The chiller tripped again at 11:40 on a Saturday, three weeks after an AI copilot helped day shift resolve the exact same fault. That first work order is closed and marked "resolved."
What it doesn't say is which sensor trend actually mattered, which of the tool's three recommendations the technician acted on, why she passed on the other two or how anyone confirmed the repair held.
So the night technician starts from zero, with the same alarm and less context than the person who already solved it. The ticket turned green, but the learning evaporated with the shift change, and the better closeout test for AI-assisted work asks a harder question: could another qualified person explain this decision and recover without the tool?
Document the maintenance decision, not the AI conversation. The next technician doesn't need a transcript of everything the copilot said; they need the story of the repair, starting with what the symptom looked like under those operating conditions, which data the first technician actually trusted and what the working theory was.
From there, it should say which of the tool's recommendations the human took and why the others lost, then what got done and how it held up. Anything still uncertain goes in too, with a validator's name at the bottom to close it out. Attach the transcript if it helps, but a raw AI conversation dumped into the CMMS forces the next reader to reconstruct a judgment call from chat logs, which nobody on a night shift is going to do.
IFMA's guidance on practical AI use in facilities keeps drawing a line between what the tool suggested and what a person decided. The closeout record's whole job is to preserve that line, briefly enough that someone actually reads it.
The evidence that knowledge transferred is a supervised teach-back, where a second qualified person walks through the evidence, picks the next step and knows when to stop. There's a real ladder here: watching the fix happen is the bottom rung, and explaining the decision, including why the rejected options were rejected, sits in the middle.
Performing or simulating the recovery under authorized conditions is the top, and it's the only rung that makes the capability real. None of this requires repeating a risky live intervention for practice; a low-risk scenario, a simulation or the next planned maintenance window does the job safely. Whatever the second technician misses becomes the most valuable output of the exercise, because it goes straight back into the record or the procedure. Facility teams already know this pattern from construction handover, where a binder full of documents has never once equaled operating readiness.
If losing the AI, its data feed or its interface could turn a manageable fault into a safety, compliance or mission problem, the team needs a degraded-mode route it has actually tested. The route is mundane on purpose, and it answers practical questions rather than architectural ones: where the relevant control physically sits, and which readings must be checked against what. It also names who may authorize action at 2 a.m., which conditions demand escalation, and how the system comes back to normal afterward.
The goal is not to recreate the algorithm by hand; it's safe recognition, containment and recovery while the tool is dark or doubtful. Not every fix earns this treatment, though. The sizing questions are practical, like how bad a failure would actually be, whether the action can be undone and how quickly qualified support can arrive.
Their answers fit the fallback to the risk better than any blanket policy will.
The NIST AI Risk Management Framework runs the same proportionality logic through its oversight, recovery and fail-safe controls; what's written here is just the wrench-level translation.
Once the diagnostic steps, manual controls and recovery decisions are written down, the remaining question is who can actually perform them, and when. Map those tasks against roles and shifts in a competency matrix, and the coverage picture stops being a feeling. What managers usually find is uncomfortable in a useful way: one person holding the entire recovery for a critical asset, or weekend nights with nobody inside two rungs of the teach-back ladder.
Sometimes it's someone still listed as capable on the strength of a course they finished two years ago. The matrix makes those gaps visible; it does not prove anyone can do the work. Demonstrated performance, observed or supervised, is what closes a gap, and pairing the person who owns the fix with the person who should is what mentorship in facility management has always been for.
The immediate stake isn't retirement statistics so much as operational depth the next time the original fixer is on vacation.
The closeout rule fits in one breath. The decision is recorded, a second person has passed the teach-back, a fallback proportionate to the consequences exists, the shift map shows real coverage and someone owns the next update. Retest when events call for it rather than on a universal calendar. In practice that means a recurrence, a meaningful change to the asset or its control logic, an updated procedure or model, or a critical name dropping off the coverage map. AI genuinely shortens the path to a correct intervention, and that's worth having. Making the intervention reusable is still the facility team's job, and it starts the moment the ticket wants to turn green.
Author's note: Fabian Sandoval is a writer covering facility operations, workplace technology and the people who keep buildings running. His work appears on IFMA's Connected FM blog and Jobillico, where he focuses on what new tools actually change for the teams expected to use them.