantic understanding. A human can see that “manual handling incident” and “lifting strain” refer to related hazards, but a SQL query won’t. A human knows that “conveyor belt entanglement” and “caught-in conveyor” might be connected, but a database usually treats them as entirely separate.
Many of us try to work around these limitations using familiar tools:
Excel
Powerful and flexible, but again, only as useful as your ability to structure and manipulate data. Filters and pivot tables help—but they still demand technical skill, and they don’t surface subtle semantic patterns.
PDFs
Excellent for data stability and archiving, but inconvenient for querying. Information in a PDF essentially becomes a fixed artefact. It will remain readable and stable, but not easily searchable in a meaningful way.
The details are ‘in there somewhere’ but the larger the document, the harder it is to find exactly what you’re looking for and the easier it becomes to miss something crucial.
These tools are all valuable in one way or another, but other solutions are needed to help us query our data more effectively.
AI as a Natural-Language Data Interface
This is where AI’s natural-language querying ability becomes transformative. It isn’t transformative because it replaces our judgement, there is ultimately no replacement for rightly formed human judgement, but AI can help us more naturally ask the questions we already know how to ask as humans.
Instead of writing:
SELECT TOP 1000 *
FROM RiskEvaluations
WHERE MachineType = ‘Milling Machine’
AND ResidualRisk >= 4
ORDER BY ResidualRisk DESC
(Which would be a very simple SQL query) we can simply ask a properly constrained AI:
“which of our milling machines have generated the highest-rated risk assessments with urgent timeframes?”
And the system interprets it correctly.
AI can parse your natural language query and compare it semantically with the underlying data. It can treat “entanglement hazard,” “caught in moving parts,” and “rotating component entrapment” as overlapping concepts—even if they are recorded differently in the dataset. It can understand synonyms, rephrasings, and contextual clues. It can use fuzzy matching rather than brittle letter-for-letter comparison.
And, crucially, it can do this very quickly.
This is not replacing SQL, properly defined data structures remain the foundation of getting anything meaningful out of such a system, but AI used in this way can seriously augment it. The database still stores precise, structured data. But AI sits on top of it as an interpreter, allowing humans to query the information with the same ease that they talk about it.
Why This Matters for Machinery Compliance
The real power of natural-language querying lies in what it enables:
1. Faster Answers to Real Operational Questions
Instead of spending hours filtering spreadsheets or running multiple database queries, a manager can get answers in seconds or minutes. That reclaimed time goes back into real safety work, not into more clerical work (unless we lack the imagination to think of any more worthwhile use of our time than clerical work).
2. Better Use of Institutional Knowledge
Much like the AI-driven risk-assessment lookup tool discussed in the last post, this approach preserves and exposes patterns in historical data that may otherwise be hidden. It helps to connect the dots that humans are looking to connect in the real world where strict querying languages would have missed the presence of the dots entirely.
3. Reduced Dependence on Technical Specialists
Teams don’t need to become expert programmers just to access information. They simply articulate the question they naturally think of. This isn’t a threat to technical specialists, it simply frees them to apply their technical expertise to larger and more impactful projects, rather than acting as a translator for basic, everyday, Health and Safety questions.
4. Maximising the Value of the Data We Already Collect
Health and safety teams invest considerable time in collecting data. That data is only useful if we can ask it the questions we need answers to. AI can help to ensure that data-collection efforts pay off.
Guardrails Still Matter
As in previous posts, it’s worth reiterating that safeguards are essential. Natural-language querying must:
- be restricted to internal, quality-controlled datasets,
- avoid fabricating answers,
- return “no result found” when appropriate, and
- be used alongside human judgement, not instead of it.
But with these constraints, AI becomes an amplifier of human effort and judgement rather than a loose, ungoverned hazard.
AI Helps to Unlock the Data We Already Have
We spend a lot of time and effort capturing information, but the real power of data lies not only in storing it, but in being able to query it, understand it, and act on it.
AI’s natural-language querying ability brings that power within reach of health and safety experts without them having to become data science experts as well. It strengthens the connection between data and action, between record-keeping and real-world outcomes.
And, as with the previous AI applications explored in this series, the goal is simple: that more people go home safe at the end of the day, and that our businesses can be more productive.
