If AI chatbots are the future, I hate it

Original Article ↗ Hacker News Discussion ↗

Nature of the Problem Discussed

  • The concrete case involved an ISP “chatbot” that was really a rigid decision tree, not an LLM.
  • It repeatedly forced Wi‑Fi troubleshooting despite clear statements about a wired, line-level speed drop.
  • After escalation, the human agent also followed a script and missed provided data.
  • A later onsite visit revealed a backend billing / provisioning downgrade to a 6 Mbps plan, likely due to messy legacy systems.

Are These Actually “AI” Chatbots?

  • Many argue the example is misclassified: it’s keyword + dialogue-tree logic, common for years, not modern generative AI.
  • Some say calling it “AI” is misleading and clickbaity; others note that expert systems and decision trees have long been labeled “AI” in industry.

Experiences with Chatbots vs Humans

  • Widespread frustration: bots waste time, block humans, repeat irrelevant scripts, and often don’t pass prior context to agents.
  • Others report positive cases: e.g., broker chatbots changing investment settings, Amazon-style bots that gather context well, an ISP with a transparent scripted flow, a Carvana bot solving a complex title issue.
  • Some users actively prefer machines to avoid exhausting human calls and long hold times.

Economics and Incentives

  • Many see customer support at large ISPs and platforms as a pure cost center, especially where there’s little competition; the goal is to be just good enough not to lose regulators or customers.
  • Debate over whether high prices imply room for better service: some cite thin margins and scale; others point to monopoly behavior and cost-cutting, not necessity.
  • Several argue most traffic is basic (“what’s my balance?”, bill pay, password reset), making automation attractive.

Design Ideas and Future of Support

  • Suggested improvements:
    • Clear “talk to a human” escape.
    • Technical “shibboleth” or quiz to fast‑track advanced users.
    • Bots that honestly declare themselves, are optional, escalate quickly, and share all gathered info with agents.
    • Better internal documentation and tools for support staff.
  • Some foresee LLM-based agents that know account history, understand natural language well, and hand off smoothly, eventually surpassing today’s human-first model.
  • Others doubt incentives will ever align to deliver that best-case hybrid, predicting further degradation before any improvement.