The Byblos Problem

Three thousand years before anyone could fathom what a PowerPoint was beyond the sharp end of a stick, the people of Byblos solved the same problem every organization is failing to solve today.

They were a small Phoenician port city on the Lebanese coast, and they had something Egypt desperately needed: cedar. The great forests of Lebanon produced timber that the Nile delta could not. Egypt had gold, linen, and grain. Byblos had the wood to build ships, temples, and coffins for pharaohs.

The trade route was obvious. What was not obvious was the organizational design required to sustain it.

The merchants of Byblos did not simply load timber onto boats. They redesigned how decisions moved through their city. They created specialized roles for negotiation, logistics, and relationship management with the Egyptian court. They built information networks along the coast so that price signals, political shifts, and fleet movements reached the right people before they reached anyone else. They separated the people who detected opportunity from those who acted on it, and they connected the two groups with clear lines of accountability and speed.

The result was not just a successful trade. It was dominance. For centuries, Byblos was the most powerful commercial force in the Mediterranean. The Egyptians named the city in their records. The word for book in Greek, biblos, comes from this place, because so much papyrus passed through its harbor that the city became synonymous with the written word itself.

Lebanon’s forests were available to anyone willing to sail there, and having the best cedar was not their competitive advantage, allowing them to win. They won because they built an organization that could move insight into action more effectively than anyone else in the region.

The technology, in their case, was maritime trade. In ours, it is artificial intelligence. The design problem is identical. MIT has been circling this for months now, and they finally published on it properly last week.

In their coverage of agentic AI and organizational design, MIT Technology Review put into words what many in technology have been feeling but not quite articulating: most companies are failing because their organizations were not designed to make decisions with AI in the loop.

I have spent the last two years building something I call Intradiegetic. It is a distributed AI infrastructure, yes, but more than that, it is a thinking system. A way of asking: what does it look like when a human and a machine build a decision architecture together, from first principles, rather than layering intelligence onto a structure that was designed for a different era?

The answer, it turns out, is that you cannot separate the technology from the philosophy of who owns what.

Let me start with the MIT observation. Their research, and the parallel work from Davenport and Bean at MIT Sloan, converges on a single finding: only 23% of AI deployments in organizations today generate a clearly measurable return on investment. The failure rate of corporate AI projects sits above 45%. And when you ask why, the answer is almost never the model. It is the organization.

Specifically, organizations are layering AI agents onto decision structures built for human-speed, hierarchical, politically managed certainty. And AI produces something different. It produces probabilistic, fast, sometimes counterintuitive insight. The structure cannot absorb it. The insight disappears. The investment disappears with it.

Jim Wilson of Accenture, speaking at MIT’s BIG.AI conference in April, framed it this way: the ROI question for AI should not be which tool to buy. It should be determined by whether the organization is structured to adopt it well.

I think that framing is close, but it still understates the problem.

The deeper issue is not just structure. It is accountability.

Here is the pattern I have seen repeatedly in corporate communications, infrastructure operations, and investor relations. An AI system surfaces a signal. A risk pattern, a sentiment shift, a crisis emerging at the edges before it reaches the center. The system is correct. The insight is actionable. And nothing happens.

People know the information. They have it in front of them. But nobody is accountable for the cost of ignoring it.

Aristotle made a distinction between two kinds of knowledge: episteme, the knowledge of facts, and phronesis, practical wisdom, the capacity to act well in the face of uncertainty. Most organizations are investing heavily in episteme. They are buying the best models, data pipelines, and dashboards. And they are treating phronesis as somebody else’s problem.

But phronesis is not a personality trait you can hire for. It is a structural condition. It emerges when the person who receives the insight is also the person who bears the consequence of ignoring it. When those two things are separated, you do not get wisdom. You get information theatre.

The crisis communications world is a useful lens here because the stakes are visceral.

Mid-market companies, the 500- to 5,000-employee organizations that form the backbone of most economies, face exactly the same crisis exposure as large enterprises. Cyber incidents are now the second-most common trigger for formal crisis plans. Reputational events move faster than ever. And yet no integrated, agentic solution exists for this tier of company.

The enterprise platforms are designed for governments and multinationals. The cost and complexity of implementation place them outside the reach of most mid-market buyers. The gap is real and widening as the pace of crises accelerates.

What does an agentic crisis communications architecture actually need to do? It needs to detect a signal before it becomes a headline. It needs to draft a holding statement in the time it takes to convene a call. It needs to route the right information to the right people without a human having to manually triage every alert at 3am.

All of that is technically achievable today with off-the-shelf open source tools. The architecture exists.

What most systems get wrong is the human gate.

In Intradiegetic, I have a rule I have come to call the distribution principle.

Automation handles speed at detection and drafting. Human judgment owns distribution consequences. Always.

It is a philosophical position much more than a liability hedge.

The philosopher Paul Ricoeur wrote about what he called the attestation of self, the idea that identity is not something you possess statically but something you demonstrate through action under uncertainty. You do not know who you are in a crisis by reading a policy document. You know it in the moment of decision, when the draft is ready, and the send button is visible, and something in you pauses.

That pause is the system working.

Every agent pipeline I build has what I call a decision authority policy: a documented answer to what the agent can do autonomously, what requires human approval before execution, and what is entirely off-limits to automated systems. I don’t particularly distrust the models, but I believe that the moment you remove human accountability from a reputational decision, you haven’t made the process faster; you have made it hollow.

MIT Sloan’s Davenport and Bean predicted, at the start of this year, that 2026 would be a year of reckoning for agentic AI. The bubble, they argued, would begin to deflate as organizations confronted the gap between what agents promise and what existing structures can absorb.

I think they are right about the reckoning. I think they are slightly wrong about the cause.

The bubble is not inflated by bad technology. It is inflated by a category error: the belief that deploying an agent is a technology decision rather than an organizational one. Companies are asking their CTO to answer a question that is actually meant for their Chief of Staff, General Counsel, or Chief Communications Officer.

Who owns the decision that follows the model?

If you cannot answer that question for every output class your AI system produces, you do not have an AI strategy. You have an AI pilot that is about to teach you something expensive.

What I find philosophically interesting about this moment, and I think it is genuinely interesting rather than just professionally convenient, is that the question AI is forcing onto organizations is one that philosophy has been asking for millennia.

What does it mean to act well under uncertainty?

Epictetus drew the line at what is within our power and what is outside it. The Stoics built an entire ethics around the discipline of distinguishing between the two. And what AI is doing, structurally, is redistributing that line. Some decisions that previously required a human now sit firmly outside human capacity, in the sense that a human cannot process the signal fast enough to act on it. The agent can. Other decisions, the ones with reputational weight, with ambiguity, with consequences that ripple in ways the model cannot fully anticipate, remain inside the human domain in a way that no level of model sophistication will change.

The organizations that figure out where that line sits and build their architecture around it deliberately are the ones that will generate that elusive 23% return. The others will generate beautiful dashboards and expensive lessons.

Intradiegetic is my attempt to work this out in public, one build at a time.

It is named for a concept from narratology: intradiegetic, existing within the story world, as a participant rather than a narrator. I chose it because I think that is the only honest posture for someone building AI systems right now. You are inside the story. The design choices you make are not neutral observations about technology. There are arguments about how decisions should be made, who should be accountable for them, and what it means to act well when the information arrives faster than the wisdom to use it.

The question I keep coming back to is whether organizations are sufficiently accountable to act on what their agents know.

That is the design problem. And until it is treated as a design problem, the investment will keep flowing, and the value will keep disappearing into the gap between insight and decision.

What does accountability look like in your organization when the model already knows what needs to happen?

The merchants of Byblos did not ask whether cedar was valuable. They asked who was accountable for getting it to Egypt before the next ship from Tyre did. That question built a civilization. It is still the only question that matters.