According to AI itself (e.g. ChatGPT) a Large Language Model, or LLM: “is a type of machine learning model trained on vast amounts of text data to generate human-like text. These models (like GPT-3, GPT-4, etc.) are capable of understanding and generating text based on the input they receive. LLMs are trained using massive datasets from diverse sources, such as books, articles, websites, etc.

The complications of using these models are (amongst others):

• Accuracy and consistency, or trustworthiness of answers.
• Answers are not stored in your systems to apply elsewhere.
• Using LLMs in the cloud requires running them over and over, pushing up costs.

How to increase the trustworthiness of AI-generated information?

The answer is: Use your knowledge and experience as input. Instead of trusting LLMs to use the entire World Wide Web of content, we want AI to use more demarcated information—specifically, information owned by the business and its domain of knowledge. When applying this demarcated information, we can substantially improve the performance of AI systems.

Image 1: The funnel for AI

In other words, we want AI to reason like us, using our experience and, therefore, our knowledge. That is why we use external data sources as ‘guard rails’.

Image 2: Adding a body of knowledge for AI to work with

How to structure your guard rails?

External sources can be anything, but we can narrow it down to structured and logical rules we want any LLM to consider before coming up with an answer. For this purpose, we use ontologies: structured information described in nodes and relationships, like a web.

Image 3: Guardrails for AI stored as a graph

ChatGPT itself says: “Ontologies help AI systems to understand the relationships between different entities and concepts in a domain, such as classes, subclasses, and properties. This makes it easier for such systems to reason, infer new facts, and make decisions based on well-defined knowledge.”

How to process AI-generated information for future use?

It is unnecessary to keep running the LLM over and over if you are using it to retrieve information for a follow-up process. But a generic AI chatbot, like the one for ChatGPT, isn’t instructed to save the information for later use. Let alone store it in an interoperable way, to use in any application you like. 

That is why we add a step of persisting the answer our AI comes up with—and not just in any format but in an Open Standard for sharing information via the Internet. This way, the answer can be used by any application anywhere in the world, by any part of your organization.

Image 4: The result of AI in a reusable (and interoperable) form

AI under control: The need for an Architecture, Framework, and Meaning of Data.

As Erik Roth, a McKinsey senior partner and global leader of growth and innovation, advised a CEO who was concerned their data wasn’t up-to-par for AI: “Yes, you need a data architecture. You need a data framework for thinking about how to tag and label things. [this would be our ontologies] …You need a bit of a curation process. Data is one of the things required for this type of technology to be useful. And getting your data in order is increasingly important if you want to take advantage of these technologies.” (see ‘What McKinsey learned while creating its generative AI platform’, November 25, 2024 | Podcast)