The Shift Report · Issue 15 · 25 October 2026
How to write up a near-miss on your phone at the spot with ChatGPT or Copilot
A near-miss written up back at the desk has already lost the forklift number, the driver’s name, the product code. AI on your phone can collect it on the spot in a couple of minutes — deciding what went wrong and what needs fixing is still yours.
You’re walking to the transport office past the warehouse when a forklift comes out of the door quite fast, without sounding the horn the SOP requires. You stop the driver, remind him of the rule and carry on. The near-miss gets filled in about half an hour later, when you’re back in the office.
A near-miss often gets written up long after the issue was spotted, back in the office or even at the end of the shift, when a lot of the detail is already lost from memory.
Sometimes I can’t recall details of the situation, like the forklift number, the operator’s name, the concentration of the chemical involved or the product code. Then the near-miss entry is not specific. It reads as generic and loses its impact.
At the moment it happens, the detail is all in front of you. AI on your phone can help you collect it. Take a photo, dictate what happened without thinking much about how you say it, and let the tool ask you what’s missing while you’re still standing there. Then ask it to write the report under the headings of your own form.
How to do it the next time you want to report a near-miss
- Set it up once, at your desk. Most people on Microsoft work accounts have Copilot Chat. I find ChatGPT Plus much better for this kind of job.
- ChatGPT: select New project in the sidebar and call it “Near-misses”. Open the ••• menu next to the project name, choose Project settings, and paste the instructions from step 2 into the instructions box.
- Copilot: open Copilot Chat with your work account and select New agent in the left-hand pane. Choose Skip to configure, paste the instructions from step 2 into the instructions box, name it “Near-misses” and select Create.
You can improve the instructions as you go.
- Give it the instructions. You can start with this one. Put the headings from your own near-miss form in the third paragraph, so the report drops straight into it.
“You help me write up near-misses on our site, on my phone, at the place where it happened, before the people involved go home. I will send you a photo and tell you what happened.
First, tell me in two or three lines what you can see in the photo that matters. Then ask me up to five questions I need answered for our near-miss form, the ones I should ask while I am still here. Do not write the report yet.
When I send the answers, write the report under these headings: What happened; When and where; Who was involved; What could have happened; Why it happened (only what I told you or the photo shows); What we did straight away; What needs fixing, and by whom.
Use only what I have told you and what is in the photo. Where the form needs something I have not given you, write NOT KNOWN. Do not decide whose fault it was. Under What needs fixing, put only what people suggested. If you have ideas of your own, list them underneath, marked as your suggestions.”
- When you see a near-miss, open the app and take a picture. Open the Near-misses project in the ChatGPT app, or the Near-misses agent in the Microsoft 365 Copilot app. Tap + by the message box and take the photo from where you were standing. If you can, get in shot what you’ll need later: the truck’s fleet number, the label on the load, the door or the sign.
- Talk it through. Tap the microphone and say what happened, in your own words. Add as many comments, observations or opinions as you like: how fast it was going, what you think caused it, what the driver said. The more you give it, the better the report. Then send the message.
- Answer its questions while you’re still there. It comes back with what it can see in the photo and up to five questions. When I tried this on a made-up near-miss, it read the truck number off the photo and asked for the driver’s name, the time and the product code on the drums. Ask the driver what you need to, check the label, and send the answers back. Then ask it to write the report.
- Check it. Ask: “Is there anything in your report that I did not tell you and is not in the photo?” It will tell you where it used its own judgement, for example what could have happened. Change anything you don’t agree with. If you need it shorter for a free-text box, ask for that.
- Put it into your form. Copy the report into your near-miss form or system, or email it to yourself and paste it in later. Check every name, number and time before you submit it.
Two things worth knowing
1. A model of the line that sees the stoppage coming and suggests the fix
On a mayonnaise line in Poznań, minor stoppages fell by up to 20% and waste by nearly 30%. At a deodorant plant in North Carolina, waste fell 20% and capacity rose 10%. Both run a digital twin: a computer model of the line. The figures are the company’s own, with no independent check.
The sensors were already there. A sensor tells you a temperature or a pressure as it is now. The model takes those live readings and predicts what the line will do next. In North Carolina it predicts 95% of blockages in product flow. In Poznań it keeps the mayonnaise’s thickness steady. On other sites it recommends temperature setpoints, fan speeds, or how much of an expensive ingredient to dose. A person approves each correction; only some low-risk ones happen automatically.
For a site your size, a full twin of a line is out of reach for now. Even this programme is going into only about 40 of several hundred factories, the most advanced ones, with sensors at the right points and the network to carry the data. The part that scales down is predicting when a machine needs maintenance from its sensor data, which is already sold as a sensor on the motor plus a yearly data subscription. (Unilever, reported by Supply Chain Dive.)
2. The barcode scanner types into the PC behind it, and so could someone outside
To the PC on the desk, a wireless scanner is a keyboard. The scanner reads a label, sends it to its base station, and the base station types it into the PC. Researchers testing a scanner from a major manufacturer found the base station did not check that the signal came from its own scanner. Someone within Bluetooth range could pick up its address by listening, then send fake keystrokes, which the PC accepts as somebody typing. That can be enough to run commands on that PC.
Radio passes through windows, gates, roofing and thin cladding, so that person could be in the car park, the next unit or a delivery bay. The researchers saw the same weakness on scanners from several makers, Honeywell among them, and think handheld terminals and label printers could be used the same way to reach the warehouse management system. No real attack has been reported.
This one sits closer to home than most cyber news: it needs someone near the building, and the weak point is kit your pickers carry every shift. The fixes the researchers list are ordinary ones: Bluetooth off where it is not used, scanners and printers on a separate part of the network, and device software kept up to date. (Kaspersky, reported by DC Velocity.)
Final thought
The detail a near-miss report needs is easiest to get right on the spot, at the time it happened. Half an hour later, back in the office, some of it has already gone. The tool helps you collect it while you’re still standing there, and it marks what it guessed. Deciding what went wrong and what to fix first is still your job.
I spend a lot of my time working out where operations lose time, especially in training, and how the new tools can help. Direct message me if you want to find out more or set up a call.
SOPwise turns a Standard Operating Procedure into a training package: task-based steps, critical mistakes, and scenario questions. £35 per SOP, no subscription. Built for UK manufacturing and chemical distribution.