Aria — an ops agent that doesn't sleep.

The brief.

A 40-person operations team was drowning in inbound requests — customer escalations, partner queries, internal exceptions. Every ticket required a human to read it, categorize it, draft a response, route it to the right specialist, and follow up. By the time tickets were triaged, response SLAs were already at risk.

The team didn't want to grow headcount. They wanted to grow capacity. We were asked: can AI handle the routing layer without losing the judgment that humans are good at?

The friction.

We spent the first week shadowing the team. We learned:

• Triage took an average of 9 steps and 4 minutes per ticket — across 4 tools.
• ~60% of tickets were repeats of patterns the team had handled hundreds of times.
• The hardest 15% of tickets required real human judgment, and those were the ones getting delayed.
• The team had institutional knowledge that lived nowhere — only in three senior reps' heads.

The diagnosis was clear: the team didn't need an AI that replaced them. They needed an AI that absorbed the repeatable 60% so they could focus on the hard 15%.

The system.

We built Aria — a three-layer agent system:

• Classification layer. Reads the ticket, infers intent and urgency, tags it with categories the team already used.
• Response layer. Drafts a reply in the team's voice using a RAG system trained on three years of past resolved tickets and the team's playbook.
• Escalation layer. Knows the 15% it shouldn't touch and routes those to the right specialist with full context attached.

Every reply Aria sends is reviewed by the team before going out. Every edit a human makes is logged and feeds back into the model's weekly fine-tune cycle. The system gets sharper every week — by design.

The outcome.

Six weeks after launch, the team reclaimed roughly 32 hours per week of human triage time. Customer response SLA shifted from "regularly missed" to "consistently beaten by 40%." The team's hardest tickets — the human-judgment ones — are now being handled within an hour instead of half a day.

The compound is what matters: Aria gets better every week from the team's own edits. The system that shipped in week six is materially smarter than the system that shipped in week one, and it didn't take a single new line of code.