Why Context Matters in AI

Over the past few months we’ve been hearing some interesting things about how investors are thinking about AI, and one concern keeps coming up. If AI makes it so easy for a brand-new company to get started and do things better, faster, and with fewer people, how do you know you’re backing the right business? It’s a fair question, because the space really is getting more competitive as the barrier to doing things drops.

I’d argue the more useful question runs the other way: what can an existing company do that a brand new one can’t? Usually the answer is everything it has built up over the years: the expertise of its people, the data it has collected, the hard-won sense of what actually works in its market. In AI terms, all of that is context, and I’d put it near the top of the list of the most valuable things a business can own right now and as one of the strongest moats it can have.

Let’s look at an example of the difference expertise and context can make. Hyble, a client of ours, builds technology for creating drinks menus. Some of their customers had already found AI on their own: they were taking a prompt and putting it into ChatGPT to generate a menu. The output looked plausible, the way AI output usually does. But it was also generic, because although ChatGPT knows what a menu looks like, it knows nothing about what makes one effective and brand-compliant.

But, run the same prompt through the AI tooling we built with Hyble and you get a completely different menu: one that can be approved and put in front of customers immediately. Not because the prompt got better - it’s word for word the same. The difference is that the system answering it knows things ChatGPT doesn’t: the psychology of how menus sell, the governance and regulatory requirements a menu has to meet per geography and per brand, and all the unglamorous knowledge built on years of data showing what separates a decent-looking menu from one a business can actually put in front of customers.

What context actually is

Context is everything the model can see beyond your prompt: your documents, your definitions, your rules, your examples, your sources of truth. If the model is who you hire, context is the onboarding. A model with no onboarding doesn’t act like a person with no onboarding. A person says ‘I don’t know’. A model guesses, fluently, and confidently. If the output reads like it was written by someone who’s never met your organisation, that’s because, in every sense that matters, it was.

The document that doesn’t exist yet

Every organisation runs on a document that doesn’t exist yet: how you define your customers, what good looks like in your work, the rules you follow (including the ones that you definitely can’t break), the tone you use, and where the real numbers live. Almost all of it sits in people’s heads. Making AI useful means writing it down and connecting it, so the right knowledge gets fetched when it’s needed rather than manually rewritten into every prompt. The systems for that part are now standard. The scarce ingredient is the writing.

Most of this work isn’t technical. Deciding what’s true about how your organisation works is leadership work, and it pays back well beyond AI. It’s something we’ve seen time and time again: AI doesn’t fix weak foundations - it exposes them. Organisations that can’t articulate their own rules and definitions discover their AI can’t either. The ones that do the writing end up with a compounding asset, because every tool and every new starter, human or otherwise, is onboarded from the same source. And to come back to those investor conversations, that asset is exactly the thing a brand new competitor can’t have right away.

Pro tip: Let the AI do the writing down

Documenting how things work is almost nobody’s favourite task, which is exactly why AI should be doing more of it. It’s also hard to articulate your own mental model by yourself; the knowledge comes out far more easily in conversation. 

I’ve sat down with the people who hold the expertise in a business, recorded and transcribed the conversation, and asked the questions that forced them to articulate what they know out loud. The transcript then goes to AI, which turns what was said into the documentation and frameworks the business keeps. The AI isn’t inventing anything new, it’s just articulating it so it stays with the business, which is a much easier ask than expecting someone to sit down and write it all out by hand.

Once the model finally knows your business, something predictable happens: simply drafting stops being enough and you start wanting and trusting it to act: update the record, chase the invoice, move the work along. What has to be true before you can safely let it act is the next topic: the harness.

Part 4: So Why Do AI Initiatives Still Fail?