For two decades, technology has expanded patient agency over information.

Search put medical knowledge in the browser. Wearables put physiology on the wrist. LLMs put explanation on demand.

All of that mattered, but none of it changed where medicine begins.

Patients became better informed participants in institutional care. They could read more, track more, and ask better questions. But the choreography around the patient stayed the same. The system initiated. The system designed. The system decided.

I spent part of my career building that first era. At Apple, we worked on the premise that if people had direct access to data about their own bodies (heart rate, activity, ECGs), they would be empowered to make better decisions. With LumiHealth in Singapore, we pushed the same idea at population scale: better information would create better participation. It did. But it left the underlying structure intact.

What is changing now is not that AI can cure disease. It is that patients and caregivers are starting to seize the first move in care design and orchestration.

In a small but growing number of cases, patients can sequence tissue, interpret variants, identify targets, line up collaborators, and assemble bespoke routes to intervention. Not because the tools are new, but because AI has made it possible to navigate across them without specialist training in each one.

Institutions still control execution. They still run the lab, perform the surgery, manufacture the therapy, and absorb the liability. But they are no longer the only place where therapeutic initiative begins.

That is the shift. The first wave of patient empowerment was about information. The next wave is about intervention.

The Stack

Sequencing has been cheap for years. Protein structures have been public for years. Genomic datasets, variant databases, biological literature — open and free for over a decade.

None of it was enough. A motivated outsider could access each tool but could not use them together. Each one lived inside a different specialist domain with its own vocabulary, its own software, its own assumptions. The workflow connecting sequencing to interpretation to target identification to therapeutic design was locked behind years of cross-domain training. The tools were available. The seams between them were not.

Since last year, AI has broken the seams.

Not by discovering new biology. By making it possible for someone without specialist training to move across domains that were previously siloed by expertise. Interpret variant calls. Navigate an analysis pipeline. Chain tools together. Reason about what the results mean for intervention. Synthesize unfamiliar diagnostics. Run literature reviews in hours. Keep pace with a disease that moves faster than any single expert’s knowledge.

AlphaFold is part of this — it turned structural biology from an institutional bottleneck into a public utility. Scientists are using it to inspect mutated proteins and ask what is targetable. That question used to require a funded lab. Now it requires a browser.

The tools were already there. AI is what made them usable together. And that is what unlocks the first move.

The First Move

One of the clearest recent examples is Paul Conyngham in Sydney. Faced with losing his dog Rosie to cancer, he worked with the Ramaciotti Centre for Genomics at UNSW to sequence healthy and tumor DNA samples, built what UNSW described as a molecular picture of Rosie’s cancer despite having “zero bioinformatics training,” and pushed toward a personalized mRNA vaccine with academic collaborators. AI was what let him cross between domains he had no training in — from raw sequencing data to variant interpretation to protein structure analysis to vaccine design — each a different specialty, navigated through a single interface. Rosie also received a checkpoint inhibitor, so this is not clean proof about which intervention drove the response. That isn’t the point. The point is that Conyngham did not just research the disease. He initiated and coordinated a bespoke therapeutic pipeline.

At the other extreme is Sid Sijbrandij. After his osteosarcoma returned and standard options ran out, he quit his day job and, in his words, went “founder mode” on his cancer. He pursued “maximal diagnostics,” published 25 terabytes of data at osteosarc.com, used single-cell sequencing to identify high FAP expression, went to Germany for targeted radioligand treatment, and later said that those efforts helped take him to no evidence of disease after radioactive treatment and surgery. He did not just ask for more options. He built what he calls a therapeutic ladder that went from zero treatments to 30. His team used AI differently than Conyngham — not as the sole bridge across domains, but as an accelerant for a team that already existed. Jacob Stern describes using AI to get up to speed across unfamiliar diagnostics, structure bioinformatic analyses, run literature reviews, and bring back summaries fast enough to keep up with a live cancer problem. AI augmented their coordination. For Conyngham, AI was the coordination. Sid is an outlier. He has unusual resources, unusual stamina, and unusual access. He is not a template. He is a leading indicator.

Then there is Baby KJ, the infant treated at CHOP and Penn with a personalized base-editing therapy developed and delivered within six months after birth. KJ is not evidence that patients can now run medicine outside institutions. If anything, it shows the opposite: the decisive work still happened inside elite clinical and research settings. But it proves something important. Bespoke intervention is no longer hypothetical. The technical stack is real. The timelines are compressing.

Taken together, these cases do not show the end of institutions. They show the beginning of a different boundary. Patients and caregivers can increasingly start the work, even when institutions still have to finish it.

Three Modes of Therapeutic Agency

This new agency is not one thing. It is emerging in three distinct modes, each with a different relationship to institutions — and a different role for AI.

Orchestrated agency is the high-resource version. A patient builds parallel infrastructure around the formal system and uses institutions as execution partners. This is Sid’s mode. AI augments a human coordination layer — it makes an existing team faster, helps them synthesize across domains, and keeps pace with a complex evolving problem. The limiting factor here is not knowledge or even AI capability. It is tissue access, sponsor outreach, manufacturing relationships, hospital process, legal paperwork, and relentless institutional negotiation. And let’s be honest, this requires real resources and time that most people today do not have.

AI-assisted agency is the version most likely to spread. A technically capable patient or caregiver uses AI to compress across domains and gets far beyond where they otherwise could. They still need clinicians, labs, and institutions. But they arrive with hypotheses, options, and a plan. This is Conyngham’s mode. Here, AI does not augment existing coordination — it replaces the coordination layer that would otherwise require a team. One person, navigating across specialist domains through a single AI interface, doing work that previously required multiple experts. It does not require billionaire resources. It does require initiative, literacy, and willing partners along with a subscription to a consumer LLM to unlock the knowledge to create a plan.

DIY agency is the volatile version: self-experimentation with minimal institutional mediation. Peptides, gray-market compounds, improvised protocols, online communities. Here, AI is often the only guide — no institutional partners, no expert validation, just a person and a model and a compound. This grows whenever the gap between what is technically possible and what is institutionally accessible gets too wide. It is also the most dangerous form. AI can break knowledge seams, but it cannot validate safety, catch edge cases, or absorb consequences. When something goes wrong, the patient absorbs the full blast radius.

These are not three stages in a neat progression. They are three responses to the same structural fact: AI is diffusing the capability to start therapeutic work faster than institutions are redesigning around it.

Coordination Inequality

The deepest change here is not that medicine is being disintermediated. It is that therapeutic initiative is being redistributed.

For most of modern medicine, the patient sat at the end of the pipeline. Institutions discovered, validated, selected, and delivered. The patient could consent, refuse, comply, or seek a second opinion. Even the most informed patient still entered a process someone else had already defined.

That begins to change when patients and caregivers can make the first move.

They can decide to sequence sooner. They can bring their own analysis into the room. They can surface a target before a local clinician does. They can identify a nonstandard path, line up outside experts, press for tissue release, and maintain a live decision tree that updates as new data arrives.

They may be wrong. They may overfit noise. They may drown in false hope. But they are no longer purely downstream.

The next health divide may not be who can access information. It may be who can coordinate action.

Who can get tissue released quickly. Who can line up the right assays. Who can synthesize the literature into a decision tree. Who can find a manufacturer, a willing investigator, an expanded-access path, or simply a clinician prepared to engage seriously with a patient-generated plan.

Call it coordination inequality.

AI broke the knowledge seams — the barriers between specialist domains that kept non-experts from navigating therapeutic workflows. But it cannot break the institutional seams — the tissue custody battles, the IRB delays, the hospital processes that lack procedures for patient-initiated care, the manufacturing requirements designed for population-scale drugs applied to single patients. Coordination inequality lives in that gap: between what AI now lets you figure out and what institutions still will not let you do.

That is why the exceptional cases matter. They show what becomes possible for the people who can build a parallel operating system around their disease.

The temptation is to treat Sid as privilege and Rosie as novelty and move on. That misses the point. These are both a privilege gradient and a structural signal. AI is lowering the cost of understanding faster than anyone expected. It is not yet lowering the cost of institutional coordination fast enough.

Where This Leads

The future here is not that every patient becomes a biohacker.

It is that every serious diagnosis acquires a patient-side operating system.

Records become machine-readable by default. Tissue and data move with the patient. AI copilots synthesize literature, diagnostics, and treatment options into a live, auditable map. Molecular profiling becomes easier to trigger and easier to interpret. External expert review becomes normal. Experimental paths move into supervised sandboxes instead of forcing people into a choice between passivity and improvisation.

In that world, the patient is not practicing medicine alone. The patient is helping initiate and coordinate it.

That changes the role of everyone else.

Clinicians become less like sole navigators and more like expert partners and risk managers. Hospitals become places of high-trust execution, not the only places where the process can begin. Regulators face pressure to evaluate platforms, protocols, and supervised n-of-1 pathways, not only fixed products. Pharma has to decide whether it is building therapies for patients or with them.

And it broadens the surface area of AI’s impact in healthcare. When Isomorphic Labs or CZI talk about using AI to cure disease, they are mostly describing AI breaking seams insider the institutional pipeline, making drug discovery faster for pharma teams. Important work. But that may not be the whole story. The more radical shift is AI breaking the same seams for people outside institutions — giving patients and caregivers the cross-domain navigation that lets them make the first move. Same technology. Fundamentally different locus of agency.

Some institutions will treat this new patient initiative as noise, nuisance, or threat. They will defend the old monopoly on the first move.

The smarter ones will do the opposite. They will make it easy to release tissue and data. They will treat patient-generated analysis as input, not insult. They will create fast paths for supervised experimentation. They will separate legitimate safety guardrails from inherited friction. They will understand that AI has already broken the knowledge barriers. The institutional barriers are theirs to remove.

Because patients are not going back to being passive.

If the front door stays closed, more people will keep using the window.

That is the real danger. Not too much agency, but badly distributed agency. A world in which wealthy patients build parallel systems, technically fluent patients improvise with AI, and everyone else waits in line for a pathway that is too slow, too generic, or already exhausted.

The better future is not no institutions. It is institutions that know how to work with a patient who shows up not only with questions, but with a draft plan.

The Choice

The first era of patient empowerment let people read the chart.

The second may let them help write the plan.

We are still early. Most patients cannot do what these frontier cases suggest. Many should not have to. The point is not to romanticize self-navigation through catastrophic illness. The point is to recognize that the old arrangement is starting to break. AI has made the tools to begin therapeutic work navigable by people who were never supposed to have access to them.

I believe medicine can respond in one of two ways.

It can build pathways that absorb this new patient agency: faster data access, cleaner tissue rights, better coordination, supervised experimentation, safer individualized care.

Or it can cling to a model in which initiative remains institutionally monopolized long after the technical basis for that monopoly has eroded.

If it chooses the second path, patients will not stop acting. They will simply act with less support, less evidence, and more risk.

This will not stay exceptional for long. The tools are too accessible, the information too available, the stakes too personal. More patients will make the first move in the coming years — not because they want to practice medicine, but because they are unwilling to wait for it.

The patient will see you now.

---