Right Sales Intelligence. Right Team. Right Context. Right Time.
Overview
Field sales representatives were spending 30 to 45 minutes per query manually searching through multiple, structurally similar plan documents just to check their Incentive Compensation status. With different teams, multiple quarters, and overlapping document formats, the answers they found were frequently inaccurate, misattributed, or simply unusable.
ICASSIST (Incentive Compensation Assistant) is a GenAI-powered sales assistant built for the Sales team, enabling field representatives to gain precise, contextual insights from both unstructured PDF plan documents and structured historical performance-payout data through a single conversational query, resolved in under 60 seconds.
The Solution
ICASSIST was built on a four-pillar vision to transform fragmented, look-alike sales documents into a secure, context-aware, team-specific AI knowledge platform.
- Precision Over Noise Metadata-filtered retrieval: enforces Team, Year and Quarter boundaries at query time eliminating cross-team content contamination entirely. Every response is scoped to the correct plan before semantic search is applied.
- Context Preservation: An advanced chunking and context enrichment strategy ensures individual chunks retain awareness of the broader document they belong to preventing the loss of meaning that plagued the previous naive RAG implementation.
- Governance and Security: The platform is SSO-enabled with controlled access and a structured feedback logging mechanism ensuring enterprise-grade security and accountability across all user interactions.
- Continuous Learning: A feedback loop captures user responses to improve retrieval quality over time making the system progressively more accurate with each interaction.
- Structured Data Access: Historical performance and payout data stored in Snowflake was made query able through natural language, enabling representatives to access earnings history and attainment data without any technical knowledge.
- Visual Content Analysis: AWS Bedrock and hosted Vision Large Language Models (VLLMs) were integrated to process image-based and visual content within plan documents ensuring no information was lost due to non-text formatting.
Technologies Used
- Dataiku: Enterprise AI development platform for collaborative data science, ML model development, and MLOps at scale
- Snowflake: Cloud-native data platform for secure, scalable data warehousing and structured payout data storage
- AWS + AWS Bedrock: Cloud infrastructure and managed foundation model services powering Generative AI inference and VLLM-based visual content analysis
- ChromaDB: Open-source vector database for storing embeddings and enabling fast, metadata-filtered semantic search
- Python: Core language for data engineering, AI pipeline development, and backend systems
- Flask: Lightweight web framework for building the conversational AI interface and API layer
Impact Created
| Before ICASSIST | After ICASSIST |
| Sales teams manually searched multiple structurally similar plan documents. Significant time was spent verifying quarter and year relevance. Cross-team content retrieval produced contaminated, low-trust outputs with no reliable path to a precise answer. | Context-aware retrieval enforces Team, Year, and Quarter at query time. Responses are accurate, filtered, and delivered through a one-question conversational workflow in under 60 seconds. |
Quantified Outcomes
- 40–60% reduction in document search time from 30–45 minutes of manual PDF searching to 30–60 seconds via conversational query.
- ~25,000+ hours saved annually across 200 daily sales users, with an average of 30 minutes saved per user per day.
- Cross-team data leakage eliminated via enforced metadata filtering at retrieval.
- Significant improvement in retrieval accuracy over the previous naive RAG baseline.
- Increased user confidence and adoption driven by SSO-enabled access and a seamless conversational experience.
Conclusion
ICASSIST demonstrates that Generative AI delivers its greatest enterprise value not simply by answering questions but by answering the right questions, for the right team, with the right context, instantly.
By replacing a naive RAG implementation with an advanced, metadata-aware retrieval architecture governed, secure, and continuously improving ICASSIST transformed a fragmented, time-consuming manual process into a precise, scalable knowledge platform trusted by the sales organization.
"The right answer, for the right team, at the right time that is the standard ICASSIST was built to meet."
