Most of the leaders I talk to think they have an AI problem. They don't. They have a context problem — and until they see the difference, no model upgrade, no better prompt, and no new vendor is going to fix it.
Here's what I mean. You rolled out Copilot, or ChatGPT Enterprise, or an internal assistant six months ago. The vendor promised productivity gains. Your team got training. The pilot looked fine. But now, six months in, adoption is uneven, the output is generic, and you're starting to wonder if you bought the wrong tool.
You didn't. The tool is fine. What's missing is context — the organisational knowledge that a human in the same seat would have absorbed in week one. Your AI can't see the conversations that shaped last quarter's decision. It doesn't know which customer has a renewal coming up or which internal process changed three months ago. It's working with a fraction of what it needs to be useful, and you're evaluating it as if it had the whole picture.
That's the context gap. And it's not a technical problem. It's an organisational one.
Why Most AI Rollouts Underperform
The research on this is consistent. Fewer than 40% of automation initiatives deliver measurable value, and only about 30% of AI pilots scale beyond the initial team. The usual explanation is that the technology isn't ready, or that people aren't using it right. But when you dig into why pilots stall, the pattern is organisational: the AI doesn't have access to the information it needs to do the work it's being asked to do.
Think about what happens when you hire someone new. They don't start contributing on day one. They spend weeks learning how things work — who owns what, where decisions get made, which systems hold which data, what the unwritten rules are. They ask questions. They sit in on meetings. They read old emails and project documents. By month three, they're useful because they've built a working model of how your organisation operates.
Your AI didn't get that onboarding. It can generate text, summarise documents, and answer direct questions — but it can't see the sprawl of information that makes those answers accurate or relevant in your context. Customer data lives in one system. Project notes live in another. Decisions get made in Slack threads that no one searches. Knowledge lives in people's heads. The organisational structure that determines who needs to know what isn't mapped anywhere the AI can access.
So when your AI produces generic output, or misses something obvious, or gives an answer that's technically correct but operationally useless — that's not the model failing. That's the model working with incomplete information.
The Five Context Layers Your AI Can't See
The context gap shows up in five predictable places. I'm naming them here so you can recognise which ones are weak in your organisation — but diagnosing them properly means asking harder questions than a blog post can hold. That's what the checklist at the end is for.
1. Conversational context. This is the history of discussions, decisions, and clarifications that happened in meetings, Slack channels, email threads, and side conversations. A human who's been in the room for six months knows what was tried last quarter and why it didn't work. Your AI doesn't — unless those conversations are searchable, structured, and accessible.
2. Structured data context. This is the information living in your CRM, your project management system, your finance software, and any other tool that holds records. Your AI can't pull customer history from Salesforce or project status from ClickUp unless you've explicitly connected those systems — and even then, it can only see what the integration allows.
3. Process context. This is how work actually gets done: the sequence of steps, the handoffs, the approvals, the exceptions. A human learns this by watching and asking. Your AI learns it only if someone has documented it in a way the AI can parse — which, in most organisations, means it doesn't learn it at all.
4. Knowledge context. This is the accumulated expertise that lives in wikis, internal documentation, onboarding guides, and — most often — people's heads. If your senior ops lead knows that a specific client always wants their reports formatted a certain way, that's knowledge context. If it's not written down somewhere your AI can access, it's invisible.
5. Organisational context. This is the structure of who does what, who reports to whom, and who needs to be consulted on which decisions. It's the difference between "generate a summary of this meeting" and "generate a summary of this meeting and send it to the three people who need to sign off on the budget."
Here's the uncomfortable part: most organisations are weak in at least three of these layers. And if your AI can't see them, you're not evaluating AI performance — you're evaluating AI performance with one hand tied behind its back.
What This Means for How You Evaluate AI
The industry has started to recognise this problem. Slack's Model Context Protocol is one response — an open standard for giving AI agents structured access to the context they need to do useful work. But a protocol doesn't fix the underlying issue, which is that most organisations don't know where their context lives, who owns it, or how fragmented it is.
That's the real work. Before you can give your AI access to better context, you need to map where that context lives. Which systems hold which data? Which conversations matter? Which processes are documented, and which ones live in someone's head? Who owns the knowledge that would make your AI useful instead of decorative?
This isn't glamorous work. It's not a model upgrade or a prompt workshop. It's the organisational groundwork that determines whether AI adoption works or stalls — and it's the part most vendors won't do for you, because it requires knowing your organisation from the inside.
I'll be honest about what this costs. Mapping your context layers takes time. It surfaces gaps you didn't know you had. It forces harder conversations about who owns what and where accountability lives. And it doesn't produce a demo you can show the board next quarter.
But the alternative is what you're living with now: an AI that underperforms because it's working blind, and a team that's losing faith in a tool that was never given a fair shot.
Download the Checklist
If you're reading this and recognising your organisation in at least two of those five layers, the next step is diagnostic. I've built a self-assessment checklist — Is Your AI Context-Ready? — that walks you through each layer and helps you score where your gaps are.
It's not a vendor pitch. It's not a sales funnel. It's a tool to help you see what your AI can't see — so you can decide what to fix first.
Download the checklist here. You'll also get Aurora Brief — my monthly briefing on what I'm learning, building, and observing — in your inbox once a month while you're at it. No spam, no upsell, just the work.
And if you finish the checklist and realise the gaps are bigger than you thought — or you're not sure where to start — that's a conversation Northlight can help with. But start with the checklist first. You need to see the problem clearly before you can solve it.