Jessica
Customer Experience Director @ SupportPro

Support that sees, understands, and solves instantly.

Jessica
Customer Experience Director @ SupportPro

"Our support teams are drowning in repetitive questions that take time away from complex issues needing human expertise. During peak periods, our average response time jumps from 2 minutes to over 15 minutes, frustrating customers and agents alike. For example, when a customer sends a screenshot of an error message or a photo of a damaged product, our agents have to manually search through knowledge bases to find relevant solutions. Our basic chatbot only handles the simplest queries and fails when customers phrase questions differently from our predefined templates. With customer expectations for instant support growing, we need a system that can understand multiple input types, access our specific knowledge bases, and provide accurate, consistent answers regardless of volume."

Expected Achievements

Reduced Wait Times82% decrease

Higher Resolution Rate37% increase

Agent Productivity65% more

Challenge

SupportPro faces significant challenges managing high volumes of customer inquiries across multiple client businesses. Support teams become overwhelmed during peak periods, with wait times increasing from 2 minutes to over 15 minutes. This creates customer frustration and impacts satisfaction scores. Approximately 70% of inquiries involve repetitive questions that agents answer multiple times daily, preventing them from addressing complex issues requiring human expertise. The company's basic rule-based chatbot fails when customers phrase questions differently from training examples or when inquiries involve company-specific knowledge not programmed into the system. When customers submit screenshots, photos of products, or document uploads, agents must manually interpret these visuals and search through knowledge bases, further increasing resolution time.

Read in detail

Our Strategy

SupportPro's challenges with response consistency, handling multimodal inputs, and managing peak volumes are clearly unsustainable. To address these issues, we designed an intelligent conversational AI system with multimodal capabilities and Retrieval-Augmented Generation (RAG). This solution understands various input types and accesses specific knowledge to provide accurate, contextual responses.

Dataset Creation & Expansion with AI

We collect SupportPro's existing customer conversations, knowledge base articles, product manuals, and resolved tickets. Our initial dataset included 75,000 text conversations and 15,000 multimodal interactions involving images or documents. We then use AI to generate additional synthetic conversations that cover edge cases and rare scenarios, expanding our dataset to over 120,000 diverse interactions.

Historical support data

REFINEMENT

AI-powered conversation synthesis

REFINEMENT

Comprehensive multimodal dataset

Fine-Tuning the Multimodal LLM

We take a large multimodal language model and fine-tune it to understand both text and visual inputs in the context of customer support. The model learns to

interpret screenshots, product images, and document uploads while maintaining conversation context.

Short Animated Video: Show a customer uploading a screenshot of an error message and the model instantly recognizing the error type, identifying relevant solution

articles, and generating a helpful response without requiring additional explanation.

Multimodal Dataset

Pretrained Model

Fine-tuned Model

Test Dataset Creation

We create a challenging test set with complex queries, multimodal inputs, and conversations requiring specific knowledge. These 5,000 test cases include difficult

scenarios like ambiguous questions, partial information, and mixed-format inquiries to ensure robust evaluation.

Evaluation & Model Refinement

We evaluate the system using metrics like answer relevance, factual accuracy, and conversation coherence. Initial testing showed 76% accuracy for complex queries.

After three iterations—including additional training on edge cases and improving the retrieval mechanism—we achieved 91% accuracy for text queries and 87% for

multimodal inputs.

Progress Bar Game: Show a bar filling up each time we improve the system's results. Label different accuracy levels as "Basic Bot,

" "Smart Assistant,

" and "Support

Expert.

REFINEMENT

RAG Implementation & Knowledge Integration

We implement a Retrieval-Augmented Generation system that indexes all of SupportPro's knowledge sources into vector databases. The system can retrieve relevant information in real-time during conversations and ground its responses in verified company knowledge.

Testing System Performance & Scalability

We test the system's response times and accuracy under various loads, simulating peak periods with hundreds of simultaneous conversations. Our benchmarks show consistent response times under 1.5 seconds even at maximum capacity, with no degradation in answer quality.

Final Solution

After completing these six steps, we deliver a fully integrated multimodal conversational AI system with RAG capabilities. SupportPro uses it to handle customer inquiries across all channels, processing text, images, and documents with equal proficiency. This solution provides: Instant Responses, Visual Understanding, Consistent Accuracy, Seamless Escalation and etc. The system now handles approximately 75% of all incoming inquiries without human intervention, achieving a 92% customer satisfaction rating that matches human agent levels. Support agents report spending 65% more time on complex cases requiring empathy and judgment, improving job satisfaction and reducing turnover. As the system continues learning from interactions, its handling capabilities are expected to expand to 85% of all inquiries within six months, while maintaining or improving quality metrics.

Read in detail