Guardrails for Agentic Customer Experience

Executive Summary

Trustworthy agentic customer service requires controls that operate during a conversation, not just after-the-fact error detection. Crescendo's approach splits responsibility into three parts: governing the interaction (instructions + enforced technical permissions + handoffs), evaluating outcomes (predictive CSAT, custom QA), and governing subsequent changes (testing, attribution, versioning).

Concierge, Crescendo's customer-facing agent, operates within an explicitly defined scope — what it supports, what policies apply, and what actions it may take — rather than an open-ended list of prohibitions. When evidence or authority is insufficient (e.g., unverified account changes), the system hands off to a human rather than improvising.

Critically, this discipline extends to the agents that improve the system. Optimization Agent, Knowledge Agent, and others can propose better instructions or surface knowledge gaps but cannot expand permissions or policy on their own. That authority stays with people. Every proposed change is tested against a persistent "Evals" suite before deployment, with full version history and change attribution.

The overarching thesis: Enterprise trust in agentic AI depends equally on constraining what the system can do today and maintaining accountable, evidence-based control over how it changes.