The draft and the gate
The model reads the thread. Software validates it. You still approve. That is how a field business is allowed to trust the inbox.
The extraction
That sentence is the product. AI turns it into fields a dispatcher can stand behind, with a score on each one, not a black box.
If afternoon is only 82% sure, you see it. If the service is not in your catalog, software rejects the guess. The conversation stays visible either way.
Draft
“Our AC isn’t working. 24 King Street. Tomorrow afternoon?”
AC Repair
24 King Street
Tomorrow afternoon
AC not working
The stack
Is this a job request, a catalog question, or someone asking for a person? Software routes the turn before a job is even considered.
Service, customer, address, window, problem, priority: each with confidence and a note on what is missing.
Deterministic rules check the draft: is that service in your catalog? Empty retrieval is better than a made-up fact.
The raw thread stays next to the draft. Approve, edit, or leave it as a conversation. The model does not invent a job.
What the AI does
What it will not do
Needs you
When the latest message cannot be answered from your FAQs, hours, or catalog, software hands the thread to a human and tells the customer, with no invented times.
Asking for a person hands off. “What services do you offer?” stays on AI. Missing address on a job request is not a handoff. The assistant keeps asking.
Queries · AI handling, then convert

Insights · ask about these numbers

Insights
Owners and admins can ask about mix, load, and completion. Answers come from deterministic aggregates. The chat does not mutate jobs, invent late counts, or dispatch anyone.
We walk a real message or call through extraction, the human gate, and the job, not a slide about generic field software.