Practical Prompt Engineering for Customer Support in Malaysian SMBs

What is Prompt Engineering for Customer Support?

For a Small and Medium-sized Business (SMB), customer support can be a constant battle. You need to be responsive and helpful, but you likely don't have a large team. This is where Large Language Models (LLMs) like GPT-4o or Claude 3.5 Sonnet can assist. However, getting useful, consistent, and safe responses from them requires a specific skill: prompt engineering for customer support.

This isn't about complex coding. It's the practice of writing clear, structured instructions (prompts) that guide an AI to perform a specific task correctly. For support, this means drafting replies that are accurate, empathetic, and aligned with your company's voice, while knowing when to escalate to a human.

At JRV Systems, we handle support queries for our own SaaS products and for client projects. We've found that a small investment in creating and refining support prompts pays for itself quickly through faster resolution times and more consistent service quality.

Choosing the Right AI Tool for Your Budget

Enterprise-grade AI support platforms can be expensive. For a Malaysian SMB, a more practical starting point is a consumer-level subscription to a powerful LLM. These tools offer an excellent balance of capability and cost.

• OpenAI's ChatGPT Plus (with GPT-4o): Costs around USD $20/month per user. It's a strong all-rounder, excellent at understanding nuanced requests and generating human-like text.
• Anthropic's Claude Pro (with Claude 3.5 Sonnet): Also around USD $20/month. Claude is often praised for its larger context window (handling more information at once) and a slightly more cautious, professional tone out of the box.
• Google's Gemini Advanced: Similar pricing. It integrates well with the Google ecosystem and has strong multi-language capabilities.

For most SMBs, the choice between them is less about raw power and more about which interface your team prefers. The core prompting techniques discussed here will work across all of them. The key is to have a human agent who reviews, edits, and sends the final message. The AI is an assistant, not a replacement.

Core Prompt Patterns for Support Agents

A good support prompt is a recipe with several key ingredients. Leaving one out can lead to a generic or incorrect response. Your base prompt should always include these four elements.

1. Persona and Tone: Explicitly tell the AI who it is. Without this, it defaults to a generic, often overly verbose chatbot. Define its role, its personality, and the language it should use.

   • Example: You are a friendly and professional customer support agent for 'Kedai Runcit Online', a Malaysian e-commerce platform. Your tone should be patient and helpful. Respond in simple, clear English.

2. Context Injection: This is the most critical part. The AI knows nothing about the customer's problem until you provide the details. Always include the full, unedited customer message.

   • Example: Here is the customer's message: [Paste the customer's full email or WhatsApp message here].

3. Knowledge and Constraints: Provide the specific information the AI needs to answer the question, and tell it what not to do. This is your primary guardrail against hallucinations (making things up).

   • Example: Use the following information about our shipping policy to answer: [Paste relevant text from your FAQ or knowledge base]. Do NOT invent any policy details. If the answer is not in the provided text, state that you need to check with the team.

4. Task and Format: Tell the AI exactly what you want it to do. Do you need a draft reply? A summary for your ticketing system? A list of action items?

   • Example: Based on the context, draft a reply to the customer. The reply should acknowledge their problem, provide a solution based ONLY on the knowledge provided, and end by asking if they need more help.

A Real Example: The JRV Systems "Friday" Bridge

We built an internal tool named "Friday" that bridges our WhatsApp support channel to our team's chat. Initially, we used a very basic prompt to summarize incoming messages.

The 'Before' Prompt: Summarize this WhatsApp message for the support team: [Customer Message]

This was too simple. The summaries were often poor, missed the urgency, and didn't suggest any action. It created more work.

The 'After' Prompt (Simplified):

You are a support triage assistant. Your task is to process an incoming customer message from WhatsApp and prepare it for the JRV Systems support team.

**Customer Message:**
"{{customer_message}}"

**Instructions:**
1.  **Summarize:** Briefly summarize the user's issue in one sentence.
2.  **Sentiment:** Classify the user's sentiment as Neutral, Positive, or URGENT.
3.  **Categorize:** Categorize the ticket into one of these categories: Technical Issue, Billing Query, Sales Inquiry, Other.
4.  **Draft Initial Reply:** Draft a polite, professional reply in English. Acknowledge their message and inform them that the team is looking into it and will respond shortly. Do NOT attempt to solve the problem.

**Output Format:**
Provide your response as a JSON object with the keys: "summary", "sentiment", "category", "draft_reply".

This refined prompt transformed the tool's utility. We now get structured, actionable information instantly. The team can see the priority at a glance, and we have a consistent, pre-approved initial reply ready to go. This is a perfect example of how prompt engineering for customer support directly improves workflow efficiency.

Guardrails Against Bad Advice

An AI giving incorrect information to a customer is a significant risk. Here are two non-negotiable rules for using LLMs in a support role:

• The Human is the Final Check: The AI's output should always be treated as a draft. A human agent must read, verify, and potentially edit every single response before it is sent to a customer. This is the most important safety measure.
• Explicitly Forbid Guessing: Your prompt must contain a clear instruction for the AI to follow when it doesn't know the answer. A simple line works well: If the information needed to answer the question is not in the provided context, you MUST respond with 'I do not have enough information to answer this question.' This prevents the model from trying to be helpful by inventing answers.

By implementing these structured prompts and strict human-in-the-loop workflows, any Malaysian SMB can leverage the power of modern AI to deliver better, faster customer support without the enterprise price tag.