Automated Customer Service Categorization with WorkBuddy Saved Enough for Two Monitors

Just read that article about WorkBuddy intelligently generating tweets. My first thought was, I haven't used that feature, but I've been using it to handle customer service messages for about a week and can't live without it anymore.

Let me describe my setup. I run a solo e-commerce store: one Shopify shop, one WeChat Mini Program shop, and occasional listings on Xianyu (Idle Fish). Across these three platforms, I get dozens of customer inquiries daily, mostly repetitive questions about logistics, sizes, and returns/exchanges. Previously, I read and replied to each one manually. Often, I'd pack shipments during the day and still be replying to messages at 11 PM, feeling completely drained.

Then I watched a WorkBuddy tutorial explaining how it breaks tasks down into agents for processing. I wondered if I could integrate customer service. Surprisingly, I managed to make it work. Now I save about two hours daily, which I use for organizing inventory, packing, listing new items, or just lying around.

Here’s how I did it, step-by-step for beginners.

Step 1: Open the left sidebar in WorkBuddy, click New Task, and select Agent Orchestration. This feature essentially lets you create multiple AI roles that work independently and collaborate. I created an agent named "Customer Service Classifier" with the task description: Deduplicate and classify exported customer service messages from various platforms into four categories: "Logistics Inquiry," "Product Inquiry," "After-sales Issues," and "Other," outputting them as a table.

Step 2: Key part here. I enabled web search functionality because it needs real-time logistics info. In the agent settings, go to Tool Configuration, enable the Web Search and Table Processing plugins, and disable unused ones to prevent erratic behavior. For data sources, I chose file upload. I dragged in the CSV file of customer messages exported from the Shopify backend. For the Mini Program, I copied and pasted text directly. Same for Xianyu.

Step 3: Click Run. It took a few minutes. The output table labeled each message with a category and automatically calculated proportions. Accuracy was above 80%. Ambiguous cases like "Can you lower the price?" were categorized as "Other," which I understood.

But classification alone wasn't enough; I wanted auto-replies. So I created a second agent, "Auto-Replier," chaining it to the "Classifier." In the Process Orchestration page, I set the "Classifier's" output as the "Replier's" input. In the Replier's task description, I specified: Logistics replies use Template 1, Product replies use Template 2, After-sales use Template 3, Others transfer to human agents and mark as pending. I wrote the templates in the system prompt to prevent hallucinations.

Here's a pitfall I must mention. On the first run, the auto-reply included terms like "Dearie" (Qinqin). I quickly revised the prompt to maintain a neutral, concise, professional tone, adding "Do not use any intimate forms of address." After fixing, it worked normally.

Regarding permissions: WorkBuddy allows setting view/edit rights. Currently, only I can see agent results due to customer privacy concerns—I don't even dare give access to outsourced CS staff. However, it has a Share function to export task results as links. I occasionally send logistics inquiry stats to suppliers to show recent issue volumes. These links have access passwords and expire in seven days.

Daily operations: Every morning, I run the "Classifier" first to aggregate previous day's messages from all three platforms. Then I run the "Replier," review the outputs, and only then batch send. I haven't fully trusted the auto-send feature yet since mistakes damage reputation. So I manually click Run before bed, check results in the morning, and manually click Send.

Avoidance tips: Don't expect WorkBuddy to do everything perfectly in one go. Initially, I tried to combine classification, reply, statistics, and daily report generation in one task, resulting in chaos. Splitting into three small tasks, each doing one thing, made it faster and more stable. Also, web search sometimes lags; if there are widespread logistics anomalies, searched info might be inaccurate, so I manually verify major disruptions.

Honestly, this workflow saves me nearly two hours daily with a lower error rate than manual replies. Calculated against market rates for outsourcing CS, I save about 1,500 RMB monthly. Over a year, that covers the cost of two ultrawide monitors.