Having implemented a multitude of business intelligence systems since the mid-90s, I cannot shake the feeling that, with the current AI revolution, we are living through a high-definition remake of the early business intelligence era.
Back in the 90s, the arrival of self-service reporting tools and desktop OLAP engines felt like magic to those of us who preferred data-driven decision-making. Suddenly, managers could conjure visually stunning dashboards from a chaotic soup of data sources. These tools birthed the corporate dashboard, created the role of the modern business analyst, and fundamentally altered corporate decision-making.
However, the introduction of these new tools did not happen without the inevitable hurdles. Executives quickly realised that a sexy report does not guarantee accurate data. Boardroom meetings devolved into debates over whose “truth” was correct, sparking a desperate need for data integrity.
Data warehouses
The solution was the data warehouse—a robust and well architected data backend which Ralph Kimball defined as “collection of data marts with conformed dimensions.” While one might argue over the best underlying architecture—Data Vaults (Dan Linstedt) versus the Operational Data Store (Bill Inmon) versus Ralph Kimball’s Dimensional Modelled Data Warehouse method—the business reality was clear: without a well-defined backend, your reports and analysis were a stand alone analysis without corporate foundation.
The concept of the data warehouse continued to evolve as the world became more digital. Data lakes (and data swamps) emerged and were merged with the data warehouse concept to form the data lakehouse. The concept however remained the same: one single trusted, documented, source to build your reports and analytics upon.
New Century, Same Problems
Fast forward to today. AI agents are the new “shiny” objects, offering even more compelling presentations than BI tools, by communicating in human-like language, on top of a great presentation. But beneath the sophisticated prose, the credibility of AI hinges entirely on its data sources.
Enter RAG (Retrieval-Augmented Generation) and its more sophisticated sibling, Graph RAG.
GraphRAG (Graph Retrieval-Augmented Generation) is an advanced AI framework that enhances the accuracy and context-awareness of Large Language Models (LLMs) by incorporating knowledge graphs into the retrieval process. Unlike traditional RAG, which relies on searching for similar text chunks (vector search), GraphRAG maps relationships between entities (nodes) and their connections (edges) to provide a deeper, more structured understanding of data.
Introduced by Microsoft in 2024, it was designed to overcome the limitations of traditional RAG, particularly when dealing with complex, multi-hop queries over large, unstructured dataset
In essence, GraphRAG requires the latest equivalent of a “data warehouse” behind the RAG. Whereas data warehouse were traditionally designed for structured data (“measurable” data that you can summarize, average, … ), knowledge graphs are designed to capture “unstructured” data and relationships between complex and hard to quantify concepts. They are the tool that provides the context and “ground truth” to LLMs – similar to data warehouses being the foundation for your corporate dashboards.
The High Price of “Good Enough”.
The rub, of course, is that building a reliable backend requires substantial investment. In the 90s, data warehouse projects spanned years and cost millions. Many companies who thought to avoid that cost by building isolated reports and stand alone data marts ended up with spaghetti data architecture, contradicting reports/dashboards and hampered decisions making.
In a similar way, many of today’s CIOs still suffer from a “zero-sum delusion,” believing they can achieve quality AI outputs without the “hassle” of a solid backend. They are mistaken. Without the right foundation:
- Complexity scales faster than value: Reports and AI outputs become increasingly chaotic and costly to maintain.
- Trust evaporates: Just as inaccurate BI reports became an areas for suspicion and discussion, so will ungrounded AI conclusions become a topic of debate.
- Hidden costs skyrocket: Whilst in the short run, building an stand alone agent for 1 single purpose might be a cheap project, in the future a multitude of agents will benefit from having access to the same contextual background in your knowledge graph. The cost improvements will not just come from code and infrastructure reusability but also from the reduced CPU cycles that a will build knowledge graph offers LLMs
History Doesn’t Repeat, But It Rhymes.
New technology exists, but we are in danger of forgetting the parallels with the past. Combining graph databases with LLMs creates an intelligence infrastructure that actually works. The lesson is simple: do it the right way from the start.
We must realise that well-presented information is not the same as factually correct information. Shiny AI interfaces are useless, and the answers of LLMs cannot be trusted without the “boring” work of a structured backend.
The investment in that backend is significant, but it pays off. It results in accurate outcomes that get cheaper to build over time, leading to the holy grail of business: trusted information and faster decision-making.
Taking the AI backend challenge seriously is the only way to make your AI powered business strategies actually stand out.
Leave a Reply