Business case for a production AI system
NDC KB Agent Executive Summary
NDC KB Agent Executive Summary
Every sales organization has the same problem: product knowledge is scattered across dozens of systems. SharePoint holds the official specs. Slack has the latest competitive intel. The wiki has outdated documentation that nobody maintains. The real answers live in the heads of three people who have been at the company for eight years.
When a prospect asks about pricing, integrations, or competitive positioning during a live call, your reps scramble through multiple systems. Responses are delayed — sometimes missing the moment entirely. The information they find may be outdated or contradictory. And there is no audit trail for the claims they make to customers.
The NDC KB Agent solves this by putting an AI-powered product knowledge assistant between your teams and your knowledge base. It provides instant, accurate, citation-backed answers about the complete product portfolio. Every response traces back to a specific source document. When sources conflict, the system knows which one to trust.
What It Does
The KB Agent serves three primary audiences: sales teams during live customer engagements, proposal writers drafting RFP responses, and customer success managers handling technical questions. For each audience, the system provides something different:
| Capability | Business Impact |
|---|---|
| Instant answers | Research time drops from minutes to seconds |
| Source citations | Every response cites specific documents |
| Authority hierarchy | Conflicts resolved by source trustworthiness |
| Multi-mode support | Tailored for sales calls, RFPs, and team handoffs |
| Integrity by design | Disputes trigger principled resolution, not hand-waving |
The workflow is straightforward: a user asks a question in natural language, the agent retrieves relevant documents from the knowledge base, ranks them by authority, generates a response with inline citations, tracks the claims it makes, and enables correction if something is wrong.
How It Differs from Generic AI Chatbots
Five architectural decisions separate the KB Agent from dropping your documents into a generic chat interface:
1. Domain-exclusive knowledge. The agent knows your products and nothing else. It cannot hallucinate features from other companies or invent capabilities based on general training data. If the information is not in the knowledge base, the agent says so.
2. Source citations on every claim. Every factual statement traces back to a specific document in the knowledge base. This is not optional. The system extracts claims from its own responses and logs the provenance chain for each one.
3. Authority-ranked conflict resolution. Enterprise knowledge bases inevitably contain contradictions. Marketing says one thing, the product catalog says another, last quarter's RFP template says a third. The KB Agent maintains an explicit authority hierarchy — product catalog overrides marketing, marketing overrides RFP templates — and applies it during every retrieval.
4. Conversation context across turns. The agent maintains session state, remembering what products have been discussed, what claims have been made, and what mode the conversation is operating in. Follow-up questions work naturally because the system carries context forward.
5. Graceful failure over confident errors. When the agent does not have enough information to answer confidently, it says so. When a user disputes a claim and the dispute cannot be resolved, the session ends rather than continuing with unverified information. This is a feature, not a limitation.
Dual-Agent Architecture
The system routes queries to one of two specialized agents:
The Lookup Agent handles fast fact retrieval — "What is HBA?", "What integrations does our analytics suite support?" It retrieves 3-5 documents, operates within a compact 2,000-4,000 token budget, and returns responses in 2-3 seconds. This is the workhorse for live sales calls where speed matters.
The Expert Agent handles complex synthesis — product comparisons, RFP content generation, competitive positioning, and multi-product analysis. It retrieves more documents, operates within a 4,000-8,000 token budget, and produces structured, citation-heavy outputs. It runs in three modes (RFP Specs, Agent Context, Sales Marketing), each optimized for a different output style.
An Intent Router examines each query and routes it to the appropriate agent. When the Lookup Agent encounters comparison keywords, synthesis requests, or multi-product-line queries, it escalates to the Expert automatically. The user never needs to know which agent is handling their question.
Trust and Safety
The 3-Branch Correction Protocol handles what happens when a user disputes a factual claim:
Branch A: The system cannot verify its claim. It acknowledges the error, corrects its understanding, and continues the conversation.
Branch B: The system finds the source document supporting its claim. It cites the source and invites the user to verify. If the user accepts, the conversation continues.
Branch C: The user rejects the citation. The session terminates and the dispute is flagged for human review.
Branch C is the most important design decision in the entire system. Terminating a session feels like a drastic response, but continuing a conversation where the system and user disagree about facts means every subsequent response is potentially wrong. In a context where product claims flow into customer commitments and RFP responses, propagating disputed information creates real liability. The system prioritizes factual integrity over conversational continuity.
Production Metrics
| Metric | Value |
|---|---|
| Automated tests | 159 |
| Lookup response time | 2-4 seconds |
| Expert response time | 4-8 seconds |
| Context window range | 2K-8K tokens |
| Knowledge graph nodes | 300-400 |
| Supported modes | 4 |
| Data retrieval latency | 80-150ms |
The system is production-ready with a REST API, OpenAPI documentation, real-time streaming responses, session persistence and export, comprehensive test coverage, and observability through logging, provenance tracking, and analytics endpoints.
Strategic Value
Near-Term
Faster sales cycles. Reps get instant answers during live calls instead of promising to follow up. The difference between answering a prospect's question in real-time and sending an email the next day can determine whether a deal moves forward.
Consistent messaging. Every rep has access to the same authoritative content, ranked by the same authority hierarchy. The new hire and the ten-year veteran cite the same sources.
RFP acceleration. Draft responses in minutes instead of hours. The Expert Agent's RFP mode generates formal, citation-heavy content that proposal writers can review and refine rather than create from scratch.
Faster onboarding. New hires ramp faster because they can ask the KB Agent the questions they would otherwise ask colleagues. The agent never gets frustrated, never gives a different answer depending on its mood, and always cites its sources.
Long-Term
The KB Agent establishes infrastructure for capabilities that extend well beyond question-answering:
Knowledge governance. A single source of truth with version control, authority ranking, and change tracking. When a product claim changes, it changes everywhere.
Competitive intelligence. The knowledge graph captures competitive positioning alongside product capabilities. Over time, you build a structured view of how your products are positioned against every competitor in every market segment.
Analytics insights. Every query is logged. Over months, you build a picture of what customers ask most frequently, where documentation gaps exist, and which products generate the most complex questions.
Continuous improvement. Logged disputes from Branch C of the correction protocol identify exactly where documentation is unclear, contradictory, or missing. Each dispute is a signal pointing at a specific documentation improvement.
Getting Started
The technical setup is minimal: a FastAPI server, a ChromaDB vector store, and Claude for response generation. The system processes queries through a REST API or streaming endpoint, supports four operating modes, and maintains session state across conversations.
The harder work is not technical — it is curatorial. The KB Agent is only as good as the knowledge base it draws from. The authority hierarchy needs to be defined. Documents need to be ingested and tagged with appropriate metadata. The knowledge graph needs to be populated with product relationships, buyer contexts, and competitive positioning. That investment in knowledge curation is what transforms a chatbot into a trusted product knowledge system.