What Generative AI practically looks like in a safety professional's work

Do you get overwhelmed when you think of AI in Health & Safety? Are you worried about AI? Do you hear AI and suddenly think ChatGPT? AI in safety is broader than that. We have computer vision for things like hazard detection on construction sites or in manufacturing, predictive analytics flagging risk patterns, and even wearables to monitor health metrics like sleep, steps, fatigue etc. Then we have generative AI - the kind of AI that reads, writes, generates, and assists like ChatGPT, the now famous Claude, etc.

My focus in the writeup is Generative AI - specifically Copilot (the one by Microsoft).

My thinking is that if an organisation is going to adopt this type of AI, they would want one that is part of their tools ecosystem which would be Copilot (for Microsoft users) and Gemini (for Google users). It is a no-brainer. I have both but somehow use Copilot more.

Can I state here…

It is how I have been able to use Microsoft and Google tools for safety work - where many are waiting for, or relying on software that just isn’t fit for purpose. The idea is to understand the safety workflow and the inner workings of the tools, as well as its limitations, to make it work for you. These are a great alternative where money for software is not available.

Yes, these tools were never built with what we do in H&S in mind, but we can make use of them simply by thinking outside the box (of what digital tools for safety should look like), and just trying it out - sometimes it works, sometimes it doesn’t.

In this piece, I will share how I have used Copilot for risk assessments, data analysis, and building automated workflows .The other AI tools are able to do exactly what I am about to share so please feel free to practice and start with what you have.

I have organised them into three categories based on how most safety professionals will find AI useful in practice. I call them my top 3.

I created a risk assessment agent in Copilot, and trained it on the five steps of risk assessment by the HSE. It identifies the hazards based on information I upload, as well as a risk rating. I then review it and where necessary, can make it more site-specific.

In this example below, I uploaded a metal master manual and it created the risk assessment, risk rating before and after controls, a risk matrix and a safety checklist. I, of course, need to review the content and cannot accept everything it dishes back as correct – that is where the Human in the loop comes in.

You can design the agent the way you work and best suits your processes. I created this simply to test what it is capable of.

Ever found Excel frustrating, especially those functions and formulas?

Well now, you don’t need to know it all. You can state what you need in plain English and watch AI go to work.

See the first image below. I opened up my sample incident data in Excel and asked Copilot to create a pie chart of the injury type. It did – the injury type column is clearly labelled so no guessing for Copilot. Then it suggested it could add a severity breakdown. I went ahead with it to see what it would do. The result? A pie chart with an inner and outer ring. It said the inner ring was “severity” data – I checked, it wasn’t. So clearly, I wouldn’t be accepting nor using the 2nd pie chart as it is kinda meaningless for what I am trying to achieve. You don’t have to be a data pro to be able to tell – look closely at what the data and chart tells you.

I could continue to tweak and ask Copilot to perform various tasks until I get the metrics and charts I need. Google’s Gemini works the same way - give it instructions and watch it work. Claiude and ChatGPT can also help with data analysis and creating these charts.

One major frustration I faced when I worked in H&S was missing important information or finding out too late. Sometimes “too late”means having to explain to enforcement why you submitted that RIDDOR late or to senior leadership why a HiPo or serious accident was not investigated on time and learnings shared in a timely manner,

So I designed a workflow on Power Automate to send a Teams message and email when certain conditions are met.

For example, if a fracture or hospital visit is selected on a report, we needed to be notified immediately. The message would include certain information about the report. It could also be set for certain information to automatically get added to a Sharepoint list, or a folder is created and images/files are saved automatically to this folder. These are some of the things I have tested on Power Automate really round some of my desires and stuff I have experienced at work.

Err what is Power Automate? It is basically Microsoft’s automation tool. And with it comes expressions and sometimes, may require some low-code information entered to help build your flow. This is where Copilot has been helpful – you describe what you are trying to do and it suggests a code or solution you can accept or reject. Copilot also helps you build flows from scratch by describing the flow you need and Copilot then recommends which tools to which and in which order.

Power Automate is a game changer for those small internal processes that do not require a full software.

Do I encounter challenges while doing all the above and more? Hell yeah! I have a solutions mindset and always strive to learn and apply these learnings. So most weekends I explore and experiment with AI to keep my knowledge current. And everything I have shared in this piece - the risk assessment agent, data analysis, Power Automate is part of what I cover in my live workshop, Build Your First Digital Safety System in a Day.

In 3-4 hours, I take a small group through data collection, analysing that data, and building automation. It is hands-on and practical. Maximum 12 people per session.

If that sounds like something you need, DM me and I’ll send you the details