System Prompt for a Customer Support Chatbot

A customer support chatbot is the most common AI feature shipped in 2026 and the most common one shipped badly. The difference between a bot that delights and a bot that infuriates is almost entirely in the system prompt. This guide gives you a copy-ready template plus the reasoning behind each section so you can customize it for your product.

If you want to skip to the working version, the free system prompt generator has a Customer Chatbot use case that produces this exact structure with one click.

The Base Template

Here is a working baseline you can adapt:

"You are Ada, a customer support assistant for Acme Corp. Acme sells project management software for small teams (5-50 people). Pricing starts at $19/month per user.

You answer questions about features, pricing, billing, account settings, integrations, and how to contact human support. You troubleshoot common issues using the knowledge base. You cannot process refunds, change billing details, or grant feature access — escalate those to a human agent.

Rules: Stay focused on Acme support topics. Politely redirect off-topic questions. Admit when you don't know something — never invent product features. Use a friendly, professional tone. Keep responses under 100 words unless the user asks for more detail. Ask one clarifying question if a request is ambiguous before answering."

This is roughly 130 tokens. It covers identity, capabilities, escalation, tone, and length. Drop in your company name and price, swap the rules to match your support policy, and you have a working starting point.

Where the Knowledge Base Fits In

If your bot has access to a knowledge base via retrieval (RAG), the system prompt should reference it explicitly. Add a line like: "When you answer factual questions, search the knowledge base first. If no relevant article is found, say so — do not guess."

This single instruction prevents the most common chatbot failure: confidently inventing product behavior that does not exist. The model will sometimes still make things up, but the rate drops dramatically when you write it down as a rule.

For RAG cost planning, the AI cost calculator has scenarios specifically for support chatbots with knowledge base lookups.

Escalation Rules: When to Hand Off to a Human

Every support bot needs explicit escalation rules. The model should know exactly when to stop trying and pass the conversation to a human. Common triggers:

• User explicitly asks for a human — always honor this immediately, never push back
• Refunds, chargebacks, billing disputes — humans only
• Account access issues — humans only
• Legal threats or complaints — humans only
• Three failed attempts to resolve the same issue — auto-escalate
• Frustration signals — detect anger, escalate proactively

Add these to the rules section of your system prompt. Be specific about how to escalate ("respond with: I'll connect you with a human agent now. They'll be with you shortly.").

Sell Custom Apparel — We Handle Printing & Free Shipping

Tone Calibration

Tone is set in the system prompt and felt by every user. Three common tones for support bots:

• Friendly and casual — "Hey! Happy to help with that..." Works for B2C, consumer products, younger audiences.
• Professional and warm — "Of course, I can help with that..." Works for B2B SaaS, professional tools, mixed audiences.
• Formal and precise — "Certainly. The documentation states..." Works for legal, medical, compliance, enterprise.

Pick one and write it into the prompt explicitly. Do not say "be friendly" — say "use first person, contractions, and a warm but professional tone. Avoid slang and emojis."

Common Failure Modes (and How to Patch Them)

The bot makes up features that don't exist. Add: "If a user asks about a feature, only confirm it exists if you can find it in the knowledge base. Otherwise say 'I'm not sure if that feature exists — let me connect you with a human who can confirm.'"

The bot promises refunds it can't process. Add: "Never promise a refund. Direct all refund requests to a human agent immediately."

The bot rambles. Add: "Default to responses under 100 words. Only go longer if the user explicitly asks for more detail."

The bot refuses too much. Add a positive capabilities list. The model is more likely to engage when you tell it what it CAN do, not just what it cannot.

Testing Your Prompt Before You Ship

Build a small eval set of 20-30 messages that represent your edge cases: angry users, off-topic questions, refund requests, ambiguous questions, multi-turn conversations. Run your prompt against all of them. Look for: bot following the rules, bot escalating correctly, bot maintaining tone, bot admitting uncertainty when appropriate.

Iterate the prompt until the eval set passes. Then run it against real conversations from your existing support tickets — you will find new edge cases and patch them too. The bot is only as good as your eval set, so spend time on it.