From: aidotengineer
The implementation of AI projects, particularly with generative AI (GenAI), presents significant challenges within enterprises. Gartner predicted that 30% of GenAI projects would be abandoned by the end of 2025 [00:00:58]. A primary reason for this failure rate is the lack of a clear business use case that solves real problems and can be monetized [00:05:59]. Furthermore, many projects struggle to reach production [00:01:26].
Case Study: Life Sciences Technology Transfer
A successful application of GenAI, leveraging knowledge graphs, was developed in a large life sciences company to address “technology transfer” [00:03:13].
The Business Challenge
In biopharma, scaling drug development from lab bench to industrial scale (making millions of doses daily across multiple factories) typically takes years [00:03:17]. This lengthy process is due to industrial teams needing to sift through hundreds of thousands of scientific documents, notes, and test outcomes [00:03:41].
An additional challenge is the significant decline in manufacturing worker tenure, from an average of 20 years in 2019 to just three years currently [00:04:00]. This means a vast amount of institutional expertise is retiring or will soon retire, highlighting the urgent need for GenAI to capture and transfer intelligence from documents and individuals’ minds to new employees [00:04:28].
The Solution: Graph-based GenAI
To address this, millions of documents were loaded into a graph [00:04:51]. Rather than loading entire documents, specific “chunks” (documents, blocks, paragraphs, lines) were loaded and structured within the graph [00:05:02]. This structuring allowed for the refinement of how chunks were stored and managed, improving the accuracy of similarity searches and eventually leading to better document chunking over time [00:05:20].
Impact
This solution not only serves as a strong business use case but also potentially saves lives by accelerating the delivery of life-saving drugs [00:06:07].
Navigating Organizational Challenges in AI Adoption
Integrating AI into business operations within large organizations (50,000+ employees) presents significant human and financial challenges [00:06:21].
Common Obstacles
- “Not Invented Here” Syndrome: Resistance from existing teams who may prefer other platforms or frameworks [00:06:26].
- Cost: GenAI architectures can be significantly more expensive than classic or cloud computing if not well-architected [00:06:50]. Convincing stakeholders to invest in an R&D-heavy, potentially more expensive system is difficult [00:07:04].
- Organizational Hierarchy: Navigating the varied perspectives of different management levels is crucial [00:11:02]:
- Top Executives (CEO): Focus on high-level, aspirational goals (e.g., “change a billion lives a year”) [00:11:05].
- Mid-Level Leaders (CDO, CSO, CPO): Translate executive vision into departmental objectives (e.g., “lead the industry in AI,” “accelerate supply”) [00:11:36].
- Operational Managers (Level Twos/Threes): Require concrete promises of cost savings, cost avoidance, earlier revenue, or balanced headcount, backed by specific numbers and timelines [00:12:15].
- Client Partners: Can either limit project scope (“Why build another search engine?”) or expand it to an unmanageable degree (“Incorporate into every tool in the organization”) [00:13:09].
- Vendors: Introduce “build versus buy” dilemmas, influencing leadership on economic realism [00:13:56].
- “Friendly Fire”: Internal colleagues may claim turf, or demand integration with their existing systems [00:14:27].
Recommendations
To succeed in such environments, it is essential to “know your audience,” personalize your message, and speak the appropriate language at each level of the organization [00:15:12].
Technical Advantages of Knowledge Graphs for Enterprise AI
A key technical challenge in AI in enterprise applications, especially with Retrieval Augmented Generation (RAG), is mitigating hallucinations from Large Language Models (LLMs) [00:15:58]. While vector databases improve results by pulling organizational knowledge, answers can still be generic and prone to hallucinations [00:18:02].
Why Choose a Graph Database?
Graph databases excel at representing complex relationships like genealogical sequences, recipes, social networks, hierarchies, and time series [00:16:29]. Beyond traditional benefits like easier traversal and improved search performance, consolidating data in a graph significantly boosts team performance. Data scientists and engineers can understand the data landscape much faster—from three months to three weeks or less for new projects [00:16:50]. In a graph, relationships are inherently present, allowing for the discovery of related information (“neighborhood of related stuff”) when searching [00:18:38].
Graph RAG for Enhanced Precision
Graph RAG (Retrieval Augmented Generation) represents an advanced approach to enterprise AI. Microsoft’s seminal paper on Graph RAG demonstrated superior results by chunking existing documents into a graph for LLM use [00:17:38].
The architecture for Graph RAG combines both vector and knowledge graph representations of data [00:19:06]:
- The vector database provides an initial answer [00:19:18].
- The graph database provides relationally close nodes, offering additional context to the LLM [00:19:21].
- This results in more contextually relevant and precise answers from the expert system [00:19:31].
Benefits of Graph RAG
- Superior Results: Achieves more precise answers compared to direct LLMs or baseline RAG with vector databases [00:18:13].
- Better Governance: Allows for controls and properties on graph nodes to manage access to information [00:19:47].
- Improved Explainability: Instead of relying on statistical probabilities from vector space, answers are derived from graph nodes and edges, which are easier for humans to reason about and understand relationships [00:19:54].
- Critical for High-Stakes Industries: Essential for business-critical sectors like life sciences and manufacturing, where accuracy is paramount and errors are unacceptable [00:18:29].