A maintenance chatbot gets requested in the same meeting as the demo. Someone pictures second shift typing a fault into a tablet and getting back the last repair, the part number, and whether the fix held. In a Macomb County stamping plant, that request makes sense. The senior millwright still knows which press reports fault codes it did not throw, and that person will not be on nights forever.
What the plant has today is a CMMS the night crew works around. Production keeps a separate spreadsheet for the morning meeting, and the paper work orders in the maintenance office stop mattering once the press is running. People open a work order because a supervisor needs a downtime code. They close it with "repaired" or "adjusted" because another call is already waiting. A chatbot pointed at those work orders will reply in full sentences that sound cleaner than the notes they came from.
Write the log a stranger can use
The log has to be enough for a new hire at 2 a.m. without a phone call: which asset stopped, what failed, what the tech did, which parts were used, and how long the machine sat, including whether that time was repair or waiting. Drop one of those and the person on the floor already has to find someone who was there. A chatbot stops in the same place. It cannot walk to the crib, and it cannot ask the operator what the screen showed at 6:40.
Names get in the way before the notes do. Production calls the machine the 400-ton. Maintenance calls it Press 4. The CMMS calls it KO-400-04, because that was the asset tag when somebody built the database in 2017. Search Press 4 and you miss the jobs filed under the other two names. Pick one name per machine, put it on the work order, and use that same name in the morning meeting.
A code that says mechanical or electrical does not say whether the hose at the cushion failed or a sensor on the feeder did. A code that says operator settles blame and leaves the next tech with no instruction. The note that helps next month names the part and the condition. A usable note reads like this: hydraulic hose at the cushion, rubbed through at the clamp, replaced with part 4410-H, press back up at 10:15. A short code list can sit next to that note if the crew will choose from it. A long list that everyone files under "other" by Thursday gets ignored. Keep the list matched to how those machines break, and leave room for the written note.
Photos have the same problem. A rub mark or a cracked connector stays on a personal phone until the phone is replaced. If the finding mattered enough to photograph, write it on the closed work order, even when the CMMS makes attachments painful.
Parts fall off the work order even when they went into the machine. A bearing comes out of a drawer in the maintenance shop instead of through the crib, so the work order shows no parts. The next time that bearing fails, there is no part number from the last repair. Write the part number down whenever a part touches the machine. That includes pulls that never hit the crib screen.
Close the work order before the shift leaves
If the work order stays open, a later reader cannot separate the repair from a plan still sitting on the board. Close it with a clock time and a name, and the next shift has a work order it can use. Techs skip the close-out because the machine is up and the radio has already moved on. Do not leave the job marked in progress overnight unless the machine is still down and the oncoming crew needs to know that. The maintenance lead reads yesterday's closes before the production meeting, while the people who did the work are still in the building.
When a close-out is wrong, correct it the same day. Add the correction on the work order and leave the original line where it is, so a later reader can see what changed.
Preventive routes need the same kind of note as breakdowns. A route marked complete from a desk, with no gauge reading and no finding, tells every later reader that the route never turns anything up. When the route finds play in a guide or heat at a bearing housing, write the finding even if the repair waits until Saturday. "Play in the ram guides, scheduled for the weekend" gives the next person a place to start.
Downtime has to mean the same thing to maintenance and to production. Production may book 40 minutes to maintenance while maintenance books those same minutes to waiting on a skid. Any later summary will blend the two, and neither side will recognize their own morning. Split the time while the call is fresh. Record whether the crew was waiting on a part, waiting on a person, waiting on production to clear the die, or working on the machine. Have that argument on the shift it happened. By the monthly review the details are gone, and a chatbot will treat whichever number it found as the fact.
The morning meeting is often where the chatbot idea starts. Someone wants a pareto of downtime, the codes are too coarse to explain the lost hours, and a vendor offers a chatbot that promises to explain the chart. A week of specific closes will change what that chart can say. A chatbot sitting on top of "mechanical" will only narrate "mechanical" in longer form.
One area for a month
Leave the rest of the plant on its current habit. Take the cell that already produces the phone calls. In a stamping house that is often a single press line. In a machine shop it is often the lathe cell that gives up an hour of second shift to the same tool alarm. For a month, only that area keeps the stricter log.
Include someone from the off-shift in that check. The people who will lean on the log are often the ones who were not in the room when the code list was written. A weekend tech should be able to read Friday's closes on Sunday night and know what changed on the press. If the notes only make sense to the day-shift lead who wrote them, Sunday night still requires a phone call.
At the end of the month, hand the closed work orders to someone who does not run that equipment. A newer tech, or a process engineer who will sit with the screen, is enough. Ask them to say, from the work orders alone, what failed last, what was replaced, and whether the same fault came back. If they can say it, a chatbot limited to those work orders has something solid to read. If they still have to ask the millwright, the chatbot will not have the answer either.
Keep the OEM manual out of that test. The manual describes how the builder thinks the press should run. The log describes what this press has done on this floor, including repairs that diverge from the book. Shops drop both into the same shared folder, then get a generic procedure when they asked what this press did last time. Load the manual later if you want help with alarms, and keep it apart from the work orders.
When the chatbot does go in, keep the boundary tight. Closed work orders for that one area. Every answer names the work order and the date. If the note is thinner than one line, the chatbot should say the work order is too thin to answer. Lockout and the decision to restart stay with the person at the machine.
The month costs the maintenance lead's attention, plus some grumbling from techs who have been closing jobs in four words for years. If a CMMS is already installed and people dodge it, correct the asset names and the close-out on that one cell before anyone prices a chatbot. If the jobs are still on paper, type the closed work orders for that cell into a sheet the crew can search. Searching a sheet the lead trusts is a smaller project than a chatbot on a system people open only when a supervisor is watching.
Skip the month and the demo still looks clean, because the sample work orders were written for the demo. During the first week a tech asks about a fault that has come back twice and receives a paragraph that could describe three different repairs. They call the millwright. They keep calling until the subscription comes up for renewal.