The Spinal Cord Problem

Why the most consequential intelligence in any institution was never in the data

The spinal cord is one of the more underappreciated achievements of biological engineering. It receives signals, coordinates responses, and keeps the body functioning — often faster than conscious thought can intervene.

It is not a lesser system. It is a different one. Precise, fast, and entirely without judgment.

I have been thinking about this distinction as AI becomes increasingly embedded within institutional workflows. Not because the concerns about it are overblown, but because the anatomy feels more accurate than most of the language currently used to describe what AI actually does — and what it does not.

AI, at its current best, is the spinal cord of institutional decision-making. It finds the anomaly, flags the exception, surfaces the pattern. These are not trivial capabilities. But the spinal cord does not ask why.

The formal layer and what lives beneath it

Over time, I have come to believe that most institutions operate through two parallel systems.

The first is the formal layer: the documented process. Models, dashboards, structured data, policy frameworks, recorded decisions. This is the layer institutions intentionally design and measure. Increasingly, it is also the layer they automate.

The second layer is harder to describe because it was never fully designed.

It is the conversation that happens because a number felt wrong. The question a relationship manager asks a client because something in the management discussion did not quite add up. The researcher who spends eighteen hours observing customers in a parking lot because the focus groups produced nothing useful. The famous milkshake research example in product management is memorable for precisely this reason: the structured research surfaced very little, while informal observation surfaced almost everything.

This second layer has always been where the consequential intelligence lives. It is rarely documented, often unrecognized, and almost never measured.

But it is the layer that asks the questions the first layer cannot generate on its own.

The cost anomaly

A few years ago, our underwriting team was reviewing an American manufacturing client during the peak of the US-China trade tensions. The business model was straightforward: equipment manufactured in China, sold into the United States. Import duties at the time had risen above one hundred percent, and the expectation was that the cost structure would reflect that.

It did not.

The financial statements showed a sharp decrease in net costs over the period. Nothing in the management discussion explained it. On the surface, the numbers were strong — improved margins, better pricing power than competitors, healthier profitability than the prior year. Any sufficiently capable model would have flagged the anomaly. Ours did.

What happened next is the part that interests me now.

We sent a question to the relationship manager over email. The relationship manager, in turn, called the client. The conversation was informal. The kind that only happens when the relationship already exists.

The client had front-loaded imports ahead of the tariff increase when a change in presidency became likely. The inventory advantage materially reduced costs during the period. They then raised prices modestly — still below competitors — and captured the spread as margin. The profitability improvement was real, but non-recurring.

In the same conversation, almost incidentally, it also emerged that the client had no exclusivity agreement with the Chinese manufacturer. This had not appeared anywhere in the documentation either.

Both pieces of information were material. Neither existed in any document we had reviewed. They surfaced because someone asked the right person an open question in a context where the answer could be honest. We flagged the profitability as one-off and rated the credit accordingly.

The judgment did not reside in any individual actor. It emerged from the interaction between systems, relationships, and context.

What AI is actually automating

AI is compressing the formal layer with remarkable speed. Document review, anomaly detection, pattern recognition, policy checks, information extraction — tasks that once consumed significant analyst time are becoming nearly instantaneous. This is genuine progress. The spinal cord should be fast.

Formal systems matter precisely because they make large-scale coordination possible. But coordination and comprehension are not identical.

The risk is not that AI is doing this. The risk is the conclusion that often follows.

When the visible layer becomes faster and more capable, it is easy to assume the underlying intelligence has scaled with it. That a workflow which now runs in minutes instead of weeks is somehow doing more than running quickly. That because the process is automated, the thinking is too.

In many cases, the process has become faster long before it has become meaningfully wiser.

The milkshake researchers made a version of this mistake first. They improved the product based on everything the structured research told them, and the sales did not move. The useful insight existed outside the structure of the research itself.

The layer beneath the workflow

The relationship manager in the underwriting case was not following a script. The question they asked the client was not generated by a workflow. It emerged from something less visible — a sense, built up over time, that the explanation in the documents did not quite fit the shape of the business.

That kind of sensing is not romantic. It is structural. It exists because relationships exist. Because context accumulates over time. Because the same two people have spoken often enough that an unusual question does not feel like an interrogation.

This is the layer institutions struggle to formalize.

A prompt can describe a question, but it cannot reproduce the conditions under which the question becomes answerable. A workflow can require a follow-up. It cannot guarantee the kind of relationship in which the answer stops being performative.

The intelligence in the informal layer is therefore not located in any single individual or any single tool. It emerges from the interaction between systems, relationships, incentives, memory, and context. That makes it extremely valuable, but also difficult to see clearly from inside the institution itself.

It rarely appears in dashboards. It does not fit neatly into audit trails. When decisions go well, it is usually absorbed into the appearance of process. When decisions fail, its absence is almost never identified directly.

And yet much of the ambiguity institutions navigate is still being resolved there — in the layer beneath the formal system, where context exists before it becomes structured knowledge.

The anatomy lesson

The institutions that will navigate this period well are not necessarily the ones adopting AI fastest. They are the ones that understand which layer they are automating.

A faster spinal cord is a meaningful advantage. It allows scarce judgment to be directed at the questions that actually require it, rather than diluted across tasks that never did. Used well, AI does not replace the layer that thinks. It protects it.

The risk lies in thinning out the sensing layer simply because the processing layer has become faster, more consistent, and easier to measure. Relationships are expensive. Context is slow to accumulate. The informal layer rarely justifies itself in the metrics most institutions track. Its contribution is usually mistaken for the process itself. By the time the difference is noticed, the layer is often no longer there to rebuild.

The invisible architecture of good decisions has never been the model, the dashboard, or the workflow. It has been the judgment operating beneath them — the capacity to notice what the data is not saying, and to know who to ask.

AI does not eliminate that distinction.

It makes institutional confusion about it far more consequential.