SignalDesk

Client / Project

SignalDesk is an internal Meta.Dev product that demonstrates AI-native customer support routing. We built it against a real ticketing workload (one of our portfolio clients) to validate that an agentic triage layer could replace the human-only escalation queue used by most B2B SaaS support teams.

Problem

The reference workload had three problems that compound: tickets were getting misrouted by keyword-based rules, tier-1 agents were spending more time triaging than resolving, and the escalation queue had a 36-hour median time-to-first-response. The team's first instinct was to hire more agents. Their real problem was classification.

Approach

Two engineers, eight weeks. The system has three components: a classifier that scores incoming tickets across intent, urgency, and required expertise; a retrieval layer over historical resolutions that suggests likely playbooks; and a routing agent that either dispatches to the correct queue or, for high-confidence trivial tickets, drafts a response the agent reviews before sending.

We made deliberate choices to keep the system debuggable. Every routing decision logs the classifier scores, the retrieved historical tickets, and the model's stated reasoning. When a routing is wrong, the support lead can see exactly why and the model can be corrected via labelled examples — no retraining required.

Outcome

Median time-to-first-response dropped from 36 hours to 2.5 hours within the first month of production traffic. Tier-1 escalations dropped 41% — not because tier-1 agents got faster, but because the trivial 30% of tickets that used to hit them now self-resolve. The team kept the same headcount and used the slack to invest in longer-tail problem categories the old system couldn't handle at all.