Somewhere in the home of every family that has passed through a serious illness, there is a binder. It holds the lab results, the referral letters, the printouts, the notes scrawled during appointments that went too fast. The binder exists because the medical system, whatever its other gifts, will not remember you from one visit to the next. So somebody in the family becomes the memory.
In the fall of 2025, an Atlanta psychiatrist named Munjal Shroff found himself in need of a very good binder. His wife had developed rashes, itching, fatigue and swollen lymph nodes. They consulted multiple doctors and went through extensive testing, and nobody could tell them what was wrong.
Shroff had used general AI tools, he told Business Insider, but he wanted something built for healthcare, and he signed up for Healz AI — a chatbot, launched in beta last year, designed to help cancer patients navigate diagnosis, treatment and recovery.
At that point her doctors suspected an autoimmune condition such as lupus was the likelier explanation, with cancer still a possibility. Healz agreed; it did not leap to the dramatic conclusion. What it did was recommend an opinion from the University of Texas MD Anderson Cancer Center, and point Shroff toward a specific specialist there for her treatment. That visit produced the diagnosis: a rare T-cell lymphoma.
“I think that pivot of getting the second opinion from MD Anderson was just absolutely crucial,” Shroff said. “I think it ultimately could have potentially happened, but it would’ve taken much longer.”
To his credit, Shroff does not tell this as a miracle story. Her local doctors were helpful, he says, and he believes they would have eventually sought the same second opinion without any software involved. What the app changed, in his account, is that it encouraged the family to advocate for the referral sooner, and spared them additional delay and uncertainty. The machine did not find the cancer. The machine said: keep asking.
The pattern held through treatment. Shroff consulted the AI about chemotherapy options, and it suggested one he could raise with her doctor. The doctor agreed it was the right course. Shroff concedes the doctor may well have recommended it anyway — but he walked into the appointment prepared, and left reassured. “I think it builds comfort,” he said, of the moments when the doctor and the software agree.
Healz’s founders and co-CEOs, Igor Demishev and Alex Plat, are careful to say the app is meant to supplement human doctors, not replace them. It can help patients understand a diagnosis, find oncologists, interpret scans and tests, and navigate recovery. They focused on cancer because it is common and the cases are complex, and because the disease has touched their own lives — several members of Demishev’s family have had it. “We are working with people who have insanely hard stories,” Plat said.
Underneath, Healz has no foundation model of its own. It layers existing models with a proprietary memory system and workflows built to digest and store large volumes of medical files — one user, the founders say, has uploaded 1,500 documents. It is also engineered against the chatbot’s worst habit, agreeableness: a “root-cause mode” makes the model ask up to 20 follow-up questions about symptoms before it will conclude or suggest anything. They have built, in other words, a chatbot instructed to be slightly difficult, which may be the most useful thing a chatbot can be.
The business figures, supplied by the founders, are of the early days: after a limited launch, recurring revenue doubled each month from June to August, reaching an annual recurring revenue of $100,000 as of early September. Doubling from a small base is, admittedly, the easiest kind of doubling. The user base is 97% cancer patients, the rest mostly doctors using it as a reference tool; most users open the app daily or weekly, and month-to-month retention stands at 70%. People do not churn out of an app that is holding their binder.
The startup’s backers include Alexey Dosovitskiy, a former Google researcher who authored a breakthrough computer-vision paper in 2020 and cofounded Recursive Superintelligence — a self-improving AI startup, founded around the start of this year, that has raised over $650 million at a $4.65 billion valuation. On one side of this story, then, a company annualizing $100,000; on the other, $4.65 billion of confidence. Between those two numbers sits the entire present moment in artificial intelligence.
“Now is the time to work on this: the underlying technology is largely there, and it needs to be made into products that help people,” Dosovitskiy told Business Insider in an email, saying healthcare is where AI can improve and save lives.
And then there is the epilogue that makes this an AI-era story rather than merely a medical one. After months of heavy use, Shroff decided to invest in Healz himself; Demishev and Plat say more fundraising is planned in the coming months. The satisfied customer has become a shareholder, which does not make his story false, but does make it the oldest kind of advertisement: the testimonial with a financial interest. Reader, note it, and then note also that everything verifiable in his account checks out modestly — the software nudged, the humans at MD Anderson decided, the doctors did the medicine.
Shroff says the AI has real limits. It has occasionally been off on specifics of her treatment and monitoring, which is why he considers it crucial to use it alongside doctors who can see firsthand what is happening. His wife is still in treatment, and he still relies on the app — and believes it would be worth even more to patients and families without medical training of their own. “It’s so easy to get lost in being a patient,” he said, adding, “It’s been an indispensable tool.”
Which returns us to the binder. For generations the binder’s job was memory and persistence: hold the records, keep the thread, do not let the question drop just because the system is tired and the appointments are short. Healz, on the evidence here, is a binder that nags. The humans did everything that mattered; the software simply refused, politely and at length, to close the file.