A support copilot that cut resolution time in half

Answering even a routine support ticket meant an agent manually searching across six disconnected systems — a help centre, a ticketing tool, release notes, internal runbooks, old Slack threads, and whatever a senior teammate happened to remember from a past incident.

As the product's surface area grew, the volume of 'have we seen this before' questions grew faster than the twelve-person team could absorb, and average resolution time kept climbing even though most individual tickets weren't actually getting harder — agents were just spending more time searching than solving.

Leadership had looked at chatbot vendors before and walked away uneasy: a bot that could confidently give a wrong answer directly to a paying customer was a bigger risk than the problem it solved. Any AI solution had to be provably safe before it went anywhere near a live customer conversation.

We built an internal copilot — agent-facing only, never customer-facing directly — grounded via retrieval-augmented generation across the same six systems agents had been searching manually, unified into one index the copilot actually searches in seconds.

Every drafted answer carries a visible citation back to its exact source passage, so an agent can verify the claim in one click rather than trusting the model on faith — which is also what got the proposal past a leadership team that had already said no to a black-box chatbot once.

Confidence scoring routes any answer below a defined threshold to a 'needs review' queue instead of drafting it with false certainty, and topics that keep triggering low confidence get flagged back to the knowledge-base team as documentation gaps — a quieter, compounding win that improves the underlying help centre over time.

Agents approve, edit, or discard every drafted response themselves; the copilot never sends anything to a customer on its own. That single design decision — draft, don't decide — was what let this ship in weeks rather than spending months building the trust a fully autonomous bot would have needed.

Average resolution time was cut in half within the first full quarter and first-response time improved even further. Just as important to the team itself: agents report spending noticeably less time hunting for answers across old tickets and Slack threads, and more time on the harder cases that genuinely need a person.