Recovery Intelligence: The Maturation of Virtual Care

What RecovryAI becomes when a million Virtual Care Assistants are deployed.

Recovery from acute surgical procedures is a high-risk window. Complications go undetected. Patients drift from their care teams. Health systems absorb the cost. A Virtual Care Assistant, an FDA-regulated and prescribed AI, closes that window. It engages patients daily, identifies emerging complications, distinguishes normal recovery from deviations in standard trajectories, and keeps patients connected to their care teams through the risky weeks. Run that capability at scale across millions of recoveries and it generates something new: recovery intelligence. The device is investigational today and not authorized for commercial use, and what follows describes where the work leads, not what it does now.

A recovery nurse looks at a photo of an incision on the fourth day after a knee replacement. The edges are red. That is information. It is true, it is specific, and on its own it does not say much. Is this the ordinary redness of healing, or the first hour of an infection that will put the patient back in a bed in three days. The nurse answers from training and from the few hundred recoveries she has personally seen. She is usually right, and she is working from a sample of a few hundred.

Now put that same photo next to a million others, each one tagged with the procedure, the surgeon, the protocol, the patient, and what happened over the following weeks. The redness stops being a single fact and becomes a position on a curve. This pattern, on this day, for this kind of patient, under this protocol, resolves on its own in 94 percent of cases, and the other 6 percent share three earlier signals that were already visible on day two. That is intelligence. It is the difference between knowing a thing and knowing what the thing means.

A town that turned bad luck into risk

In 1948, in the town of Framingham, Massachusetts, investigators enrolled 5,209 residents and began following them. The goal was to understand why so many Americans were dying of heart disease, which at the time was treated as a matter of luck and age. A person had a heart attack the way a person was struck by weather. Doctors had little to offer in the way of prevention because no one had assembled the record that would show what came before the event.

Framingham assembled it. The same people, examined every two years, measured across many dimensions at once, blood pressure and cholesterol and weight and smoking and family history, followed for decades. Out of that record came an idea that did not exist in medicine before. The investigators called it a risk factor, a term Framingham coined in the early 1960s and that the field has used ever since. High blood pressure was no longer a normal part of aging to be left alone. It was a measurable signal that forecast an event years away, and once it could be seen coming it could be changed. Framingham moved a whole field from treating heart disease when it arrived to predicting it before it did.

The lesson is in the shape of what made it work. Not a smarter doctor looking at one patient. A defined population, followed over time, measured across many dimensions, large enough that patterns invisible in any single case became plain in the aggregate. Recovery has never had its Framingham. No one has ever followed a large population of surgical recoveries closely enough, across enough dimensions, for long enough, to say what the risk factors of a recovery even are.

What a million recoveries know

A Virtual Care Assistant, deployed at scale, is both a clinical tool and a data collection engine. Every recovery it engages is a patient followed every day rather than every two years, reporting pain and mobility and wound condition and medication and sleep. Each data point carries more than what the patient reports. It carries the context it was collected in: the surgeon who prescribed it, the protocol they follow, the procedure type, the patient's demographics and geography, the facility where care was delivered. One recovery is a constellation of data points. A million recoveries, each tagged with that context, become a pattern.

Those tags are where the intelligence lives, because recovery is not one variable moving in a line. It has structure. A hip replacement recovers on a different curve than a knee replacement. A cardiac catheterization follows a different arc than a shoulder arthroscopy. A minor office procedure generates a different recovery signature than open surgery. The same procedure recovers differently across demographic and geographic cohorts. The same patient recovers differently under different protocols.

And the protocol itself traces back to a specific surgeon, trained at a specific program, who made a specific set of choices: when the patient may first shower, how the wound is dressed and for how long, when weight-bearing begins, how pain is managed in the first week. Two fellowship-trained arthroplasty surgeons will make different calls on all of these, and both will be defensible. Today no one can say which set of choices actually produces the smoother recovery, because no one watches recoveries at scale across surgeons and protocols. A followed population of recoveries can begin to answer that. It can hold a surgeon's protocol next to the recoveries it produced and next to the recoveries a different protocol produced for similar patients, and let the outcomes speak.

That is a question no single clinician can answer, no matter how skilled, because the answer does not live in any one recovery. It lives in the pattern across all of them.

Borrowed intelligence, and native intelligence

There is an important line here, and it is easy to miss. The Virtual Care Assistant we are building today harnesses powerful general-purpose models, constraining them through procedural guardrails and safety mechanisms to keep the system within its clinical domain. That intelligence is borrowed and carefully bounded. It is general knowledge applied one recovery at a time, under physician oversight. It is the right tool for the job the device does now, but it is not the same thing as understanding recovery.

Understanding recovery at the level of prediction is a different threshold. It requires a model trained on recovery, and only on recovery, built on the firsthand record of what a million recoveries actually did rather than the general world's secondhand account of what a hip replacement feels like on day six. A model built on that record would know things no general model can, because the knowledge was never written down anywhere to be trained on. It exists only in the recoveries themselves, and until now those recoveries were never collected at scale.

This is the move from triage to prediction. Triage is a question asked in the present. Is what this patient feels part of normal healing or the first sign of something wrong. It is answered after the signal has appeared. Prediction is the same question asked earlier, and asked in more dimensions. Given this patient, this procedure, this surgeon's protocol, this cohort, here is the recovery this person is likely to have, here is where the risk sits, and here is the day to watch. The first is a tool that responds. The second is a system that anticipates. The bridge between them is the data, and the data is the byproduct of the first tool doing its job.

The study that runs itself

Framingham cost decades and a fortune, and it had to. Someone had to fund it, recruit the town, pay for the visits, and wait fifty years. It was a study built alongside care, at great expense, precisely because the record did not otherwise exist.

Recovery intelligence would be generated by care that is already being delivered. The record is a byproduct of the Virtual Care Assistant doing the clinical work it was built to do. Every recovery it supports adds to what the system knows about recovery, at no additional burden to a patient or a nurse, because the engaging is the product and the learning is the exhaust. This is the second result from a single move. The first result is the one RecovryAI's pivotal trial, 2025 through 2026, is measuring: a patient engaged through the risky weeks and a nurse freed for the cases that need her. The second result, building with each deployment, is the record that turns the next recovery from a thing watched into a thing understood.

The clinical tool does not stop being a clinical tool. The wedge stays the wedge. Recovery intelligence is what the wedge produces once it is running at scale, and it makes the clinical tool better in a loop, because a system that has understood a million recoveries reads the next one more precisely than one that has seen a few hundred.

What a health system plugs into

Run it forward and the shape of the claim comes into view. A health system that adopts the Virtual Care Assistant is buying, at first, relief. Fewer readmissions, lighter phone queues, recoveries covered that used to go dark after the two-week visit.

What arrives with it is larger. A surgeon, before an operation, could see the recovery this specific patient is likely to have, drawn from the recoveries of patients like this one under a protocol like this one. A service line could see which of its protocol choices actually move recovery and which are habit. A rural hospital could borrow the pattern learned from a hundred thousand recoveries it never had the volume to learn on its own. The intelligence flows back into care, under clinician oversight, informing the judgment rather than standing in for it, the same principle that governs everything RecovryAI builds. Adoption becomes a way for a health system to understand its own recoveries for the first time.

That is the story the pattern predicts. It is not the story yet. RecovryAI's pivotal trial settles the clinical question first, and the intelligence described here is where the work leads if the execution holds and the scale arrives.

The discipline this requires

Recovery intelligence built carelessly is surveillance. The same discipline that separates a regulated medical device from a consumer chatbot separates recovery intelligence built responsibly from recovery data scraped and sold. The record has to be governed and de-identified, held to the standard that regulated clinical data is held to, used to inform clinicians rather than replace their judgment, and generated with the patient's understanding of what their recovery contributes and protects. The value of the record and the discipline of holding it are the same commitment, not competing ones. A record assembled without that discipline would erode the trust that lets the record exist at all.

This is the harder, more rigorous version of the company, and it is the inevitable maturation of virtual care built with discipline. The Virtual Care Assistant earns its place one recovery at a time. What it becomes, once it has earned that place a million times over, is a system that understands recovery in a way medicine never has, because physicians have never had the capability to extend clinical care into the home through a prescribed AI that collects and learns from each recovery.

Regulatory status

Our Virtual Care Assistant is an investigational device, limited to investigational use under U.S. law, granted FDA Breakthrough Device Designation, with a De Novo Class II submission pending, and not authorized for commercial distribution. Everything described in this essay concerning recovery intelligence is forward-looking and describes a direction of development, not a current capability or a commercial claim. The evidence is still being compiled.

---

---