A Real Customer Support Workflow with Agentic AI: Where It Helped and Where It Didn't

说明与用途

将代理式 AI 部署到客户支持中，实现了分类、信息检索和草稿生成的自动化，但在政策、过时文档和情感识别方面出现问题。通过工具调用限制、确定性升级规则和来源要求等安全措施进行改进。

问题

Integrating an agentic AI into a live customer support workflow to automate routine tasks while ensuring safety and reliability.

结果

The AI handled triage, context gathering, drafting for repetitive questions, and summarizing long threads, improving efficiency. Failures included policy edge cases, stale docs, tool loops, misreading emotion, over/under‑escalation, and hallucinations. Guardrails such as tool‑call limits, deterministic escalation triggers, and source requirement mitigated risks.

工作流

• 1. Receive ticket
• 2. Agent classifies intent, urgency, sentiment
• 3. Agent performs read‑only account/order lookup
• 4. Agent retrieves relevant knowledge‑base articles
• 5. Agent drafts reply with confidence and sources
• 6. Policy check: auto‑send if low risk/high confidence, otherwise human review
• 7. Human reviews or escalates as needed

证据

The agent answered refund questions fluently, including cases where our policy had exceptions.

来源

When two articles disagreed, the agent picked one without telling us.

来源

On vague tickets, the agent sometimes called the same lookup tool repeatedly with slight variations.

来源

A frustrated customer who wrote "great, another broken update" was sometimes read as positive.

来源

Early on, the agent escalated almost everything, which defeated the purpose.

来源

When an order lookup returned nothing, the agent sometimes filled the gap with plausible but made‑up details.

来源