One of the more interesting things we’re learning about AI is how much its ability to consistently produce the intended outcome depends on data cleanliness, not in a single field, but across the entire lifecycle of the work.
A useful way to think about it is a two-hour story cut into five sections. Remove any one of those sections and large parts of what remains stop making sense. A repair behaves the same way. It is not a ticket. It is a narrative with a beginning, a middle, and an end.
That is why intake consistency and taxonomy matter as much as the models themselves. A taxonomy is a shared naming system that takes the many ways a repair can be described and places each one under a consistent label.
Without that structure, one request may be recorded as a dripping faucet, another as a leaking kitchen tap, and a third simply as a plumbing issue. They may represent the same kind of work, but connecting them reliably becomes much harder.
Why maintenance has historically been difficult to structure
Maintenance data hasn’t historically been created for this purpose. Customizable fields allow organizations to describe the same work differently. Human intake introduces variation based on who enters the request. Residents describe symptoms rather than diagnoses, while technicians and vendors may document the resolution differently than the original problem was described.
None of that necessarily prevents a repair from getting completed. But it makes it much harder to connect millions of repairs into a consistent dataset that models can reliably learn from.
Our models now map incoming work into a repair taxonomy of roughly 6,700 subcategories. As more becomes known about the service required, that structure connects the original issue to what follows: diagnosis, work performed, time to completion, costs, and resolution.
That’s what makes the record useful beyond simply documenting that a repair happened.
Intake becomes more important as AI does more
Diagnosis isn’t only valuable because it removes a task from someone’s plate. Better information at the beginning can support intelligence later around expected completion times, expected costs, repair history, documentation, and what actually occurred across the life of the repair.
Verify Agent is one example of what becomes possible when that full cycle is intact. Rather than treating each note, image, or data point in isolation, it can reference information across the repair to help understand what happened and where an operator’s attention may be needed.
The same principle will matter as agents begin supporting more decisions throughout the coordination lifecycle.
What this changes for maintenance operations
As agentic coordination becomes more capable, “good operations” will increasingly depend on the quality and continuity of the information surrounding the repair.
The question isn’t simply whether a resident interacts with a human or AI. It’s whether the repair has enough context to get the right person involved at the right urgency, with the right capability and an appropriate understanding of what the work should cost.
That makes data cleanliness much bigger than keeping fields and notes organized. It becomes a property of the entire repair lifecycle.
And somewhat counterintuitively, as AI takes on more of the coordination surrounding maintenance, intake will matter more, not less.