Case study · 18 March 2026
A copilot on service tickets cut first-response time
RAG over SLA contracts and repair history. Dispatchers get a drafted reply with a citation — not a chatbot for the end customer.
A field-service network maintained HVAC in retail sites. Tickets arrived by email, portal and call centre. The first useful reply depended on who sat at the inbox and whether they remembered a 2022 contract addendum.
The problem
Mean time to a substantive first response: 11 hours. Not because of headcount. Because the truth lived in SLA PDFs, old work orders and senior technicians’ heads. A new dispatcher took six months to become useful.
What we deployed
A copilot on the dispatcher’s side — not a public chat for the tenant. On each ticket: a suggested class, a contract quote (with clause number), similar jobs, a draft first reply. Nothing sends until a human accepts.
Sources: SharePoint contracts, CMMS history, call-out price lists. If the model cannot find a citation, it says “I don’t know” and queues a senior. That product rule was harder than the model.
The result
First response fell to 2.4 hours. Wrong SLA citations dropped close to zero, because a reply cannot leave without a quote. Seniors stopped being the search engine.
Custodianship: once a month we ingest addenda and learn from dispatcher rejects. Without that the index would rot in a quarter.
