Many hopes and dreams are being fueled by the wonders and hype of AI systems. Yet those of us that have been through these technology-hype cycles before recognize the inevitable. AI technologies will help in some areas and fail in others. Just like the bygone eras of Big Data and BI, AI technologies will find their place in the pantheon of useful tools and systems. We will learn how to use and apply these new tools and techniques, just as we have with previous generations of wondrous technologies.
And one thing that we learned in previous generations is just as apparent today. Garbage In leads to Garbage Out (GIGO). If you don’t feed BI with the right data there is no way it can provide useful insight. If we think that just collecting enough data will lead to insight – well, many Big Data systems founded on that approach have struggled to achieve the results they promised.
As we look at AI systems today, we see that the rapidly evolving AI tools and techniques show that producing relevant, useful answers is not so simple. We have:
- LLMs – Large Language Models – that are great at synthesizing information – but expensive to train (and was that the most relevant data?), fairly static in their source data, difficult to keep focused, and questionable privacy.
- RAG – Retrieval Augmented Generation – Incorporates more domain specific and localized information, easier to keep more current; but still can expose confidential information, requires ongoing maintenance, and may not support sufficiently relevant answers.
- MCP – Model Context Protocol – direct interactions with other systems – can support near-real-time information and supports automated actions; but security and privacy are still being worked out, something has to figure out what MCP tool to invoke with which parameters for what purpose, and when.
- Agents – Agentic systems are the current darling – they can choreograph activities, give your AI applications memory, decide which RAG or MCP or LLM systems to call; but the mechanisms to make these decisions in a useful way is still emerging, there are Agentic Frameworks that support building applications and evaluating their behaviour is an additional challenge.
As you can see, there are a lot of pieces to fit together to build a useful AI system. But wait, what did we learn from bygone eras? We learned that if you don’t feed the right data in, we can’t trust what we get out. So even with this new technology, we still need to care about the data we feed the LLMs and RAG systems. Is it relevant and trusted? Is the data quality appropriate? What MCP Tools (functions) should we use and how do we know what data to ask for? How do agents know what information sources to use and how? Do all users get the same information? Does it matter if the user is an Accountant or a Sales Manager?
Addressing these issues requires matching the context and intent of the user with the context of the data sources and reasoning.
Context is a complex topic. Pragmatic Data Research and the Egeria open source project are actively researching and developing new capabilities to capture, manage, and deliver context to AI systems and monitor its effects. We call this Context Intelligence.
Context Intelligence augments the latest buzzword in the AI lexicon, Context Engineering. Context Engineering incorporates the user’s perspective, organization and needs. Context Intelligence adds our understanding of existing systems, that drive a business and organizations. For example, the AI system should know that if we want this kind of sales data it should get it from System A and if it wants a different kind of sales data it needs to call System B. And by the way System C is outdated and used as an archive only.
Context Intelligence builds on the best practices in information management, and focuses attention on the most relevant information to capture, link, manage and integrate as a component of the AI and analytics ecosystems; details of which we will explore in future blogs.