What Are Knowledge Agents?
We all know that AI is changing the world and adding new capabilities to our businesses and our lives. Most of us are already familiar with how it can create marketing content, do deep research for a holiday, or help navigate complex legal material. We also know that AI can achieve goals by creating multi-step processes — and that one day it may even make the tea and iron our shirts.
Knowledge agents, however, are much less well known.
That’s a shame, because they are one of the most practical and high-impact ways AI can make a real difference very quickly. They’re such a strong use case that I wanted to write this article to introduce you to them.
Why Data Warehouses Still Matter — and Where They Don’t
Over the last 20 years, we’ve built a lot of data warehouses — and we still do. They remain a valuable way to structure complex data for reporting, particularly when accuracy is mission critical. Booking data is a good example: when the answers drive financial or operational decisions, they need to be correct.
But creating a data warehouse is a labour of love.
You need to import data from multiple sources, understand those sources in detail, and transform them into structures suitable for reporting. That requires investigation, analysis, and careful data modelling. Uncovering the transformation rules takes time and implementing them takes effort. AI has helped shorten development cycles by assisting with code generation, but building a robust warehouse is still far from simple.
For me, the real opportunity for knowledge agents sits one level below this — with data that isn’t quite so mission critical.
That’s the sweet spot where knowledge agents can deliver information in near real time, at a fraction of the cost and development effort of traditional approaches.
Before we can properly talk about knowledge agents, though, we need to introduce an important concept: the MCP server.
MCP Servers: The API Layer for AI
In the early days of the web, most applications were “two-tier” systems — a set of web pages connected directly to a database. That architecture was fragile and difficult to maintain. Any change to the pages or the database could easily break the system.
The next major step forward was “three-tier” architecture, which introduced an API layer between the user interface and the database. APIs allowed systems to communicate in a controlled, predictable way and completely transformed how the internet worked. That model has remained largely unchanged for decades.
An MCP server is essentially the AI equivalent of an API — but designed for large language models.
MCP servers allow AI to retrieve data from a wide range of sources: APIs, web pages, file systems, databases, and more. They form a layer on top of your data that tells AI how to interact with it — how to extract data, what’s allowed, and what isn’t.
You can think of an MCP server as a wrapper that sits on top of multiple, very different data sources and makes them all accessible to AI in a consistent way.
Strictly speaking, you don’t need an MCP server for AI to access a data source. But having one makes it dramatically faster and safer, because the AI doesn’t need to infer how the data works. An MCP server is essentially a structured text document that tells the AI how to behave — a form of context engineering rather than traditional programming.
Why MCPs Matter Right Now
Over the last 18 months, companies have been creating MCP servers at a remarkable pace. Today, there’s a good chance an MCP server already exists for most of the data sources you care about. For the few that don’t, it’s usually only a few days’ work to create one.
MCPs have not yet emerged as the default mechanism for AI data access. In reality we have found that they are well suited to small solutions with a limited number of data sources, but their limitations become apparent as solution grow both in size and complexity. As larger implementations demand greater scalability, alternative approaches are likely to become the standard.
One recent approach is the use of skills, which may ultimately be used instead of the direct use of MCP servers for larger solutions. We will explore this shift in a future article. This rapid evolution is characteristic of the AI landscape, where best practices can change in a matter of weeks rather than years.
For the time being, however, MCP servers remain a practical and effective choice for building focused, small-scale AI solutions.
With that foundation in place, we’re now ready to talk about knowledge agents themselves.
So, What Is a knowledge Agent?
At its most basic level, a knowledge agent is created by pointing AI at the data sources you care about, giving it rules about how that data should be interpreted and used, and then allowing people to ask questions in plain language.
The result feels a lot like ChatGPT — but instead of drawing on the public internet, it’s grounded in your specific data, your business rules, and your objectives.
That’s what makes it so powerful.
Rather than building complex pipelines and reports up front, you allow AI to reason over your information on demand, using MCPs and context to guide its behaviour.
Practical Examples
The easiest way to understand the value of knowledge agents is to look at real-world use cases:
HR and policy support
Feed your staff handbook, HR policies, and relevant internal data into AI, and your employees can ask HR-related questions and get clear, consistent answers — without wading through pages of dense and arid documentation.Training and onboarding
Combine system documentation, training manuals, and usage guidelines to create a systems training agent that supports new staff during induction and beyond.Product or System FAQ
A knowledge agent can serve as a first-line support tool by answering common questions about a product or system using curated reference material.
Where Knowledge Agents Really Shine
Knowledge agents are ideal for problems where information needs to be available quickly and cheaply, but where the data isn’t critical enough to justify months of design and engineering effort.
They can be hardened over time into more robust solutions — but their real magic lies in speed. You can create something genuinely useful in days rather than months and immediately make people’s lives easier.
That’s the use case I’d recommend starting with.
Used thoughtfully, knowledge agents are one of the fastest ways to turn AI from an interesting experiment into something that delivers real, everyday value.
Just bear in mind that knowledge agents currently work well for small solutions. If you have large volumes of sources, then design and architecture of the solution will still be as critical as it ever was.
