An agent that handles 60% of support — and paid for itself in a quarter.

What they came to us with.

Frame's support team was three people answering the same fifteen questions, over and over. The volume was growing twice as fast as the team. Every plan to hire ahead of it kept getting deferred for product work that would generate the volume.

How we approached it.

Most support 'AI' is a chatbot wrapped around a knowledge base. We pitched something else: a real agent with tool use. It could look up customer accounts, check shipping status, file tickets, route to humans when the answer wasn't clean. The metric we cared about was deflection rate — tickets the agent fully resolved without human touch.

What we shipped.

Six-week build. Claude as the backbone, OpenAI for embeddings, a RAG pipeline grounded in Frame's docs and product database. Three tool integrations: account lookup, ticket creation, escalation. Evaluation harness with 200 real historical tickets. Hard human-in-loop gate for refunds, account changes, anything irreversible.