Think about the last serious accident at your place, or a client's, or one you read about in the trade press.

Now ask yourself: looking back, were there signs it was coming?

Nearly always, there were. A machine that got flagged but never fixed. A crew worn out from shift after shift. A job that had been done the right way ninety times, until the pressure was on and someone took a shortcut. A near miss a few weeks back that got reported, talked about in a meeting, then buried under everything else.

The information was there. It just wasn't being read properly.

That's the real problem with how most places do safety. It's not a dig at the people doing it. It's built into the usual way of working.

For years, safety has been measured by things that count what already went wrong. How many injuries. How many reportable events. These are the final score of a game that's already been played. They tell you how often something went wrong. They don't tell you where the next one is coming, or how close you are to it right now. By the time these numbers move, someone has already been hurt.

The usual response to a serious accident follows the same path every time. Accident happens. Investigation. Find the cause. Raise the actions. Write up the lessons. Update the training. Then, bit by bit, the lesson fades, and the next accident happens somewhere slightly different for much the same reasons.

I've sat in plenty of those reviews. They're done with good intentions and they do change things, but they always look backwards. The aim is always to stop the next one. The method, too often, is to pick apart the last one.

AI turns that around.

Predicting risk isn't magic. It's spotting patterns in the kind of information safety teams already collect. Past accidents. Near misses. Inspection findings. Which jobs got closed and how fast. Permits. Readings from the site. Shift patterns and tiredness. Machine service records.

On their own, these sit in different places and get looked at by different people at different times. The safety manager sees the inspections. The ops team runs the service schedule. HR handles absence and shifts. Nobody joins it all up to see what it says together.

That's what an AI tool can do. Not replace the judgement of the people on the ground, but find the links they don't have time to hunt for. If near misses in one part of the site tend to rise in the week before a big deadline, that's worth knowing. If one type of job done by one crew goes wrong more often than most, and you can see it before it happens, you can act.

This is the real shift: from counting what went wrong to measuring the things that tend to come before it goes wrong. How often near misses are reported. How often people step in and stop a job. How many inspections get done. How quickly jobs get closed out. Whether the crew bothers to report at all. These move before an accident, not after. That's where you want to be looking.

For this to work, you need a few things.

You need good information, which means a place where people actually report things instead of letting them slide. That's as much about people as technology. It starts with making reporting easy, and making sure that when someone does report, they see something happen as a result.

You need that information joined up, not scattered. An accident in one spreadsheet and a machine fault in another are just two loose facts. Put them in the same place and they become a pattern.

And you need a tool that has learned from your site, your crew and your risks, not just a generic one off the shelf.

None of this is far off. The technology is here and it works. The question is whether the way you do safety is set up to use it, or whether you're still working backwards.

The best time to stop an accident is before the things that cause it have taken hold. That's the whole point of predicting risk. Not to see the future in some clever way, but to read the information you already have and act on it before someone gets hurt.