Designing the way we offer context so AI can reason instead of guess

Article — Jan 12, 2026

By Piers Fawkes — 4 min read

Yesterday, I was on an informational call with an old friend from London who’s now in stealth mode for a fascinating healthcare startup. We were talking about real data and a regulatory environment where getting things wrong really matters.

As we talked through my suggestion that they consider knowledge graphs, I realized I kept coming back to the same four reasons knowledge graphs are so useful in practice.

None of them are especially theoretical but when I speak to people about Fodda, these are the themes I hear myself talk through:

1. Knowledge graphs reduce cognitive load on the LLM

A lot of AI systems today ask the model to do too much work. We give it raw data, loosely related documents, or embeddings, and expect it to infer what connects to what, what matters most, and which paths are valid. When it fails, we call it hallucination. In reality, it’s guessing because we’ve pushed all the reasoning into runtime.

What knowledge graphs do differently is shift that work upstream. By making relationships explicit in the data itself, the model doesn’t have to infer them on the fly. It doesn’t need to work out what’s related, what’s hierarchical, or what should be considered first. The structure already exists.

2. Knowledge graphs enable controlled disclosure

Another thing that came up in the conversation was data access. In the past, if someone wanted insight, the default was to hand over a CSV file or a full dataset, even if they only needed a handful of answers. That’s inefficient, risky, and commercially awkward.

With a graph-based approach, you don’t give access to everything. Your system returns precise answers to specific questions: the system queries the graph, retrieves only what’s relevant, and passes that to the model. The underlying data remains protected, and the output is easier to explain and audit.

3. Knowledge graphs encode expert rules without hard-coding logic

Instead of hard-coding rules into prompts or code, expert judgment is embedded in the structure of the data itself. What connects to what, and what matters more than something else, is decided upfront. The model doesn’t invent the logic. It follows it.

4. This is about designing the data correctly at the start

The last point is probably the most important. Once you start thinking this way, the focus shifts from “How do we train the model?” to “How do we design the data so the model doesn’t have to guess?” If the relationships are clear, the model can do what it’s good at: synthesizing, explaining, and interacting in natural language.

We no longer pre-train our AI systems and instead provide them with in-session and API-called data.

By connecting AI platforms (and enterprise versions of LLMs) to knowledge graphs like the ones Fodda offers - we remove mental workload from the AI systems to allow them to excel in other areas.