How much of my support volume can a bot realistically handle?

Well-built bots deflect roughly 40 to 70 percent of inbound tickets, concentrated in repetitive how-to and account questions your docs already answer. The remainder — edge cases, angry customers, anything touching money or policy — should escalate to humans. Aim to remove the boring, high-volume tickets, not to eliminate your support team.

How do I stop the bot from making up answers?

Ground every answer in retrieval and constrain it in the system prompt to use only the retrieved context, cite its source, and say it does not know when the context is thin. Pair that with confidence-based escalation so a weak retrieval triggers a human handoff rather than a confident guess. Grounding plus escalation is the core anti-hallucination pattern.

When should the bot escalate to a human?

Escalate on three triggers: low retrieval confidence (the docs do not cover the question), sensitive topics such as billing, refunds, cancellations, security, or legal, and any explicit request to talk to a person. Always pass the full conversation so the customer never has to repeat themselves to the agent.

How do I integrate with Intercom or Zendesk?

Use their APIs and webhooks. Subscribe to new-conversation events, send the message into your bot, post the reply back through their messaging API, and on escalation assign the conversation to a human queue with a note containing the transcript. Most platforms also expose a native bot or app slot you can plug into directly.

How do I measure whether the bot is working?

Track deflection rate (tickets resolved without a human), customer satisfaction on bot conversations, escalation rate, and time-to-first-response. Watch satisfaction closely — a bot that deflects tickets while frustrating customers is a net loss. Sample real transcripts weekly to catch wrong answers your metrics miss.

How to Build an AI Customer Support Bot

What you are building

This tutorial builds a production AI customer-support bot — the kind that answers “how do I reset my password?” or “what’s your refund window?” instantly, around the clock, from your own documentation, while quietly escalating anything it shouldn’t touch. The architecture is retrieval-augmented generation wrapped in safety rails: the bot retrieves relevant help content, answers strictly from it, cites its source, and hands off to a human the moment it is uncertain or the topic is sensitive. Done right it removes the repetitive, high-volume tickets so your team handles only what genuinely needs judgement.

How it works

Three pieces. Knowledge ingestion: chunk and embed your help centre, FAQs, and past resolved tickets into a vector store, so the bot’s answers are grounded in content you actually wrote. The answering loop: embed the customer’s message, retrieve the closest chunks, and prompt the model with a safety-first system prompt that says use only this context, cite the article, never invent prices or policy, and admit uncertainty. Escalation and integration: when retrieval confidence is low, the topic is sensitive (billing, cancellations, security, legal), or the user asks for a person, the bot hands the conversation — with full history — to a live agent through your Intercom or Zendesk API. The combination of grounding plus confidence-based escalation is what keeps it from confidently inventing answers.

Tips and the ROI estimator below

The single biggest quality lever is the escalation policy, not the model. Over-escalate at first, then tighten as you watch transcripts — a bot that says “let me get a human” is recoverable; one that confidently states the wrong refund policy is a liability. Always pass the full conversation on handoff so customers never repeat themselves. Track deflection rate and satisfaction together, because deflecting tickets while annoying customers is a net loss. The estimator below turns your ticket volume, deflection rate, and per-ticket cost into a monthly savings and payback figure so you can size the opportunity before you build.