AI triage for a field-service helpdesk inbox
An HVAC and refrigeration service business in Sydney receives roughly 200 customer emails a week. The office coordinator was reading every email, deciding whether it was a warranty call, a maintenance booking, a billing query or a sales lead, and then manually creating a job in their service-management system. It was consuming four hours a day.
We built an automation on Microsoft Power Platform that watched the shared inbox. When a new email arrived, Azure OpenAI extracted the customer name, the asset described, the fault, and the requested response. It then drafted a service ticket in the existing system with a suggested category and priority.
The coordinator did not lose control. The automation drafted the ticket; the coordinator reviewed and approved. Over three months we measured the draft accuracy and tuned the prompt so that, by the end, roughly 80% of tickets were approved without edit. The remaining 20% were warranty claims that required a human judgement call.
Out-of-hours emails now queue overnight so the first technician in the morning already has a list of jobs to allocate, rather than the coordinator sorting through email first thing.
The business is now extending the same pipeline to SMS messages sent to the service line.
The service business had twelve field technicians and a helpdesk inbox that received between forty and sixty emails a day. The dispatcher read each email, decided what it was about, drafted a ticket, and assigned it. This took between ninety minutes and two hours every morning. We deployed an Azure OpenAI workflow that reads incoming emails, categorises them by problem type, drafts a ticket with the customer's name, contact number, vehicle or job reference, and the issue described, and queues it for the dispatcher's approval.
The workflow runs on Microsoft Power Automate, which is already part of the company's Microsoft 365 subscription. No new SaaS product was introduced. The dispatcher receives a list of pre-drafted tickets each morning and approves, edits or rejects each one. The average time per ticket dropped from four minutes to under one.
We set a rule that the system never books a technician without a human approval. The system drafts. The dispatcher decides. Every AI automation follows the same sequence: read, summarise, suggest, draft, human approves. The automation reduces the repetitive admin. The human retains the judgment.
The dispatcher now spends forty minutes a morning on ticket triage instead of two hours. The technicians receive better-structured tickets, with the customer's problem already summarised, so they arrive on site knowing what they are dealing with. In the first month, the business closed 15 percent more tickets per day without adding staff.