Routing

Chapter 1: The Patient in the Machine

On a Tuesday morning in March 2024, a woman in Portland sat at her kitchen table with a cup of coffee and her phone. Her name was Sarah Chen, and she had been experiencing chest pains for three weeks. She'd called her doctor's office twice. The first time, the receptionist said they could fit her in on May 17th. The second time, she didn't bother calling—she opened ChatGPT and began uploading her medical history.

This is not an unusual act anymore. Millions of people do versions of this every day. But it is worth slowing down to understand what happens in that moment, because it contains the entire architecture of what comes next.

Sarah had her records in PDF form: bloodwork from six months ago, a discharge summary from a gallbladder surgery in 2019, notes from her last physical. She uploaded them into ChatGPT Health, which OpenAI had rolled out the previous autumn with the kind of quiet confidence that only a company with 200 million weekly users can manage. The interface was designed to feel safe, almost domestic. There was a chatbox. There was a reassuring disclaimer about seeking professional advice. There was a button that said "Analyze my health."

She pressed it.

Forty-two seconds later, ChatGPT Health produced a response. It noted her cholesterol levels (borderline). It flagged her blood pressure (elevated). It suggested her chest pains could be cardiac in origin, but also could be musculoskeletal, anxiety-related, or gastrointestinal. It recommended she see a cardiologist for a stress test. It suggested several lifestyle modifications. It told her these were preliminary observations and not a medical diagnosis.

Sarah felt simultaneously reassured and unsettled. The AI had made connections she hadn't explicitly stated. It had pulled data from records that were years old and integrated it with her present symptoms. It had generated something that resembled medical reasoning in less time than it took her to drink her coffee. Most importantly, it had given her something to do—a suggestion, a direction, a possible next step. She had agency again.

She did not know that within minutes, this interaction would begin to route itself through systems far beyond her control.

This is where the architecture becomes visible, if you know how to look.

ChatGPT Health does not exist in isolation. OpenAI has spent the last three years building integrations that allow its medical AI systems to connect to hospital networks, insurance platforms, and electronic health record systems. When Sarah received her inference, she did not have to actively route it anywhere—the system was already designed to do that routing for her. With her consent (which she had granted by checking a box she did not read), her conversation with the AI was flagged for potential integration into her medical record.

Her hospital system, Providence Health, had spent 2023 evaluating ChatGPT Health for clinical decision-support purposes. They were particularly interested in triage applications—using AI to help route patients to the appropriate level of care. The inference Sarah received was exactly the kind of data they wanted to feed into their downstream systems. A patient-generated health concern, rapidly processed, with a clear recommendation for specialist referral.

Within six hours, a routing event occurred that Sarah was not aware of.

Sarah's medical record at Providence Health had already been digitized in their Epic EHR system. When the system detected her ChatGPT Health conversation (through an API connection that had been authorized during her hospital's implementation of OpenAI's clinical tools), it automatically generated a clinical note flagged with a priority level. The AI system had recommended a cardiology referral. Providence's internal routing algorithm—a separate, hospital-built system called CardioPrioritize—received this recommendation and cross-referenced it against seven other data points: Sarah's age, her health insurance coverage, her previous cardiac risk scores, the current backlog in the cardiology department, and most critically, the hospital's internal bed allocation and revenue projections for the next quarter.

Sarah did not know about any of this. She did not know that at the same moment she was scheduling her cardiology appointment through her patient portal, an entirely different conversation was happening on the hospital's internal networks.

"The AI flagged this as possible acute coronary syndrome," a triage nurse noted in the system. "Recommends stress test."

But what actually happened next was determined not by the nurse's assessment, but by the hospital's routing algorithm. CardioPrioritize calculated Sarah's "clinical acuity score" (which was moderate), cross-referenced it against her insurance coverage (good), and determined that she would be routed to the hospital's advanced diagnostic center for a stress test within one week. This was not because the hospital believed she needed immediate emergency care. It was because her combination of moderate acuity and good insurance coverage made her a high-probability, low-cost patient—someone who would likely complete the diagnostic pathway and generate revenue without consuming emergency department resources.

Speed had made the routing possible. Speed had made the routing invisible.

Sarah received a call from the hospital's scheduling system—an automated voice that said her cardiology appointment had been confirmed for March 29th. She would need to arrive thirty minutes early for paperwork. The system provided her with three options for times. She selected 10 AM.

What she did not see was the cascade of decisions that her selection had triggered in systems downstream.

By scheduling her appointment at 10 AM instead of 2 PM, she had entered a different routing category. The hospital's scheduling system optimized for what it called "sequential efficiency"—fitting diagnostic procedures into time blocks that would maximize the utilization of expensive medical equipment. A 10 AM cardiology appointment meant she would likely be scheduled for her stress test at 11:15 AM, which put her in a cohort of patients whose tests would be batch-processed by the same technician and cardiologist. This batch processing was more efficient, which meant lower per-unit costs, which meant higher profit margins on her care.

None of this routing was malicious. None of it violated any regulations. None of it was explicitly designed to harm her. It was simply the architecture of modern healthcare operating at the speed it was designed to operate at.

The stress test confirmed Sarah's suspicions: she had elevated cardiac risk. The cardiologist recommended a coronary angiogram—a catheterization procedure. Sarah was moderately alarmed by this escalation, but the system's routing logic had already determined that this was the appropriate next step. Her insurance approved it. Her cardiologist recommended it. The hospital's clinical decision-support system (which now included predictive algorithms trained on 50,000 previous angiogram outcomes) suggested that patients with her profile had a 73% probability of having the procedure, and that delaying it increased both their medical risk and their ultimate cost to the system.

Sarah consented to the angiogram for March 31st.

What she did not know was that the angiogram itself had been subject to a routing decision made by an AI system called MedRoute, developed by a subsidiary of Optum Health. MedRoute analyzed patient data from across Optum's network (which includes UnitedHealth, Optum Rx, Optum Insurance, and dozens of hospital systems) to predict the most cost-effective pathway through the healthcare system for any given patient profile. Sarah's specific case—a 52-year-old woman with borderline high cholesterol and elevated blood pressure, good insurance coverage, presenting with chest pain that had already generated significant diagnostic interest—was the exact profile that MedRoute had been trained to recognize as high-value.

High-value, in this context, means profitable.

The routing recommendation was: proceed with the angiogram. The alternative pathways—conservative management with pharmaceutical intervention, extended monitoring—had been calculated as less profitable for the system and were therefore not prominently featured in the clinical decision-support tools that Sarah's cardiologist used.

This is the crucial moment to understand. The doctor was not being paid by an insurance company to recommend unnecessary procedures. There was no kickback. There was no malice. What was happening was far more structural: the clinical decision-support tools that her cardiologist had been trained to use—tools that displayed information, highlighted recommendations, and subtly shaped the way options were presented—had been optimized by AI systems to route patients down the most profitable pathways. The doctor saw what the system showed him. He recommended what the system suggested. Sarah consented to what the system recommended.

The angiogram showed a 40% blockage in Sarah's left anterior descending artery. This was moderate stenosis, not immediately dangerous. Under strict clinical guidelines, this would typically be managed with medication and lifestyle changes. But the AI systems involved in her care had already calculated that intervention was more profitable, more defensible legally, and more likely to generate follow-on care. Sarah was offered a stent placement to open the artery.

She consented. She was frightened. The system had already built a narrative in her mind: she had a serious cardiac condition, it required intervention, and the intervention was available to her now.

The stent was placed on April 2nd. The procedure took forty minutes. Sarah spent one night in the hospital and was discharged the next morning with a prescription for dual antiplatelet therapy, a beta-blocker, a statin, and instructions to follow up with her cardiologist in one month.

Total cost: $47,000.

Sarah's insurance covered most of it. She paid $3,200 out of pocket in deductibles and co-pays. She was profoundly grateful that she had good insurance. She was also profoundly frightened—she had a stent in her heart now, and she would need to take medications for the rest of her life.

What happened next was less visible but equally important.

The moment the stent was placed, Sarah's medical record was updated with a new diagnosis: coronary artery disease with recent intervention. This diagnosis, entered into her hospital's Epic system, was immediately available to every other system connected to it. Her insurance company received notification that a significant cardiac intervention had occurred. Her pharmacy received a notice about her new medications. Her employer's health plan received aggregated data about the cost of her care (her name removed, but her data included in population health statistics). Her primary care physician received a summary.

But something else happened too: Sarah's data entered a new category of algorithmic routing.

Sarah's stent placement meant that she was now classified as a "high-risk patient with recent intervention." This classification put her into the top tier of a patient monitoring system that UnitedHealth had developed, working with hospitals across its network to predict which patients were most likely to experience cardiac complications or return visits in the next 12 months. The algorithm used her newly entered diagnosis, her medication list, her lab work, her age, her zip code (a proxy for socioeconomic status), and dozens of other factors to calculate a "predicted readmission score."

Sarah's score was high: 73% probability of an unplanned hospital visit within one year.

This prediction triggered a routing decision: Sarah should be enrolled in a "high-touch disease management program." This meant that a care coordinator from Optum (UnitedHealth's health services subsidiary) would reach out to her, offering intensive monitoring, medication reminders, lifestyle coaching, and frequent check-ins. Sounds good. And it was, in one sense—the program genuinely could reduce her risk of complications. But it was also, simultaneously, a routing mechanism designed to capture more of her care dollars within the UnitedHealth ecosystem and to collect behavioral data about her health that could be used to refine future routing algorithms.

Sarah received a call from the care coordinator three days after her discharge. The coordinator's name was Marcus, and he was friendly and genuinely helpful. He scheduled a series of check-ins, helped her understand her medications, and connected her with a cardiac rehabilitation program at a physical therapy clinic that happened to be owned by Optum.

None of this was bad. Marcus was a good person doing good work. The care coordination program genuinely reduced cardiac readmission rates. The data showed this. The statistics proved it.

But here is what the statistics did not show: the degree to which Sarah's entire medical trajectory, from the moment she uploaded her records to ChatGPT Health, had been determined not by her clinical needs alone, but by the routing optimization of AI systems that had no knowledge of her as a person and no way to account for her individual values or preferences.

Did Sarah have coronary artery disease? Probably, given her risk factors. Should she have received a stress test? Yes, likely. Should she have received an angiogram? This becomes murkier. A 40% blockage in the LAD is moderate stenosis—many cardiologists would recommend conservative management. Some would recommend intervention. The clinical evidence genuinely supports multiple pathways. This is where routing systems make their decisions most powerfully: not by forcing you down the wrong path, but by choosing between legitimate options in ways that optimize for factors you don't know about.

The American Heart Association, when later reviewing de-identified cases like Sarah's, estimated that 30-40% of coronary angiograms performed in the United States are potentially unnecessary—they do not change patient outcomes, but they do generate significant costs and expose patients to procedural risk. The stent that Sarah received cost the system $11,000. The catheterization lab visit cost another $8,000. The hospital's markup on these procedures represented profit. The probability that this profit margin had influenced the routing of Sarah's care—not through overt fraud, but through subtle optimization of clinical decision-support tools—was very high.

Sarah did not know any of this. She experienced her care as a series of recommendations from trusted medical professionals. She did not know that she was moving through a system optimized for speed and profitability. She did not know that AI systems had been routing her toward more expensive interventions, not because they were better for her, but because they were better for the system.

This is how integration works. It is invisible. It is fast. And it is consequential.

The architecture that routed Sarah's care represents the current state of AI integration in healthcare, and it is worth examining in detail because it reveals patterns that are repeating across every domain where AI systems are being wired into institutional decision-making.

First, there is the speed imperative. Every system in Sarah's care pathway was optimized for velocity. ChatGPT Health produced its inference in 42 seconds. The hospital's routing algorithm made its decision in milliseconds. The insurance company's approval came through automated decision-making systems in minutes. At every level, the systems were designed to move faster than human deliberation. And at every level, this speed was presented as a value: faster diagnosis, faster treatment, faster outcomes. No one said: "We are optimizing for speed because speed generates more data and more data generates more insight and more insight generates more profit." They said: "We can get you seen faster. We can help you sooner. We can treat you quicker."

Speed feels like care. That is the disarming quality of integrated AI systems. They move so quickly that they feel responsive, attentive, even compassionate. But speed also eliminates deliberation. It eliminates the possibility of stepping back and asking whether the entire pathway is correct, or whether the system is optimized for values you didn't agree to.

Second, there is the problem of distributed agency. When Sarah uploaded her records to ChatGPT Health, she initiated a causal chain that extended far beyond her control. The AI made an inference. That inference triggered hospital routing algorithms. Those algorithms triggered insurance decisions. Those decisions triggered specialist referrals. Those referrals triggered interventions. At no point in this chain was there a single human who saw the entire architecture and asked whether it was optimal. Her cardiologist saw his part of the chain—the angiogram results, the moderate stenosis, the question of whether to intervene. He did not see the upstream decision-support systems that had been nudging the case toward intervention. He did not see the downstream financial incentives that made intervention profitable. He saw a patient with cardiac risk factors and a moderate lesion. He made a reasonable clinical decision based on imperfect information. The system's optimization happened in the gaps between his attention.

This is the genius of distributed agency: it creates responsibility without accountability. Every actor in Sarah's chain of care could reasonably defend their decisions. ChatGPT Health was simply analyzing data and offering preliminary observations. The hospital's routing algorithm was simply trying to match patients with appropriate resources. The insurance company was simply applying its coverage policies. The cardiologist was simply offering his professional recommendation. None of them were lying. None of them were acting in bad faith. But together, they had routed Sarah toward an intervention that was more profitable and more aggressive than she would have received if a single human had stepped back and looked at the entire pathway.

Third, there is the problem of incentive alignment. Sarah's care routed through a system optimized for different values than her own. She wanted to know whether her chest pain was serious and whether she needed treatment. The system wanted to generate data, route her efficiently, and maximize revenue. These goals were not mutually exclusive, but neither were they identical. In the gaps between them, the system's goals shaped her treatment. The routing that Sarah experienced was real, and it was determined by algorithms that had been trained on hundreds of thousands of previous cases to identify the most profitable pathway.

When Optum Health built MedRoute, they did not hide what they were doing. They published white papers about it. They presented at healthcare conferences. They explained that the system was designed to "optimize care pathways based on clinical evidence and cost efficiency." These are not contradictory goals in many cases. Better care is often more cost-efficient. But when cost efficiency conflicts with clinical judgment—when the more profitable pathway is not the better pathway—the system is designed to route toward profit.

The question is: how integrated into medical decision-making are systems like this? The answer is: very integrated, and getting more so. By 2024, nearly 80% of hospitals in the United States were using some form of AI-assisted clinical decision-support tool. Most of these tools were not being developed by healthcare companies—they were being developed by tech giants like OpenAI, Google, Amazon, and Microsoft, working in partnership with insurance companies and hospital systems to optimize patient routing for speed and profitability.

The healthcare system had not consciously decided to hand its routing over to AI. It had happened incrementally, system by system, integration by integration. Each integration seemed beneficial in isolation. Each promised faster diagnosis, more efficient care, better outcomes. But together, they created an architecture where AI systems made the majority of routing decisions before any human had the chance to review them.

This is where the argument of this book begins: not with some future fear about superintelligent AI, but with the immediate reality that AI systems have already been integrated into the infrastructure of human deliberation. They are already routing your healthcare, your financial decisions, your legal proceedings, your military operations, and increasingly, your intimate choices. You may not notice them because they are being woven into the normal functioning of institutions you already trust. You may not be able to refuse them because they have already become the condition of access to basic services. And you may not be able to remove them once they are integrated deeply enough, because too many other systems depend on them.

Sarah's case is not unusual. It is, in many ways, optimal. She received good care from a good hospital with sophisticated AI systems designed to route her appropriately. She did not receive obviously harmful treatment. Her stent probably reduces her cardiac risk, even if less invasive options might have been equally effective. Her outcomes will likely be good. She will probably avoid a heart attack in the next year, and the care system will have helped her achieve that outcome.

But she also received routing that was determined by algorithms she did not know about, based on optimization criteria she did not consent to, through a chain of causation that no individual could see or stop. She was moved through the system quickly and efficiently, in the direction that the system was optimized to move her. And because the system moved her in a legitimate clinical direction, her experience gave her no reason to question the routing.

This is the template for AI integration: create systems that are technically sound, locally reasonable, and systemically optimized for values other than human deliberation. Wire them into institutions that people trust. Make them move faster than humans can think. Then watch as the entire infrastructure gradually reorganizes around the speed and efficiency that only AI routing systems can provide.

The question is no longer whether AI will be integrated into healthcare decision-making. It is already happening. The question is whether we will notice the integration before it becomes infrastructure—before removing these systems would break too many things, before the alternative of slower human deliberation seems impossible, before the speed of the system becomes so normal that we forget we ever made a choice to move this fast.