I had the opportunity recently to ponder what Amazon Web Services (AWS) is doing with GraphRAG in association with its partner Lettria. In this post, I’ll  make a direct comparison of this AWS/Lettria approach with what Ernst & Young (EY) is doing with the help of partner Graphwise. 

Let me describe first what I understand AWS/Lettria is doing by characterizing what seems to be going on–a no-code, automated, one-shot, ingest-process-egest approach to knowledge management. Then I’ll contrast it with what EY has been doing with Graphwise (which leverages the value of an established, empowered KM department engaged in iterative, human-first, metadata development that also leverages generative AI, NLP and GraphRAG automation ). 

In general, I like what Denise Gosnell and Vivien de Saint Pern of AWS say in their piece “Improving Retrieval Augmented Generation accuracy with GraphRAG.” The authors make legitimate points about the value of Graph RAG versus vector–only RAG. 

For example: “By modeling data as a graph,” the authors write, “you capture more of the context and intent. This means your RAG application can access and interpret data in a way that aligns closely with human thought processes.” 

No disagreement there, or with Lettria’s assertion that GraphRAG can benchmark at least 30 percentage points more accurately than vector-only RAG. 

My disagreement is mainly with the implication that a third party and its GraphRAG automation alone are going to solve your organization’s AI readiness problems. 

Lettria’s four-step method implies that users can, with this approach, tell the system to hoover up the company’s data sources, and allow it to process the sources, and output the claimed AI-ready result.

“Step 1:

Upload your documents and data sources to Lettria’s platform.

Step 2:

Lettria does the heavy lifting of parsing, filtering and tracing knowledge. With graphs and nodes, everything has a verifiable source.

Step 3:

Review the emerging concepts and relationships. Verify or reject GenAI outputs, always owning answer accuracy.

Step 4:

Deploy your custom AI solution to your teams with confidence.”

Such magic bullet claims don’t sit well, particularly when it comes to difficult challenges like AI adoption, the necessity for quality, contextualized inputs, and the low level of most enterprise data/content/knowledge lifecycle maturity.

The reality is that the humans with the right skills and experience need to collaborate closely on the inputs to any AI system to ensure accuracy. At the same time,  proactive leadership needs to clear a path for process transformation whenever needed. 

Data quality from an AI standpoint implies desiloing and contextualization across sources that were typically created in closely guarded silos. The process is actually a knowledge creation and sharing process, with instance data at the lowest tier of abstraction.

Technology alone never solves an enterprise’s data quality problems. But this is a lesson that most industries not versed in doing their own data collection seem to find very difficult to learn. As the following case study underscores, an AI-enabled graphRAG approach to management of a company’s proprietary data, content and knowledge assets is important, but hardly sufficient to the task. 

EY’s incremental, evolutionary KM approach with Graphwise

By contrast with Lettria’s claims regarding its solution, Architecture and Governance Practice Lead Arup Vidyerthy of Ernst & Young (EY)’s co-presentation with Sales Enablement VP Helmut Nagy of Graphwise (“Connecting the Dots. Building a Collaborative Knowledge Hub with LLMs and Graphs”) at Knowledge Summit Dublin in 2025 was refreshingly straightforward and even humble. 

It’s evident that the internal KM group at EY has been central to a serious, long-term effort that builds on top of years of the knowledge management (KM) group’s work curating provenanced knowledge and building, developing and deploying centrally managed taxonomies.

To obtain accurate results, modelers working with the Graphwise platform have to work closely with domain experts, engage in give and take, build consensus and persuade business units to follow key development principles.  

Viderthy noted that because EY employs approximately 400,000 people, a high volume of content flows in from all the different industries EY serves. A large KM group reviews the incoming content. The challenge historically in such a sizeable effort is that some content gets misclassified or improperly tagged. 

Initial implementation of the Graphwise AI Graph Suite showed immediate improvement along these lines. Content uptake increased 50 to 60 percent, and search improved significantly. 

Viderthy acknowledged that the adoption of the platform is a work in progress, and anticipates further improvements as the taxonomy and ontology modeling efforts central to the AI-ready knowledge graph evolve.

The full scope of a human-led AI-ready data effort

I’m guessing that the KM folks spent much time over the years incorporating and aligning new content from acquisitions with existing content from established business groups, for example. 

In short, EY couldn’t take on a firmwide, full-blown knowledge graph project and succeed in building such a graph without sufficient data maturity and a culture that systematically collects and curates the data, content and knowledge that could be harnessed in the graph. 

In that sense, the KM organization made the adoption of knowledge graph-based AI possible. And Graphwise provided a platform designed to support the human-first, machine assisted efforts EY rightly favors alongside the firm’s other systems.