# Anchoring Enterprise GenAI with Knowledge Graphs

Jonathan Lowe, Pfizer & Stephen Chin, Neo4j | AI Engineer Summit 2025 | 20:59

Source: https://www.youtube.com/watch?v=OpVkWc3YnFc
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/anchoring-enterprise-genai-with-knowledge-graphs
Published: 2025-04-07
Tags: embeddings, enterprise, rag, team-adoption

## TL;DR
- A real business problem, such as speeding technology transfer in biopharma, gives an enterprise GenAI project a reason to exist beyond experimenting with models.
- Knowledge graphs can connect document chunks and organizational data so GraphRAG returns more precise, explainable context than direct LLM use or baseline vector search.
- Enterprise adoption depends on adapting the proposal to executives, business leaders, client partners, vendors, and colleagues who may all have different goals.

## Summary
Jonathan Lowe describes a Pfizer technology-transfer project that uses GenAI to help move drug development from laboratory scale to factories. The work involves millions of documents, notes, and test outcomes, while manufacturing expertise is becoming harder to retain as average employee tenure falls. The team stored document chunks in a knowledge graph and refined its chunking by examining which structures produced useful search results. Stephen Chin and Lowe connect this technical approach to the organizational work required to ship enterprise systems. Lowe explains how an idea must be tied to executive goals, department-level measures, and user needs such as accuracy and speed. Chin explains GraphRAG as a combination of vector retrieval and graph context. The graph adds relationships, access controls, and explanations to the retrieved information. Lowe also argues that the graph improved how quickly technical teams could understand and prepare data for new projects.

## Key ideas
### A concrete business problem gives enterprise GenAI a reason to ship
[02:20](https://www.youtube.com/watch?v=OpVkWc3YnFc&t=140s)
Lowe frames the project around technology transfer in biopharma. Teams must move from laboratory work to industrial production across multiple factories, and that process can take years because people have to examine hundreds of thousands of documents, notes, and test outcomes. The project aims to help new manufacturing staff recover knowledge that is stored in documents and in experienced workers' heads. Chin connects this example to Gartner's reported failure mode: projects often lack a business use case that solves a real problem and can be monetized. In this case, faster transfer of drug-making knowledge could help get life-saving drugs to people sooner.

### Document chunking becomes more useful when its structure is stored in the graph
[04:42](https://www.youtube.com/watch?v=OpVkWc3YnFc&t=282s)
The team loaded chunks of documents into a graph instead of treating each document as an undifferentiated object. They represented levels such as the document, block, paragraph, and line. This structure let them examine which chunks returned the results that users wanted when searched with similarity search. Lowe says the team used what it learned to improve how documents were chunked in the first place. The graph therefore held information about the source and structure of retrieved text, giving the team a way to refine ingestion rather than accepting one fixed chunking strategy.

### Enterprise adoption requires a proposal for each layer of the organization
[07:30](https://www.youtube.com/watch?v=OpVkWc3YnFc&t=450s)
Lowe describes an idea moving through a large company with more than 50,000 or 100,000 people. Users care whether a tool removes boring work, produces accurate results, and responds fast enough for practical use. Senior executives communicate broad goals, such as leading an industry or accelerating supply. Their direct reports translate those goals into cost savings, cost avoidance, earlier revenue, or balanced headcount. Client partners may apply yet another department-specific view. Lowe says proposals need numbers, times, and promises that match the people being addressed. The same project may otherwise be reduced to a small integration or expanded into an impossible company-wide scope.

### Internal politics can block an AI project after the technical work begins
[13:48](https://www.youtube.com/watch?v=OpVkWc3YnFc&t=828s)
Lowe lists resistance that can appear even after an idea has found users and executive support. Vendors may argue that the company should buy rather than build. Colleagues may claim that AI search is already their area or ask for integration with tools they have built. Client partners in different departments may disagree about whether the company needs another search engine or one shared capability in every tool. His advice is to know the audience, personalize the message, and speak to each group in the language and at the level it uses.

### Graphs improve team understanding as well as data retrieval
[16:47](https://www.youtube.com/watch?v=OpVkWc3YnFc&t=1007s)
Lowe says Pfizer used graphs for data types that included genealogic sequences, recipes, social networks, hierarchies, and time series. He also found a benefit outside query traversal and search performance. As more data was consolidated in the graph, data scientists, engineers, developers, and SREs could understand the data landscape faster. Work that had taken three months to consolidate, understand, and clean up took three weeks or less for a new project. The graph gave the team a shared structure for seeing how data fit together.

### GraphRAG combines vector similarity with connected context
[18:10](https://www.youtube.com/watch?v=OpVkWc3YnFc&t=1090s)
Chin explains GraphRAG as an architecture with both vector and knowledge-graph representations of data. The application can ask the vector store for relevant material, then retrieve relationally close nodes from the graph database and pass that extra context to the language model. Direct LLM use can lack enterprise context, while baseline retrieval with a vector database can produce answers that remain generic and hallucinate. Graph context adds relationships that may not have been made explicit through joins in a relational database. The approach can also evolve as the knowledge graph grows.

### Graph structure supports access control and explainability
[19:31](https://www.youtube.com/watch?v=OpVkWc3YnFc&t=1171s)
Chin says graph nodes can carry controls and properties that govern who can access information. The retrieved answer can then be inspected through nodes and edges rather than only through statistical probabilities in vector space. People can reason about why information was included and which concepts relate to a topic such as manufacturing. This matters for life sciences, manufacturing, and other business-critical settings where an incorrect answer has serious consequences. The graph gives the application a way to expose relationships behind the context sent to the language model.

## Notable quotes
- "We all want to be successful with Gen, we all want to do amazing things, we're getting asked to do amazing things, but we need to have the right way of approaching this in our organizations with leadership to sell this internally." (01:20)
- "I love building applications that delight the people that use them." (08:32)
- "What used to take 3 months to consolidate, understand, clean up took three weeks or less for a new project." (16:50)
- "You can get better results where now it's actually pulling in organizational knowledge, but the answers tend to be a little bit generic." (17:52)
- "You can get better explainability now because when you're getting an answer from the LLM you're no longer looking at statistical probabilities in the vector space, you're actually looking at graphs and nodes and edges which we can reason about." (19:50)

## Tools & references mentioned
- Gartner
- Microsoft
- Neo4j
- Pfizer
- Dirty Dancing
- IBM

## Who should watch
- You are trying to move a GenAI prototype into production inside a large enterprise and need support from groups with different measures of success.
- Your application searches sensitive, connected business data and needs more context, access control, or explanation than a vector-only system provides.
- You are building a knowledge graph and want to understand how document chunking can be refined through search behavior.

## Related talks

- [Knowledge Graphs & GraphRAG: Techniques for Building Effective GenAI Applications](https://aietalks.com/talks/knowledge-graphs-graphrag-techniques-for-building-effective-genai-applications) (Zach Blumenthal & Andreas, Neo4j, 1:39:52)
- [Intro to GraphRAG](https://aietalks.com/talks/intro-to-graphrag) (Zach Blumenfeld, Neo4j, 1:18:35)
- [Practical GraphRAG: Making LLMs Smarter with Knowledge Graphs](https://aietalks.com/talks/practical-graphrag-making-llms-smarter-with-knowledge-graphs) (Michael Hunger, Jesús Barrasa & Stephen Chin, Neo4j, 19:46)
- [GraphRAG: The Marriage of Knowledge Graphs and RAG](https://aietalks.com/talks/graphrag-the-marriage-of-knowledge-graphs-and-rag) (Emil Eifrem, Neo4j, 19:15)
- [Agentic GraphRAG: Simplifying Retrieval Across Structured and Unstructured Data](https://aietalks.com/talks/agentic-graphrag-simplifying-retrieval-across-structured-and-unstructured-data) (Zach Blumenfeld, Neo4j, 15:25)