Unique Photo partnered with Space Dinosaurs to reduce growing customer support volume and improve response speed without increasing operational overhead. By identifying high-impact automation opportunities, we implemented an AI-powered chat solution that scaled support, improved consistency, and significantly reduced reliance on manual workflows.
The Challenge
Unique Photo was experiencing increasing customer inquiries that required manual support, creating:
- Operational inefficiencies
- Slower response times
- Limited scalability in customer service
They needed a solution that could automate common interactions while maintaining a high-quality customer experience.
The Approach
Space Dinosaurs conducted a comprehensive AI assessment to:
- Identify high-value automation opportunities
- Develop an AI adoption roadmap
- Validate potential solutions
The assessment revealed that an AI-powered chatbot could significantly reduce support overhead while improving response speed and consistency.
The Solution
Space Dinosaurs designed and implemented a Live AI Chat solution tailored to Unique Photo's needs, focused on accuracy, reliability, and user experience.
Architecture
Every incoming message is classified by intent and routed to the appropriate handler: human handoff, an order or package helper, or the AI. When the AI handles a reply, the system first retrieves relevant documents, then produces a single answer grounded in them.
Chat service
- Dedicated chat GraphQL API (Apollo) for starting chats, sending messages, assigning agents, and signed-URL access, layered on the existing product application
- Server-Sent Events with Redis pub/sub for cross-pod real-time delivery of messages, activity indicators, and status changes
- JWT-signed URLs for secure, time-limited external access to a chat
- Chat data model (tickets, messages, audit metadata) persisted in the product database
AI pipeline
- Azure OpenAI for chat completion and embeddings
- Azure AI Search hybrid retrieval (keyword + vector, with optional semantic ranking) over two purpose-built indexes: product catalog and site content
- The AI is constrained to answer from the retrieved documents, reducing hallucinations and keeping replies on-brand
Reliability and safety
- Prompt injection detection at the chat entry point, before any message is stored or sent to the model
- Centralized prompts and model parameters in a single configurable module
- Per-message audit trail recording which path handled each reply (RAG, general inquiry, helper, or handoff)
- Graceful fallback from RAG to general inquiry on configuration or retrieval failure, with activity indicators that never get stuck
Key Capabilities Delivered
Advanced AI Chat Functionality
- Intent recognition for smarter query handling
- Dynamic search routing based on user input
- Automated fallback responses for edge cases
Search & Retrieval Optimization
- Hybrid full-text + vector search implementation
- Improved response relevance, speed, and accuracy
Reliability, Security & Monitoring
- URL validation and prompt injection detection
- Audit logging for traceability
- User sentiment tracking for continuous improvement
User Experience Enhancements
- Image attachment support within chat
- Activity indicators and improved responsiveness
- Reconnection handling and refined error messaging
Documentation & Enablement
- Comprehensive implementation documentation
- Delivered as both PDF and in-code Markdown for ongoing maintainability
The Results
- Enabled "Always-On" Support: By acting as a highly capable front-line concierge, the new chatbot absorbed the massive influx of traffic that comes with an always-active widget, engaging with 89.5% of all new conversations (1,000 tickets in the first 33 days).
- 4.8x Increase in Daily Automated Resolutions: Prior to the new bot, the system only fully resolved an average of 2.8 tickets per day. Post-launch, the AI successfully resolved an average of 13.7 tickets per day end-to-end (455 total tickets in 33 days) without ever requiring a human agent to step in.
- Massive Workload Shielding: In the first month of "always-on" operation, the AI autonomously handled 2,323 individual chat messages. By absorbing this high-volume tier-1 communication, the AI effectively shielded human staff from chat fatigue, allowing them to step in only when necessary.
- Automated High-Friction Tasks: The system's intent routing successfully intercepted and processed 535 specific "Order Status" inquiries in the post-launch window. By handling these via a dedicated backend helper, the AI eliminated one of the most historically repetitive manual workflows for Unique Photo staff.
- Collaborative Handoff: The AI sustained an average of 2.29 back-and-forth messages with the user to gather context and attempt resolution before seamlessly transferring the user to an agent for complex, high-value inquiries.
