How to find out why your machines stopped, using Copilot in Excel on your downtime log

Every stoppage on your site ends the same way. Somebody gets the line running again, walks back to the terminal or the clipboard, and has about two minutes to write down why it stopped.

They write “jam”. Or “usual”. Or just the name of the machine.

It is a fair thing to write at 3am. The person writing it knows what happened, whoever is on tomorrow probably knows too, and there is a queue of pallets waiting.

Once a month somebody sits down with four weeks of downtime reasons and tries to decide where the fitters’ time and the spares budget should go. The minutes are all there, logged to the nearest five. The causes sit with whoever was stood next to the machine.

In my experience you’ll find something like “agitator failure” in the log. It could be a failed motor, a gearbox, a coupling, a bearing or a shaft.

Operators write better reasons when the box gives them a short list to pick from, in words they already use, and when somebody visibly does something with what they picked. Writing the list and acting on it are jobs for a person. Microsoft changed the counting part in August.

Microsoft’s release notes for 11 to 25 August say Copilot’s editing mode in Excel can now run Python on your workbook. You hand it the log, ask how much of the lost time has a cause you could act on, and it does the sums. The drop-down list goes better with the night shift in the room.

Seen working

I built a downtime log for a chemical filling site that does not exist: 97 stops over four weeks, two shifts, four machines, with the blank cells and typos a real one has. The log is made up. The analysis is real. I ran it through Claude with Python rather than inside Excel, so Copilot’s version of this is untested here.

It split 2,291 lost minutes, about 38 hours, by machine and shift. Then it found that 979 of them, 43%, sat behind reasons like “usual”, “jam”, “wrapper” and “operator adjust”. On the stretch wrapper, 36 of 43 stops had no usable cause. Nights lost nearly twice as long per wrapper stop as days, 18.6 minutes against 9.9, and the log cannot say why.

It also caught a mistyped asset number, two rows where the minutes did not match the start and end times, and one entry filed under “jam” that said a guard had been left open while clearing it.

It fell short in two places. It called “waiting fitter” vague, then left those minutes out of its 43%. Counted in, the figure is 56%. Either number is the tool’s judgement of what counts as vague, so treat it as a rough size. And the drop-down list it suggested for the wrapper was generic. The most useful line in its answer was telling me to rename the list in the night shift’s own words.

Two things worth knowing

1. The Copilot formula in Excel stopped working on 14 September

Microsoft retired the =COPILOT() worksheet function on 14 September, about a year after it appeared in preview, The Register reported on 17 August. Microsoft’s roadmap now says it has decided not to go ahead with the feature. The Copilot side pane stays.

2. Make UK found most manufacturers’ AI sits in the office

Make UK’s report on AI and skills, covered by The Manufacturer on 8 June, found 83% of manufacturers using AI in HR, finance and admin, against 11% in production and 6% in quality control. Over half named skills gaps as the main barrier, especially at technician and operator level, and half said staff have no time to train. Make UK is asking for shift-friendly training.

One thing to try this fortnight

Open last month’s downtime log in Excel and open Copilot from the Home tab. If you cannot see it, ask IT before assuming you do not have it. According to Microsoft’s March notices to admins, firms with fewer than 2,000 Microsoft 365 users kept Copilot inside Excel for staff without the paid licence, on what Microsoft calls standard access. Whether it will run Python for you depends on your licence and your settings, and looking takes ten seconds.

Paste this in:

“This is our downtime log for the last four weeks. Use Python to work out where the lost minutes are going, by machine and by shift. Then list every reason entry that is too vague to act on, and tell me how many minutes sit behind them. Count entries like ‘waiting fitter’ as vague too, because they describe a delay rather than a fault.”

Write down one number: the share of lost minutes with no usable cause. Then ask the second question:

“For the machine with the most vague minutes, suggest a drop-down list of no more than 8 stop reasons, and tell me which ones you are unsure about for our machine.”

Print the list and take it to the two people who log that machine most often, on their shift. Something like: “You two log more wrapper stops than anyone. Which of these do you actually see, and what would you call them?”

Put their version in the sheet as a drop-down. Run the same prompt again in four weeks and compare the number.

About 20 minutes for the first part. The conversation on nights takes as long as it takes.

Final thought

The count gives you two things: the share of your lost time that has no reason attached, and the names of the people who know the reason. The wrapper still needs a fitter. That is a smaller result than August’s release notes suggest, and you can act on it before the next month-end.

If you want to know what your people are already doing with AI, including the sheets nobody told you about, I run a free 20-minute Shadow AI Check. I will not try to sell you anything on the call. Reply and I will send you a time.

Getting what the person at the machine knows into the record, in their words, is the problem I am building SOPwise around.