01
Problem
Ticket volume tripled in 12 months. Hiring couldn't keep up. Quality slid. Churn was creeping up. Leadership had bought "AI support" pitches before and been burned.
B2B SaaS · Customer support
AI support that handles 70% of tickets — and the customers love it.
A Claude-powered agent built on Salesforce Service Cloud, with strict tool calls, golden datasets and a kill-switch for when reality drifts from the eval set.
STACKSalesforce Service CloudAnthropic ClaudeAWS FrankfurtLangSmithStripeVector DB
- Tickets auto-resolved
- 70%
- Retention uplift
- +18%
- Design → production
- 4 weeks
02
Approach
No AI demo. A four-week production rollout: golden dataset first, deterministic guards in code, structured outputs, full conversation logging, kill-switch tied to a human approval queue.
03
Solution
Claude agent integrated into Salesforce Service Cloud. Tool calls into the customer's product (subscription, plan, usage), RAG over years of documentation and resolved tickets, automated evaluation harness running on every prompt change.
04
Outcome
70% of tier-1 tickets resolved without human touch. Retention up 18% — driven by sub-minute response time on simple issues. CSAT held above 92%. Zero "critical" incidents in production thanks to the kill-switch design.