The copilot that took over the busywork.
A fintech's support team was drowning in tickets that all looked slightly different but were mostly the same. We built an agent that drafts the reply and routes the rest.
The problem
Ticket volume had tripled in a year. Most tickets were variations of the same forty questions, but each one still needed a human to read it, find the right macro, adjust it, and route the exceptions. Agents spent more time triaging than actually helping the customers who needed a person.
What we built
An assistant inside their existing helpdesk — not a new tool to learn. It reads each incoming ticket, drafts a reply grounded in their help docs and past resolutions (with sources shown), and routes anything it isn't confident about to the right specialist queue. Agents review, edit, and send; nothing goes out without a human.
We spent the first two weeks on evals before touching the workflow: building a test set from six months of resolved tickets so we could measure draft quality honestly instead of guessing. That test set still runs on every change.
The outcome
Manual triage dropped 70%. First-response time fell from hours to minutes on the routine queue. The interesting part: CSAT went up on complex tickets too, because agents finally had time for them.
Knowledge search for a law firm
Drowning in tickets too?
Start with a two-week AI Audit. We'll tell you if a copilot is worth building — and prove it with your own data.