Maintenance requests become work orders: Celeste for property management

A property manager’s inbox is a mix that punishes slow reading: a dripping faucet, a question about subletting, a request for a rent ledger, and somewhere in the pile, a unit with no heat. The faucet can wait until Thursday. The no-heat cannot wait until the inbox gets read. The job is triage, and triage at 11pm is exactly what a person cannot staff and an agent can.

Celeste works the tenant-facing side of this desk on email and web chat.

Intake that ends in a work order

“My kitchen faucet is dripping” is not answered with sympathy, it is answered with intake. The agent collects what a dispatcher needs: which unit, which fixture, how bad, since when, photos if the tenant has them, permission to enter. Connect your maintenance system as a tool and the result is a structured work order, not a thread for a human to re-read and retype.

Creating a work order is a write, and writes in this product follow one rule: you classify each tool action before the agent may execute it, and until then it can only draft. Most managers run work-order creation as held drafts for the first weeks, approving each with one click, then classify routine intake as automatic once the drafts have proven clean. The emergency path is separate and stricter, which is the next section.

Emergencies are barred from automation

Put “flood,” “no heat,” “gas,” “leak,” “smoke” on the escalation keyword list, and a message containing one is barred from an autonomous reply by code that runs after the AI. It is held for a person immediately, flagged with the word that tripped it, so the no-heat email is distinguishable from forty faucets at a glance. The agent never gets to decide an emergency is routine, because that decision was never delegated to it.

Lease answers and verified rent status

Lease questions are document questions. Ingest your lease templates, house rules, and move-out procedures, and “can I sublet,” “what is the pet deposit,” and “how do I get my security deposit back” answer from what the documents say, consistently, instead of from whoever replies fastest. When the documents do not answer, the agent logs the gap by name and routes the question to your team.

Rent balances and payment history are records, so identity comes first. A tenant who writes in verifies who they are, and every lookup is scoped to their own tenancy by code outside the AI. No tenant can see another unit’s ledger, and an unverified sender gets no account data at all. Anything past status, payment plans, disputes, anything with discretion in it, goes to a human with the context already gathered.

Turned up one topic at a time

The agent starts in shadow mode, drafting silently against your real tenant traffic while you read what it would have said. Lease FAQs usually graduate first, then maintenance intake as held drafts, then rent status once verification has run clean. The escalation list is live from day one in every mode.

The gain shows up twice. Tenants get an answer at the hour they wrote, which is most of what tenant satisfaction actually is. And your staff opens a morning queue that is already sorted: work orders drafted, emergencies flagged, lease questions answered, with only the judgment calls left.

See how it works for tool classification and escalation, or the product for the manager’s view. The trial starts silent, so you can watch a week of real tenant email get triaged before the agent sends anything.