The era of AI agents is rapidly advancing, exemplified by 11x's groundbreaking work in building "Alice," an AI Sales Development Representative (SDR) designed to learn and operate with human-like nuance. Sherwood Callaway, Tech Lead for Alice, and Satwik Singh, AI Lead at 11x, illuminated their journey in developing Alice's sophisticated knowledge base at the AI Engineer World's Fair in San Francisco. Their presentation highlighted how this innovation is transforming the traditionally laborious process of sales outreach.
An SDR's core responsibilities typically involve sourcing leads, contacting them, and booking meetings, with success measured by positive replies and scheduled appointments. Alice, 11x's AI SDR, now handles these tasks at an unprecedented scale, sending approximately 50,000 emails daily across 300 different business organizations—a stark contrast to the 20-50 emails a human SDR might send. This massive output necessitates an equally robust understanding of both the seller's offerings and the leads' needs.
Previously, 11x's sellers manually input details about their products, services, pain points, and value propositions into a "Library" dashboard to "train" Alice. This manual process was inherently tedious, creating significant onboarding friction for new customers. As Sherwood Callaway noted, "The user would have to enter details about every individual product and service and all of the pain points and solutions and value props associated with them in our dashboard." This cumbersome approach often resulted in suboptimal emails, either lacking sufficient relevant context or being overwhelmed with too much information, hindering Alice's ability to craft truly smart, personalized outreach.
The pivotal shift came with the introduction of the Knowledge Base, designed to emulate how a human SDR is onboarded. Instead of sellers manually "pushing" information to Alice, the AI agent proactively "pulls" all relevant context about the seller's business. Satwik Singh emphasized this human-like learning: "If you're training a human SDR, you would kind of get them in, and then you will basically dump a bunch of documents on them, and then they ramp up throughout a period of like weeks or months. Similarly, Knowledge Base is basically a centralized repository on our platform for the seller info, and then users can kind of come in, dump all their source material." This centralized repository accepts diverse data types, including PDFs, images, websites, audio, and video, ensuring comprehensive understanding.
To facilitate this, the pipeline first involves "parsing" non-text resources into text format, making them "legible" to Large Language Models. 11x opted to leverage third-party vendors for this complex task rather than building from scratch due to the sheer volume of resource types and the specialized expertise required. For documents and images, they chose LlamaParse (a LlamaIndex product), while Firecrawl handles website parsing. For audio and video, Cloudglue was selected for its ability to extract information directly from the video content, not just the transcript. After parsing, "chunking" breaks down long texts into logical units, which are then stored in Pinecone, a vector database chosen for its well-known reputation, cloud-hosted nature, and bundled embedding models.
The retrieval phase utilizes a "Deep Research" agent, built with Letta, which accepts lead information as input. This agent formulates relevant questions, queries the Pinecone vector database using specialized tools, and generates concise answers. This multi-step, iterative process allows for more nuanced and context-aware responses. To build trust and transparency with sellers, 11x developed an interactive 3D knowledge graph visualization. This innovative feature allows users to "peek into Alice’s brain," as Callaway described, by clicking on nodes representing data points and viewing the associated chunks of information. This visual representation reassures customers that Alice truly understands their business.
The development of the Knowledge Base was a transformative project for 11x, significantly enhancing Alice's capabilities and streamlining the user experience. Key lessons learned included the inherent complexity of RAG (Retrieval-Augmented Generation), the strategic advantage of prioritizing production deployment before extensive benchmarking, and the critical importance of leaning on specialized vendors. By embracing these principles, 11x has successfully moved Alice beyond being a mere "email-otron" to a truly intelligent digital worker capable of sophisticated, human-like sales engagement.
