When Everything Is Urgent, Nothing Is: Fixing Maintenance Triage

Every work order your team gets is an emergency. Ask the resident. The dripping faucet is an emergency. The flickering porch light is an emergency. The squeaky cabinet hinge, somehow, an emergency. And buried in that pile, at 11pm on a Saturday, sits the one that actually is: water coming through a ceiling, a gas smell, a lock that won't lock.

That's the trap. When every ticket arrives flagged URGENT, the flag stops meaning anything. Lula calls this out by name: "everything comes in as urgent," which masks the real emergencies and turns your coordinator into a human switchboard fielding the same question all day. What's the status. What's the status. What's the status.

So the job is to sort first, then act. Answering faster comes second. Here's the five-step workflow that does it, and the numbers that come out the other side.

Step one: read the ticket and name the problem

A resident doesn't file a "Category 3 plumbing event." They text "water everywhere help." The first move is natural-language classification: the AI reads the message and figures out what's actually broken. Plumbing, electrical, HVAC, appliance, lockout. No dropdown menu the resident ignores, no mistyped category a coordinator has to fix later. It reads the words like a person would, and tags the issue.

Classification alone is where most tools stop. It's also where most of the value is still sitting on the table.

Step two: score the urgency on safety, ignoring volume

This is the part that fixes the "everything is urgent" problem. Each ticket gets scored Emergency, Urgent, or Routine, and the score is built on safety criteria rather than on how many exclamation points the resident used (Lula). Gas, flooding, no heat in January, a security failure: Emergency. A running toilet: Routine, no matter how the message reads.

Now the flag means something again. The Saturday-night gas call jumps the line because it earned the spot, while the running toilet waits its turn no matter how the message reads.

Step three: route to the right vendor automatically

Once the issue is named and scored, it goes to the best-fit vendor without a coordinator playing telephone. Lula reports scheduling kicks off within 10 minutes of submission. Compare that to the usual choreography: coordinator reads it, picks a vendor, calls the vendor, waits for a callback, confirms with the resident, updates the file. Every handoff is a delay and a chance to drop it.

The payoff shows up in two numbers. Response runs about 25% faster, and 80% of jobs get resolved in a single trip (Lula). One trip matters more than it sounds. The repeat truck roll is where margin quietly leaks out of a maintenance operation.

Step four: close the status loop before anyone calls

Remember the human switchboard. The resident texts "what's the status?" because nobody told them anything, so they assume nothing's happening. Automatic status updates kill that loop. Vendor assigned, here's the window, here's the confirmation. The resident stays informed, and your coordinator stops fielding the call that exists only because the last message went unanswered.

Across the whole workflow, Lula puts coordinator time spent on coordination down 60-65%, with operational cost down 10-30%. That's most of a role handed back to the work that actually needs a human.

Step five: feed the result back in

Every completed job is data. Was the urgency score right? Did the vendor solve it in one trip? Feed that back and the scoring sharpens over time. The system that's been running for six months triages better than the one you switched on yesterday.

The line you don't let AI cross

Here's where to be careful, because this is the part a vendor will gloss over. AI owns the routine band. The judgment call stays with a person.

Maintenance urgency scoring decides what auto-dispatches and what a coordinator looks at first, and the after-hours emergency is exactly the case that wants a human in the loop (Lula). The pattern across support AI backs this up. Payment and account questions deflect cleanly at 80-90%, but complex lease and legal issues only land at 20-30% and should escalate (AI Consulting Network). Angry or high-priority messages get routed to a person automatically on sentiment (Fini Labs). Gartner frames the whole strategy the same way: autonomous coverage of the common stuff, plus explicit exception handling for the rest. Full automation was never the goal.

Call it act-with-a-gate. The AI clears the flood of routine tickets, dispatches them, keeps everyone updated, and routes the genuine edge cases to your staff with full logging. The machine handles volume. The human handles the call that actually carries risk.

That's the whole trick. Stop treating every ticket like the worst one, so when the worst one shows up, somebody's actually looking.