A client came to us drowning in repetitive support tickets. Here's exactly how we shipped a Claude-powered triage and response agent over a single weekend — stack, prompts, guardrails, and the rollout plan that made it stick.
The stack
A thin API layer in front of Claude, a vector store of their existing help docs, and a simple escalation rule: if the model's confidence dips below a threshold, it hands off to a human instead of guessing.
The guardrails that mattered
We spent more time on what the agent shouldn't do — refunds, account deletion, anything irreversible — than on what it should. Clear boundaries are what made the client trust it enough to actually turn it on.
The result
Sixty percent faster first response, and a support team that now spends its time on the 20% of tickets that actually need a human.