Helio — Autonomous support agents for a high-volume SaaS

A deterministic AI layer in front of Zendesk that resolves the majority of incoming tickets without human intervention — and escalates the rest with full context.

Helio runs a project management platform with eighty thousand paying teams. Twelve thousand support tickets a week, an average first-response time of eleven hours, and a CSAT score sliding toward 3.2. Hiring more agents wasn't economically viable. Hiring fewer wasn't operationally viable.

We built a policy-bound agent layer that classifies, routes, and resolves tickets against a typed interface to Helio's product database, billing system, and knowledge base. When the agent is uncertain, it escalates with a structured handoff — customer context, attempted resolutions, a confidence score — that human agents can act on without re-investigating from scratch.

We chose deterministic over probabilistic by design. The same question gets the same answer. Refunds follow a published policy. Escalation paths are auditable end to end. The interesting engineering wasn't in the model — it was in the constraints around it.

The orchestration runs on Cloudflare Workers, with Durable Objects holding ticket state and a typed RPC layer binding tool calls to internal services. Observability flows through OpenTelemetry into Honeycomb, so the support lead can see, in real time, what the agent is doing and why.

Six months in: 67% of tickets resolved without human touch. First response under ninety seconds. CSAT at 4.6. The support team grew by two people instead of the projected fourteen.