Question 1

Why can AI now handle most support tickets, and should it?

Accepted Answer

LLM agents can now read your help docs, past conversations, and backend data, which is why Fin (the company formerly called Intercom, now part of Salesforce) reports a 76% average resolution rate and Decagon deployments at companies like Bilt report 75%. The 'should' is more nuanced: Klarna replaced the work of 700 agents with AI, then publicly rehired humans in 2025 after quality complaints. Let AI own the repetitive tier while guaranteeing customers a real human path for complex or emotional cases, and prepare for customers' own AI agents calling on their behalf, which Sierra and Meta's Personal Agent Protocol is built for.

Question 2

How do startups set up AI support (Fin, Decagon, custom bots) in a week?

Accepted Answer

The fastest path is buying, not building: point Fin, the agent built into your helpdesk (Zendesk, or Freddy in Freshdesk) or Chatbase at your help centre, write plain-language guidance and escalation rules, test it against real past conversations, then go live on chat first. Fin also plugs into the helpdesk you already run. Your knowledge base quality is the ceiling on resolution rate, so most of the week goes into fixing docs, not configuring software. If you build custom with Claude or an n8n workflow you get control and lower per-ticket cost, but budget more than a week.

Question 3

How do founders use support conversations as product research with AI?

Accepted Answer

Every ticket is a free user interview, and AI finally makes the whole corpus readable: founders pipe exports into Claude Code or tools like Enterpret to cluster complaints, surface feature gaps, and quantify which bugs actually cost revenue. Even your support bot's failures are research, an unresolved-questions report is effectively a ranked list of where your docs and product confuse people. The one discipline that matters is verification, forcing the AI to quote exact customer words so you act on real patterns instead of plausible slop.

Question 4

What does AI support actually cost, and how should a founder think about the ROI?

Accepted Answer

Two pricing models dominate: per-outcome, where Fin charges $0.99 ($1.99 on voice) only when it resolves a conversation and nothing if the customer asks for a human, and helpdesk bundles, where Freshdesk includes 500 AI agent sessions from $19 an agent a month. Enterprise players like Decagon tie contracts to outcomes. Compare any of these to your fully loaded cost per human-handled ticket, and read the fine print on what counts as an outcome, handoff fees, and volume minimums. Klarna's $40M savings claim and later walk-back is the reminder that ROI includes retention, not just deflected tickets.

Question 5

How do you stop an AI support agent from hallucinating and breaking customer trust?

Accepted Answer

Ground the agent in your real docs, explicitly bound what it may make claims about, label it clearly as AI, and give it a fast exit to a human. The failure cases are instructive: a court made Air Canada honour a refund policy its chatbot invented, and Cursor's support bot fabricated a login policy that triggered public cancellations. Before launch, test against hundreds of real past conversations, then monitor accuracy and hallucination rate as an ongoing operations job, not a one-time setup.

Question 6

When do customers still need a human, and how do you design the AI-to-human handoff?

Accepted Answer

Klarna's reversal settled the debate: AI keeps the volume, but customers must always be able to reach a person, especially for emotional, high-stakes, or account-critical issues. Good escalation fires on explicit requests, detected frustration, and risky keywords like refund or legal, and it transfers the full conversation context so the customer never repeats themselves. The team changes too: instead of a ticket queue, small startups end up with humans running AI operations, knowledge, and the hard 30% of conversations.

How founders use AI for Customer Support

The questions