Meridian is a B2B customer-success platform used by over 600 mid-market and enterprise teams. When their support queue crossed 4,000 tier-one tickets a week, the team knew a staffing ramp wouldn't keep pace with growth. They needed the agent to handle the predictable half of the queue — and sound like Meridian while doing it.
The brief: brand-faithful, not brand-adjacent
Meridian's CX team had tried two other AI tools before Canary. Both deflected some tickets but generated a steady stream of escalations from customers who felt they were talking to a generic bot. The quality review team was spending more time correcting AI tone than it had spent answering questions before. The brief for Canary was blunt: sound like us, or don't bother.
"Before Canary, our tier-one deflection sat at 31%. Eight weeks in it's 58% — and our quality review scores actually went up."
Jordan Tanaka, Head of Customer Experience, Meridian
Week one: loading the voice layer
The implementation team spent the first week on identity, register and refusal style — Meridian's three core brand-voice layers. The identity layer established what the agent is (a Meridian support specialist), what it is not (a general-purpose AI), and what it cares about (time-to-value for the customer, accuracy over speed). Register defaulted to concise and direct, with a warmer override for billing and cancellation topics. Refusal style — how the agent says 'I can't help with that right now' — was the most invested layer, drawing on 200 hand-picked escalation examples from the prior six months.
The results after eight weeks
- Deflection rate: 31% → 58% (week 8, steady-state)
- Quality review pass rate: 84% → 91% (agent-authored responses)
- Median time-to-resolution on deflected tickets: 4 min 20 sec
- Human escalation rate on agent-handled volume: 9%
The deflection improvement alone returned the equivalent of 1.8 FTEs to focus on complex, high-value interactions. The quality review improvement was the bigger surprise: because the agent's voice layer was tuned to Meridian's own standards, reviewers were approving agent responses at a higher rate than they were approving junior-agent responses from six months earlier.
What Meridian would do differently
Jordan Tanaka's one note for teams starting out: invest in the refusal-style layer early. 'The first two weeks we under-specified how the agent should decline feature requests that aren't on the roadmap. We got technically correct answers that sounded evasive. Once we gave the agent explicit language for that scenario, quality scores jumped four points in a week.'