The Gap AI Didn't Create
Every new AI session forces you to re-explain what your organization already knows. That is not a modeling problem. It is a structural one, and it was there long before any model showed up.
There is a small, specific frustration that anyone who has spent real time working with AI models will recognize immediately. You open a new chat. You start explaining the project. And at some point, usually a few minutes in, you realize you are reconstructing something that already exists: a decision made three weeks ago, a document someone else wrote, a piece of context another conversation already established.
You have not learned anything new. You have just spent ten minutes rebuilding what you already knew, so that a tool with no memory of it could catch up.
Do that daily, across a team, across every new session and every new project, and it stops being a minor annoyance. It becomes a tax.
One organization eventually stopped treating this as an AI problem and started treating it as an organizational one. That shift began while Guy Derry was building out a content operation, and it led him toward a more interesting question than “how do we get better outputs from AI.” His question was closer to: why does an organization keep losing the same context, over and over, and why did it take a chatbot to make that loss visible.
The problem was never the model
Guy’s account of the moment it clicked is less a single scene than a slow accumulation. He was rolling out a content operations setup, using AI tools more heavily than before, and kept running into the same wall: every new chat session meant pointing the model back to previous conversations, previous documents, previous decisions already made.
The gap did not arrive with AI. The models just removed the cover that had been sitting on top of it.
This is worth sitting with, because it runs against instinct: when an AI project underperforms, the instinct is to blame the model for hallucinating or missing something obvious. But a model that hallucinates a detail was usually never given that detail, guessing because it has no memory of your organization.
Humans have always covered for this gap without noticing they were doing it. A new hire asks a colleague instead of searching a wiki. A manager remembers a decision’s rationale even though it was never written down. Continuity gets carried in people, quietly, and organizations have depended on that for as long as they’ve existed. An AI agent cannot do any of that. It cannot ask around. It cannot remember a hallway conversation from eighteen months ago. It is the first participant in an organization’s history that cannot be quietly folded into its tribal-knowledge economy. So it fills the blank the only way it can, and the fill is visible in a way a human’s improvisation never was.
This reframing matters because it changes where the fix has to live: not in next quarter’s model release, but in whether context can keep moving, or dead-ends in someone’s memory.
Why this took a chatbot to notice
The first response to the problem was, by his own account, more manual than it needed to be. Like a lot of people experimenting seriously with AI tools, he started with a saved library of prompts and supporting material, maintained by hand. It worked, in the sense that it produced results. It also meant he was doing the same recovery work the model itself was supposed to be saving him from, just one layer up.
There was no single morning when it broke. It wore through instead, the way a stair tread wears down one footstep at a time. Another session. Another round of copying context back in. Another reference to a conversation that should have still been live. “It’s when you’ve done it for the umpteenth time,” he says, by which point the tread is worn thin enough that everyone can feel it give.
One working system revealed the others.
What actually gets built once you take the gap seriously
Every organization eventually builds systems for finance, for customers, for hiring, for engineering. Context is usually the exception: it accumulates wherever work happens, then quietly disappears when people change roles, projects end, or conversations fade. Most organizations stop once they recognize the Context Tax.
Recognition is cheap. The harder question is what an organization actually builds once it accepts that the problem is structural rather than technical.
The first version was a self-contained space for one function: everything content operations needed in one place, operating guidance, prior research, historical work, plus a review step that checked each draft from several angles before anything was called finished. What changed wasn’t the model. It was that context stopped being assumed and started being explicit.
A self-contained space works right up until you notice that nothing in an organization is actually self-contained: content connects to product decisions, which connect to engineering constraints, which connect to strategy. Build one working island and the others become visible, none of them built to talk to each other. A single pricing decision, for instance, gets reconstructed four separate times: once by marketing writing about it, once by product building around it, once by an engineer implementing the edge case it created, and once by an AI agent asked to summarize “why we charge what we charge,” each starting from zero and arriving at a slightly different answer.
What replaced the islands was a central index: decisions stopped living only inside documents and started living somewhere every function could reference, so a decision made anywhere is discoverable, not just documented, elsewhere.
The mechanics stay undisclosed. What disappeared once it existed doesn’t: a recurring meeting whose only purpose had been re-explaining a decision to people who’d already been told it once.
Almost every organization experimenting with AI has diagnosed the same gap and stopped there, running pilots that never connect into anything that compounds. An organization full of disconnected experiments looks identical, from the outside, to one that has not started.
The metaphor that survives contact with the argument
What remains surprising, months into building this, is less an argument than an image.
“The metaphor I keep coming back to is water,” he says. “Knowledge wants to flow. Context wants to flow. Our job isn’t to force that flow. It’s to remove the blockages that prevent it.”
It is a small claim dressed as a large one, and that is exactly why it holds up. Nobody has to be persuaded that water flows downhill; the only real design decision is whether you dig the channel or leave the water to pool. Applied to an organization, the same logic says something uncomfortable: most of the effort currently spent on AI adoption is effort spent trying to make water flow uphill, one prompt at a time, instead of removing whatever is damming it in the first place. The same logic extends past AI entirely: “it’s really about bringing every team inside an organization forward in a context-preserving way,” whether what’s arriving next is a new hire, a new model, or an autonomous agent.
This is where the piece has to be honest about its own limits. Guy’s evidence here is one company, his own, observed from the inside. That is a real data point, not a survey, and it should be read as one. This is one organization’s experience. Whether the pattern holds beyond it is exactly what Orchestrate’s reporting exists to find out.
Disclosure: Every publication has to begin somewhere. We chose to start with ourselves, examining Kernel, the company behind Orchestrate, before turning the same reporting approach toward others.