Automate Ecommerce Customer Support with Custom AI

Teams often rely on the rule-based automation in Gorgias or Zendesk. These tools are great for simple triggers like tagging tickets containing "Where is my order?". They fail when logic depends on data outside the helpdesk, like a customer's lifetime value in Klaviyo or specific line items in a Shopify order.

Consider a return request. A simple rule can check if the order is within 30 days. But what if your policy is different for final sale items, or if the customer is a VIP who gets exceptions? Gorgias rules cannot look up product tags in Shopify or a customer segment in Klaviyo. This forces agents to manually check 3 different systems for every return, a process that takes 5-10 minutes per ticket.

The core problem is that helpdesk macros are stateless. They cannot chain multiple data lookups and conditional checks together. To decide on a return, you need to check the order date, query the product SKU for a "final_sale" tag, and check the customer's email against a VIP list. This multi-step, data-dependent logic is where rule engines break down, leaving your team with manual work.

Syntora would approach this problem by first conducting a discovery phase to understand your current support operations. This would involve connecting to your existing Shopify and helpdesk APIs, such as Gorgias or Zendesk, utilizing their official Python clients. We would then ingest relevant historical data, typically the last 6 months of ticket history and order information, to thoroughly understand your most common support requests and business context. This initial data ingestion and analysis phase typically requires 3-5 business days. This historical data would then become the foundation for crafting precise prompts, instructing the AI on your specific business rules, brand voice, and desired responses.

The core system architecture would involve a FastAPI service, acting as an orchestration layer for calls to the Claude 3 Sonnet API. When a new ticket is received via a Gorgias webhook, for instance, this service would fetch the customer's full order history from Shopify's GraphQL API. It then dynamically constructs a comprehensive prompt, combining the customer's message, order details, and your specified return policy. Claude's API then generates a categorized response and a draft email reply based on these inputs.

The FastAPI application would be deployed as a serverless function using AWS Lambda. This architecture is designed for cost-effectiveness and scalability, with operational costs for processing up to 10,000 tickets per month generally being minimal. For secure management of API keys, AWS Secrets Manager would be utilized. Supabase would serve as the logging destination for every request and AI response, providing a comprehensive audit trail and facilitating debugging. The entire deployment process would be managed using Vercel for continuous integration and efficient updates.

The delivered system would incorporate structlog for structured JSON logging, transmitting these logs to Supabase. A customized Retool dashboard would be developed on top of this log data, providing clear visibility into ticket volume, automation rate, and average processing time. For proactive alerts, monitoring would be implemented, such that if the automation rate were to drop below a predefined threshold, for example 70% for more than an hour, a pre-configured CloudWatch Alarm would dispatch a notification to a designated Slack channel.