You are in a meeting. Not a particularly important one. A quarterly review, maybe. Or a client check in. The kind that fills the middle of your week and rarely produces anything that was not already known by the people in the room.
But someone is there who should not need to be. Not because they are unwelcome. Because they are too valuable to be spending an hour on this.
You know who they are. Every organisation has them. The person who carries the client history in their head. Who remembers what was promised three years ago and what was quietly dropped. Who knows that the decision maker’s real concern is not what they put in the email but what they said in the corridor afterwards. Who can feel when something is going wrong before the data shows it.
They are in the room because nobody else has what they have. Not their opinion. Their context.
Remove that person and the conversation becomes generic. The slides still work. The data still appears. But the intelligence leaves with them.
Now hold that image. We are going to come back to it.
The great convergence
If you read Parts One and Two of this series, you have seen how four decades of human work built the foundation for the intelligence era, and how the architecture of computing is inverting around it. The models, the orchestration, the interfaces: each layer of the stack is either commoditising or going open.
The frontier models are now so close in capability that the smartest builders treat them as interchangeable utilities. MCP, the protocol that defines how agents connect to tools, has been donated to the Linux Foundation. Interfaces are being rebuilt over weekends.
So if everyone has access to the same intelligence, the same plumbing, and the same tools, where does competitive advantage actually live?
There is only one answer. And it is the one layer most organisations have barely thought about.
Context.
Not data. Data tells you what happened. Context tells you what it means. Context is the difference between knowing a client spent two million pounds last year and knowing they nearly left six months ago, that the relationship was saved by a conversation your CEO had at dinner in Milan, and that their new procurement lead has a completely different set of priorities to the person you have been working with for a decade.
No foundation model knows that. No foundation model ever will, because that knowledge was created by your people, through your relationships, across your specific history.
That is the context layer. The only layer in the intelligence stack that cannot be bought, downloaded, open sourced, or generated from scratch.
This is the moat. The last one standing.
The floor is rising
Here is the part that should sharpen the mind.
The foundation models are getting smarter every month. Not incrementally. Substantially. Industry expertise, market dynamics, regulatory frameworks, best practices: the models are absorbing all of it from publicly available information.
Two years ago, a senior person with twenty years of experience had knowledge that was miles ahead of any AI system. That gap is compressing. Not because your people know less. Because the models know more. Every quarter, the baseline rises.
The industry analysis your team spent weeks producing can now be generated in minutes by anyone with a subscription. The only knowledge that gives you a lasting advantage is the knowledge the models cannot acquire from public data. Your proprietary context. Your institutional memory. The things that are true about your organisation and no one else’s.
If you are not actively identifying and structuring that layer, your advantage is eroding with every model update. The floor is rising to meet you.
The context advantage has a shelf life. And the expiry date is closer than most people think.
The thing nobody wants to talk about
Building a context layer sounds like a technology project. It is not.
The most valuable context in any organisation does not live in databases. It lives in people. And for decades, holding that knowledge gave people standing. Knowledge was scarce. Access to knowledge was power. The person who understood the legacy system had job security. The person who held the client relationships had leverage.
Now imagine asking those people to make their knowledge accessible to a system. To externalise the very thing that made them indispensable.
Of course there will be resistance. You are asking them to change the terms of a contract they have operated under for their entire career. If you cannot show them what replaces the old form of value, you will get compliance at best and quiet sabotage at worst.
Here is the reframe that matters.
You bring people into meetings because you need their context. But what you actually need is their thinking.
These two things have been bundled together for so long that most organisations cannot tell them apart. The person who remembers what happened is the same person who knows what to do about it. The context and the wisdom have always arrived in the same package.
The context layer separates them. For the first time.
When the system holds the information, the human is freed to do the thing that was always more valuable: apply judgment. Exercise taste. Make the call that requires wisdom rather than recall.
The person who used to be invited because they remembered is now invited because they think. The account director who held the relationship history becomes the person who decides what the relationship should become.
The context layer does not diminish those people. It reveals what was always the more valuable part of their contribution. The knowledge was never the real asset. The thinking was.
The flywheel
Context compounds.
Every day an organisation spends structuring and connecting its proprietary context, its intelligence gets sharper. Sharper intelligence produces better outcomes. Better outcomes generate new context. That context feeds back into the system and makes it sharper still.
This is a flywheel. And flywheels have a specific property that matters enormously: the longer they spin, the harder they are for anyone else to start.
An organisation that began building its context layer twelve months ago is not just twelve months ahead. It is compounding months ahead. A competitor starting today does not face a linear gap. It faces one that has been widening every single day and will continue to widen from here.
This is why context is not just an advantage. It is the advantage. The one that, once established, becomes nearly impossible to replicate. Not because the technology is proprietary. Because the context is yours, and rebuilding it from scratch would require living your organisation’s history again.
The hardest problem in the stack
Now here is where I owe you some honesty.
Everything I have described might make it sound like building a context layer is a strategic decision followed by a straightforward implementation. It is not. It is the single hardest unsolved problem in the entire intelligence stack.
The obvious counterargument: context windows are getting enormous. Some models now handle a million tokens or more. So why not just feed everything in and let the model figure out what matters?
It does not work.
A bigger window increases capacity, not comprehension. The model can see more. That does not mean it weighs everything correctly. Long context introduces noise, distraction, dilution. It is the equivalent of inviting every person in the company to the meeting because the room is bigger. The room holds more people. The meeting does not get smarter.
More information can actually reduce accuracy. When relevant context is buried inside hundreds of thousands of tokens of tangentially related material, the model misses things. It gives equal weight to a critical client conversation and a routine status update from eighteen months ago.
And if the wrong context is retrieved, or retrieved without the right framing, the model reasons confidently on flawed inputs. Not obvious failure. Plausible, well structured, subtly wrong responses. That is where the real damage happens.
Context management is not one problem. It is five.
Storage capacity is what the context window provides. Scaling fast. Necessary but not sufficient. Selection logic is how the right context gets retrieved at the right moment, through vector search, knowledge graphs, and increasingly sophisticated retrieval architectures. Sequencing logic is how context gets assembled in the right order and relationship, which is where protocols like MCP are pushing: formalising how agents access and structure context rather than blindly injecting it. Persistence logic is how context survives across sessions and time, so a conversation from last quarter informs today’s intelligence. And governance logic is how you ensure context is used appropriately: who can access what, what constraints apply, what should never be surfaced in certain situations.
None of this is cheap. Every token of context costs compute. Every retrieval operation costs time and money. At scale, context management becomes a genuine economic decision. Organisations building context layers are making a deliberate trade off: paying more per interaction for dramatically better outcomes.
Here is why this should excite you rather than discourage you.
The difficulty is the moat. If this were easy, everyone would have solved it already and there would be no advantage in it. The organisations that figure out disciplined context management will have solved the problem that makes agentic AI actually work at scale.
Because the truth the industry is only beginning to acknowledge openly is this: agents without reliable context are unreliable. They hallucinate. They lose track. They act on incomplete information. The reason most enterprise AI deployments still feel like demos rather than production systems is not that the models are insufficient. It is that the context feeding them is insufficient.
Solve context, and you solve the reliability problem holding back every serious deployment. That is why this layer matters more than any other.
The gift of open infrastructure
The reason the context layer is so powerful as a competitive investment is precisely because everything else is converging or going open.
The models are commoditising. The orchestration layer is open. The interfaces can be built cheaply. This means you do not need to invest in plumbing. You can concentrate your investment on the one layer where it creates actual differentiation.
A law firm does not need to build a foundation model. A hospitality group does not need to create its own orchestration protocol. All of that exists, and most of it is open. What they need to do is structure what they know and make it accessible to intelligence systems that can act on it.
That is the work. And it rewards the organisations with the deepest knowledge, not the deepest pockets.
Where to start
This is not a transformation programme. Do not let anyone sell you one.
Here is the question that starts the conversation: if you gave a competitor access to every AI tool you use today, what would they still not be able to do?
That gap is your context advantage. Now ask the harder question: is that gap growing or shrinking?
Pick one domain. One team. One workflow where proprietary context makes the obvious difference. Map what that context actually is. Where does it live? Who holds it? What would be lost if the person who carries it left tomorrow?
Structure it. Not perfectly. Not comprehensively. Just enough to connect it to an intelligence layer and see what changes.
The moment people see their own knowledge making an AI system dramatically better than the generic alternative, the resistance recedes. It stops being a theoretical argument and becomes a visible demonstration of what their expertise is worth when it is actually working for them.
One team. One domain. One proof point. Then expand. Each one adding to the flywheel.
The person in the room
Come back to the meeting we started with. The person who carried the context. The one who made the conversation intelligent just by being there.
The context layer does not replace that person. It changes why they matter.
They are no longer in the room because they remember. They are there because they think. Because they can look at what the system has assembled and say “yes, but...” And that “but” is the thing no model can generate. It is taste. It is judgment. It is the wisdom that only comes from years of caring about something deeply enough to know what the numbers do not capture.
For 40 years, we asked people to carry the knowledge and do the thinking. Now the systems can carry the knowledge. What remains is the thinking. The work that was always the point.
The window
The models will keep getting smarter. The orchestration will keep getting more open. The interfaces will keep getting cheaper. None of that is your edge.
Your edge is what you know. The institutional wisdom your organisation has built over years. The relationships. The decisions. The things that are true about your business and nobody else’s.
The context layer is how you turn that into something an intelligence system can use. And the organisations building it now are creating an advantage that compounds with every passing day.
Part One of this series explored how four decades of human work created the foundation for the intelligence era. Part Two unpacked the architecture of what is being built and why seeing it clearly is the prerequisite for every decision that follows. This piece is about the layer where the advantage lives, and why building it is both the hardest and the most valuable work in the intelligence era.
The models are the same for everyone. The tools are the same for everyone. The context is yours.
The question is not whether it matters. That is settled.
The question is whether you are building yours while the window is still open.
This is Part Three of the Intelligence Era series.
Part One, “We Taught the Machines Everything. Now What?“ explored how four decades of human work, every form filled, every workflow built, every process documented, created the foundation that intelligent systems now run on.
Part Two, “Welcome to the Intelligence Era“ unpacked the five layer intelligence stack, why the architecture of computing is inverting, and why understanding the shift is the prerequisite for every decision that follows.
Craig Hepburn is Co-Founder & CEO of RAIN Ventures, an AI venture studio building intelligent systems and “useful intelligence” for businesses. He writes about the strategic implications of AI at the intersection of technology and business transformation.