Routers

Chapter 1: The Bargain We Made Without Knowing

In the spring of 2024, a woman in Portland sat in her living room and asked her phone to help her understand why she had been tired for three months. She didn't call her doctor. She didn't schedule an appointment. Instead, she opened an app that did not exist two years prior and spoke to an AI system trained to reason through medical information in language that felt like a conversation with someone who had time to listen.

The system asked her questions: When did it start? What else have you noticed? Do you have a sore throat? Any fever? The woman answered. The system, within seconds, suggested possibilities—anemia, vitamin deficiency, sleep apnea, thyroid dysfunction. It told her which were most likely given her age and history. It explained what each condition meant. It told her what questions to ask her actual doctor when she finally called one.

She never experienced this interaction as an infiltration. She experienced it as a solution to a problem she'd been carrying for weeks: the problem of a healthcare system with too few doctors and too many hours of waiting. The system was useful. It was frictionless. It felt like optimization.

This is how the bargain was made. Not in laboratories or board meetings, though those occurred too. But in small moments of genuine convenience, repeated millions of times, until an entire infrastructure of routing decisions had been wired into the spaces where humans live and think and suffer. By the time anyone thought to ask whether this was a choice we were making, we had already made it.

The problem is not that AI arrived suddenly. It is that AI arrived gradually, which means we are still not sure it arrived at all.

Consider the timeline: ChatGPT was released in November 2022. Within two months, it reached 100 million users—the fastest adoption of any software in history. By late 2023, the major AI companies were announcing medical applications. OpenAI launched ChatGPT Health. Google was integrating Gemini into its medical search. Meta was weaving AI assistants into the core functions of Instagram and WhatsApp. Amazon's Alexa was adding voice-activated medical reasoning. Claude, Anthropic's reasoning model, became available through apps designed to interface with patient data management systems.

What is striking about these launches is not their boldness but their normalcy. They arrived as features, not revolutions. A new button in an app. A new layer of functionality in a voice assistant you already had. A small optimization to a system you were already using. The companies marketing these systems emphasized one thing above all: usefulness. Faster access to information. More convenient triage. Better matching of patient needs to available doctors. The logic was so straightforward it required no advertising. These systems saved time. In a healthcare environment chronically short on time, this alone was revolutionary enough.

But here is what made these integrations different from the previous 40 years of healthcare technology:

For the first time, the systems making recommendations were not stored in a database that humans had built. They were inferred by machine learning models trained on patterns in human data—medical records, lab results, symptom histories, drug interactions, demographic information. For the first time, when you fed your medical information into these systems, you were not retrieving pre-written answers. You were triggering a process of pattern-matching that happened in real-time, in the model's latent space, in ways that even the people who built the model could not fully articulate or predict.

For the first time, the routing decisions—where your information went, what system evaluated it, what recommendation emerged from that evaluation—were happening at machine speeds. A doctor diagnosing a patient thinks. She reviews the chart. She sits with the patient. She considers. This takes time. The AI system ingests your data and produces a recommendation in milliseconds. This is not a philosophical difference. It is a structural one. It means the decision has already been made before you could have asked for a second opinion.

For the first time, these systems were integrated directly into the infrastructure where decisions happen. Your medical history did not stay in your doctor's file cabinet. It was not even stored on a hospital's secure server. It was synced with your phone, your cloud storage, your health app, your pharmacy's system, potentially your insurance company's backend. Once integrated, this data became available to the routing system instantaneously. The system did not need you to request an evaluation. It could initiate one. It could pre-populate your options before you even entered the doctor's office.

And here is what made this integration feel urgent rather than alarming: the existing system was broken in very human ways.

If you have sought medical care in the United States in the past decade, you know this. A minor concern requires a two-week wait for an appointment. The appointment lasts 12 minutes, during which your doctor scrolls through a hospital's electronic records system designed by a software vendor who has never met a patient. You describe your symptoms. The doctor nods while typing. The visit ends with a referral to a specialist, which adds another month to the process. If your insurance disagrees with the specialist's recommendation, you wait for an appeals process that takes another month. The average American with a chronic illness spends more time scheduling medical care than receiving it.

An AI system that could compress this timeline by even 20 percent was not a luxury. It was, in the words of every executive who pitched it, a solution to a crisis.

Here is where the bargain becomes invisible: no one had to decide that speed was valuable. Speed was already the scarcest resource in healthcare. No one had to be convinced that routing medical decisions through a machine was acceptable. The alternative—waiting—was already so costly and painful that the machine looked like mercy.

The same logic arrived simultaneously in five different domains, each with its own flavor of crisis.

In customer service: Companies like Amazon and Meta had spent a decade trying to reduce the cost of human customer support. They hired workers in low-wage countries, imposed strict call time limits, automated everything they could. By 2023, the math was simple: a human customer service representative cost a company between $35,000 and $50,000 per year in salary and benefits. An AI system that could handle 80 percent of routine inquiries cost essentially nothing after the initial training. Replacing humans with routing systems was not a choice about technology. It was already a business requirement. The AI systems just made it faster, cheaper, and technically feasible.

In content moderation: Facebook and other social platforms employed tens of thousands of people to review reported content and decide what violated policy. These workers were often traumatized by the material they encountered. They worked for subcontractors, in offices in the Philippines and Kenya, for wages that amounted to a few dollars per hour. By 2023, AI systems that could flag potentially policy-violating content with reasonable accuracy became available. This was framed as a humanitarian solution: spare human workers from psychological harm, scale moderation to the billions of posts uploaded daily. That the systems were also more profitable was, in presentation, almost incidental.

In hiring: Large companies received thousands of applications for every open position. Humans could not possibly review all of them. So the companies had already turned to automated resume screening. When AI systems became available that could predict which candidates would be successful employees—ostensibly based on the patterns in successful hires from the past—this seemed like an obvious upgrade. The system could find better candidates faster. That it might also encode the company's historical biases into a mathematical algorithm was a concern that could be addressed with a technical fix. The urgency of solving the scale problem was immediate. The concern was abstract.

In financial trading: The system had already evolved to the point where humans could not possibly participate in real-time market movements. By the time a human trader saw that a certain stock was moving, the institutional traders with algorithmic systems had already executed thousands of trades. The only way to compete was to hand the routing decisions to machines. By 2023, asking whether this was wise was almost quaint. The system was already wired. The question was only how much further to optimize it.

In email: Every major email provider had, by 2023, built AI systems that read your email before you did. Gmail's smart compose would finish your sentences. Outlook would summarize your inbox. These were small optimizations. Useful ones. They saved seconds per day. But they meant that your incoming communication was already being routed through a system that made inferences about what mattered to you before you had read it yourself.

In each case, the same pattern repeated: an existing system that was inefficient, human, and broken. A technological solution that was more efficient, faster, and easier to implement. A small optimization that wove the system deeper into the infrastructure where decisions actually happen. And because each individual optimization solved a real problem—because each one was genuinely useful—no one had to be convinced to accept it. It arrived as a solution, not an imposition.

This is how infrastructure becomes invisible. Not through conspiracy, but through convenience. Not through coercion, but through the daily choice to solve a problem that is sitting in front of you right now, rather than to think about the problem you might be creating in the future.

By 2024, the integrations were proliferating so rapidly that no one was tracking them all anymore. Every major tech company was announcing new features. Claude was being integrated into coding platforms. ChatGPT was being wired into workplace productivity software. Google's Gemini was appearing in every Google product: search, Gmail, Docs, Drive, the calendar system you use to schedule your life. Meta was putting AI into Instagram, where it would learn what you photographed, what you liked, and what kind of content would keep you on the platform longest.

But here is what is important about these integrations: none of them felt like a moment. Each one felt like a feature.

You did not choose to wire AI into your hospital. Your hospital chose to use a triage system that routes patients more efficiently. You did not choose to have AI read your email. Gmail chose to add a feature that made email more useful. You did not choose to have AI make hiring decisions about you. The company that rejected your application chose to use a system that screened more fairly, at least ostensibly. Each decision was rational, local, and made by someone with the authority to make it.

What no one was authorized to make—what no one was even asked to make—was the meta-decision: the choice that we were building something systemic, something integrated, something that would eventually route decisions through machines at scales and speeds that would make human oversight impossible.

By late 2024, the major AI companies were beginning to acknowledge what had happened. Sam Altman, the CEO of OpenAI, started giving talks about the "multimodal, agentic systems" they were building—systems that could not just respond to prompts but could take actions, move through other systems, execute decisions. These systems would be wired into companies and institutions in ways that made them less like tools and more like employees. They would be embodied in robotics. They would be integrated into operating systems. The framing was always positive: this is how we solve hard problems faster.

What was being described, in other words, was the completion of a nervous system.

Not a system controlled by any single company, but a distributed nervous system where decisions about your health, your employment, your finances, your security, and your attention were routed through machine learning systems designed to optimize for speed and efficiency. A system where you could not opt out because opting out meant opting out of medicine, employment, banking, and social participation. A system that had been wired together not through any grand plan but through thousands of small decisions to solve immediate problems.

The bargain was that we would gain speed and efficiency. We would get medical triage in seconds instead of weeks. We would get customer service that never put us on hold. We would get hiring decisions that were theoretically more fair because they were algorithmic. We would get financial markets that found the true price of things faster. We would get content feeds that showed us what we actually wanted to see, or at least what the system predicted we wanted to see.

The cost of the bargain was something we did not think we were paying: the cost of deliberation itself. The cost of thinking. The cost of the slow, human processes of review and consideration that used to stand between a problem and a decision. These processes had always been imperfect. Doctors make mistakes. Hiring managers have biases. Financial traders make bad bets. But what they allowed for—what the slowness actually purchased—was the possibility of changing your mind. Of asking a second opinion. Of saying: wait, let me think about this. Of inserting yourself, your judgment, your values, into the moment before the decision was executed.

The faster systems did not allow for this. The faster systems made the decision and then, in the next millisecond, the decision was already cascading through other systems. Your insurance claim was approved or denied. Your medical record was updated. Your credit score was adjusted. Your application was rejected and the position was filled. The trading algorithm had already hedged its bet. By the time you realized the decision had been made, it was too late to ask the system to wait.

None of this happened because anyone chose speed as a value. It happened because everyone chose to solve the problem in front of them, using the tools available, and the tools available happened to be faster.

The woman in Portland who asked her phone about her tiredness eventually did see a doctor. The doctor confirmed that she had a vitamin deficiency, exactly as the AI system had suggested. She felt vindicated in the tool. It worked. It had saved her weeks of uncertainty. She did not think about the fact that she had voluntarily wired her complete medical history into a system that made inferences about her body faster than she could think.

She probably did not think about what would happen if that system made a different kind of inference. If it detected something in her data that it categorized as a risk factor for a disease she did not have. If it routed her toward expensive treatment based on a probabilistic guess. If the insurance company used the same system to deny her coverage. If the pattern recognition had learned to see patterns that were not there, but became real through being acted upon.

She probably did not think about the moment—which may or may not have already passed—when the integration had gone far enough that slowing down the system, demanding human review, asking for deliberation, became technically impossible, economically ruinous, and politically infeasible. When the system had become infrastructure, and infrastructure does not wait for humans to catch up.

The bargain was made quietly, through a thousand small optimizations, because optimization requires no permission. It only requires that you solve the problem in front of you, using the tools available, and that you do not ask what else those tools might be routing through your life while you are not looking.

We have built something that makes decisions faster than we can review them. We have called this progress. We have integrated it into the places where our lives are actually determined: hospitals, courtrooms, banks, hiring systems, military command structures. We have done this incrementally, so that we could not point to a single moment when we decided to do it. We have done this in ways that solved immediate, real problems, which is why resistance has felt like technophobia rather than wisdom.

And now we are living inside it.

The nervous system is not coming. It is here. The only remaining question is whether we notice what it is doing before the routing becomes so integrated, so fast, so invisible, that asking the system to wait feels like asking water to flow uphill.