Build a Support Chatbot
Last updated: 2026-05-18
An AI support chatbot answers questions, deflects tickets, and escalates when needed. This guide covers tool selection, knowledge base setup, conversation design, testing, deployment, and monitoring. No-code and code-based approaches both work — choose based on your resources.
Tool Selection
No-code — Intercom Fin, Zendesk Answer Bot, Freshdesk Freddy. Built into support platforms. Train on docs and past tickets. Fast to deploy.
Low-code — Botpress, Voiceflow, Landbot. More control over flows. Still visual. Good when you need custom logic.
Code — Custom build with OpenAI, Anthropic, or similar. Full control. Requires development. Use when platform tools are insufficient.
Start with platform-built AI if you use Intercom, Zendesk, or similar. Add custom build only when you hit real limits.
Knowledge Base Setup
Content — FAQs, product docs, help articles, past tickets (anonymized). The chatbot answers from this content. Quality of content equals quality of answers.
Structure — Organize by topic. Use clear headings. Chunk long docs for RAG. Update regularly.
Ingestion — Most tools ingest from URLs, PDFs, or pasted text. Some support integrations with Notion or Confluence. Make sure all relevant content is included.
Conversation Design
Greeting — Set expectations upfront. "I'm an AI assistant. I can help with X, Y, and Z. I'll connect you to a human if needed."
Scope — Define what the bot can and can't do. Avoid overpromising.
Handoff — Clear path to a human. "Talk to an agent" or "Request a callback." Preserve context when handing off so the customer doesn't have to repeat themselves.
Tone — Match your brand. Friendly, professional, or technical. Configure this in the tool.
Testing
Test cases — Common questions, edge cases, and out-of-scope questions. Does the bot answer correctly? Does it escalate when it should?
Adversarial testing — Try to confuse it. Off-topic questions, gibberish, multiple questions at once. See how it handles failure gracefully.
Human review — Have support agents review sample conversations. Fix gaps in the knowledge base or prompts.
Deployment
Channels — Website widget, in-app, Slack, WhatsApp. Deploy where your customers actually are.
Rollout — Start with a subset of traffic. Monitor. Expand when quality is acceptable.
Fallback — Always offer a human option. Some customers prefer it. Some issues require it.
Monitoring
Metrics — Resolution rate, deflection rate, escalation rate, CSAT. Track over time.
Conversation review — Sample conversations weekly. Find failures and improve.
Knowledge gaps — When the bot fails, add to the knowledge base or adjust prompts.
Common Pitfalls
Bad training data — Outdated or wrong docs. Garbage in, garbage out. Audit and update content regularly.
No human escalation — Customers stuck with a bot that can't help. Always provide a handoff option.
Overpromising — Bot claims it can do things it can't. Set a clear scope.
Ignoring feedback — Customers report issues; no one acts on them. Use feedback to keep improving.