By Amy Morris
Your AI can sound convincing and still send you digging in the wrong place. Could the tools promising better safety be amplifying the flaws in your data? Before you invest, ask these question about your systems, people and processes.
Artificial Intelligence has become available in nearly every software product we use, but are we ready to use it? The 2025 SaaS Benchmarks Report1 found only 36% of surveyed software companies had designed their products with AI as a core feature. The remaining 64% had simply bolted-on AI as a supporting feature. The longer the software has been around for, the greater the odds its AI features are simply an add-on rather than built-in.
Most organisations now have AI tools available, but availability isn’t readiness. McKinsey’s 2025 Technology Trends Outlook found 78% of organisations are already using AI in at least one business function, yet only 1% describe their AI deployment as fully mature. Using AI to automate a workaround because your existing safety reporting dashboard doesn’t show you what you need is just a bandaid. Using a large language model to look for trends in data your system was never designed to capture can produce confident-sounding conclusions from the wrong evidence. This is not AI readiness. It’s a rush to use AI. Buying shovels before you know where to dig or what you’re looking for.
Before you dig – Know what you’re looking for and why
Know where you need to dig and what you’re digging for before you buy a shovel.
Before committing to any AI solution, first understand what decisions you will make from the output, and what questions you are trying to answer. The question is not “What can AI do for us?”, the real question is “Considering our objective, can AI help achieve it?” Skip this and your AI adoption could undermine safety:
- Identification of incidents and near misses with AI technology increases the volume of reports, it does not reduce the risk. Without a mechanism to intervene when the risk arises, you’ve built a faster way to observe the problem, not fix it.
- AI analysis of safety data can amplify the most frequent issues whilst obscuring the most serious. Without clear parameters and good data to ensure the higher risks are highlighted, they can be drowned out in the noise of everything else.
Such risks arise when AI is layered onto legacy data and systems that were never designed for this. Even as confidence in AI strategy grows, organisations are reporting falling confidence that their infrastructure, data and people are prepared for AI.
As you dig – Human-in-the-loop verifying input, not just output
A prospector who doesn’t know the terrain doesn’t get better results with a better shovel.
Good AI governance ensures there is a human-in-the-loop, and we must always verify AI outputs, but these are not the same thing.
For example, if your safety reporting system follows Safe Work Australia’s Type of Occurrence Classification System (TOOCS)4 to classify injuries (as many do) you may have an over-representation of “Falls from height” incidents. Workers recording a dropped or falling object often misclassify it as a fall, because “Being hit by falling objects” is a sub-category under “Being hit by moving objects”. If your reporting system only offers “Being hit by moving object” and “Falls from height,” which would you choose for a falling object?
AI can’t fix a data problem underneath it, it just digs faster in the wrong spot. The misclassification stays buried, deeper and harder to find, until fixing it later costs more than fixing it now. Verifying outputs alone can create a false sense of security. A study of radiologists5 using an AI diagnostic assistant found when AI gave incorrect answers the accuracy of even the most senior radiologists fell. Trusting the output had quietly replaced verifying it. We all have this automation bias, the tendency to favour suggestions from automated decision-making systems. So when our inputs contain errors, like a misclassification of dropped objects as falls from heights, even the most experienced safety professional could easily trust an inaccurate output.
Are you ready for AI in safety?
Readiness isn’t about how quickly you can start, it’s knowing where to start and what you are looking for. Before you buy another shovel in the AI goldrush, test your readiness.
By Amy Morris, Bega Group
Amy Morris is the Senior Safety Systems and Standards Manager at Bega Group, specialising in safety data, systems and risk management across high-potential hazards and psychosocial risk. She holds a Master of Work Health and Safety (Risk Management) and is the Women in Safety Community Champion. Amy co-hosts Safety Unpacked, a monthly podcast with Women in Safety exploring safety research and cross-industry learnings.
