How Assembled Automates 50%+ of Support Cases Using Claude AI

Assembled was built by engineers who spent time at Stripe and saw firsthand how customer support fails to scale. As a company grows and its products get more complex, the gap between what customers need and what a support team can deliver widens rapidly. Assembled’s founders set out to close that gap not by replacing human agents, but by giving them better tools and coordinating them more intelligently with AI. Today the platform serves enterprise customers including Brooks, Warner Brothers, Robinhood, and Stripe.

The support industry’s prevailing response to scale has been deflection: use AI to intercept and block low-complexity tickets from ever reaching a human. Assembled took the opposite bet. Their view was that the real problem is Tier 2+ tickets — cases that require multi-step reasoning, access to account context, and something close to empathy. These are the tickets that define customer relationships. Deflecting easy questions while leaving hard ones unaddressed was not a strategy; it was a way to erode trust at scale.

Assembled Assist deploys Claude as the core reasoning engine. The system uses meta-prompting to dynamically adapt ticket categorization to each customer’s unique taxonomy. Routing decisions — which cases go to AI, which go to human agents, and which require escalation — are made in real time based on case complexity and customer sentiment signals. The architecture is multi-model, enabling seamless failover between providers. For complex reasoning tasks, Claude 3.5 Sonnet handles the cognitive load; Claude 3 Haiku runs high-speed, lower-stakes operations. Every agent response flows through an evaluation system that builds a golden dataset over time, measuring accuracy against customer-specific standards.

The clearest test of that architecture came during a competitor outage. Assembled migrated all LLM workflows to Claude in real time, sustaining service for enterprise customers through five to six hours of downtime without interruption. The reliability forced a broader reassessment: after comparing performance across tasks, Assembled began reevaluating its entire model infrastructure in favor of Claude. For one major customer, accuracy on AI-generated support responses climbed from 62% to 90% across hundreds of ticket categories.

Assembled’s current trajectory points toward making language barriers irrelevant in global support delivery — using LLMs to let agents respond fluently across languages without separate staffing by region. The broader argument their results make is that AI in support does not have to be a cost-reduction play. At their best customers, satisfaction went up while support spend went down simultaneously. That combination, difficult to achieve without the right coordination layer between AI and humans, is what Assembled is building toward.