Beyond the hype cycle: An investor's guide to AI in clinical workflows

The model is the easy part. Where value will actually accrue in clinical AI.For years in intensive care, one of us heard the same sentence whenever he proposed changing how a patient was treated: this is not how we do things. It is probably the most expensive sentence in medicine, and it has outlived every technology sent to retire it.In 2020 it fell silent. Supply chains broke, familiar drugs vanished, and consultants who had resisted a new protocol for a decade adopted one in an afternoon. Telemedicine told the same story. A telemedicine company the firm Amit co-founded had backed years earlier found insurers unwilling to pay and hospitals unwilling to let doctors work off site. India had no rules for teleconsultation until guidelines appeared on March 25, 2020, the day the national lockdown began. Within months, telemedicine was ordinary. Nothing about the medicine had changed; the price of the alternative had. A workflow moves when standing still becomes more expensive than moving, and almost never before.That is the variable the clinical AI debate keeps missing. Capability is no longer scarce, since most clinical AI companies rent the same handful of frontier models and the distance between the best of them and the fourth-best is measured in months. What is scarce is a workflow willing to change, and value that stays where you put it. We have sat on both sides of this pitch. This is how we tell the two apart.Start with the minute. Ambient scribes, which listen to a consultation and write the note, are the most widely adopted clinical AI product, which makes them the honest place to look. A JAMA study published in April followed more than 8,500 clinicians across 5 health systems. Adopters saved about 16 minutes of documentation for every 8 hours of patient time, visit volume rose by half a visit a week, worth roughly US$167 per clinician per month, and after-hours charting did not move. The minutes went into the inbox and into chart review, which may be where they belonged. Scribes do leave doctors less depleted, and in a shortage that is worth a great deal, but it is a retention story and should be priced as one. A minute that cannot be traced to a patient seen, a clinician retained, or a claim paid evaporates, and so eventually does the renewal.The clearest money has instead come from billing, because richer notes support higher codes. The Peterson Health Technology Institute found top-level visits rising 5% at one health system, worth more than US$1,000 per provider per month. Insurers noticed, and several now use AI to downcode high-level claims at scale unless the documentation plainly supports them. Two sets of algorithms argue over the same patient record, and the Institute concluded that AI is not yet lowering system-wide costs and may be raising them. Revenue that depends on out-arguing the counterparty is a lead, not a moat. An arbitrage against a payer is a transfer, not a creation, and transfers get repriced. In India the point is sharper, because the counterparty is usually the household, and a family paying out of pocket cannot be out-argued by better documentation. It can only be persuaded, or it walks away.Then there is the tax every product levies in clicks, logins, and altered routines, which doctors resent more than any group we know. In a randomized trial at UCLA, physicians given a scribe used it in about a third of their visits. A pilot, in our experience, is often how a hospital says no without hurting anyone's feelings. What survives demands the fewest new habits, living inside the note the doctor already writes and the thread the patient already uses, which across most of Asia means WhatsApp. The number worth watching is not benchmark accuracy but the edit rate, how often a doctor rewrites what the machine drafted. Accuracy tells you what the model knows; the edit rate tells you whether the doctor believes it, and belief is the only adoption metric that compounds.Belief also decides who the real incumbent is. In America it is usually Epic, which 71% of executives at Epic-based health systems say they prefer by default, and a product on Epic's roadmap is a feature with a countdown clock. Asia has no Epic, which founders mistake for open ground. The incumbent here is a prescription pad, a WhatsApp thread, a cash counter, and a waiting room with 40 people in it. About half the world's population gets 5 minutes or less with a primary care doctor, and in India closer to 2. In America, clinical AI is sold as a cure for burnout; in much of Asia, the disease is the queue. The only question worth asking about a product here is whether it fits: voice notes in the language spoken at home, records that live partly on paper, families who decide together.One conviction falls out of all this. The largest opportunity in clinical AI sits outside the visit. The visit is where healthcare gets billed; the weeks between visits are where it succeeds or fails. A gastroenterologist in Texas reckons only 3 of every 8 procedures he recommends actually happen, and a specialist in the Dominican Republic answers his patients on WhatsApp until 11 at night. Both describe the same gap, and no scribe will close it, because it opens after the patient walks out of the room.That gap is widest where doctors are scarcest, which is why some of this decade's most consequential clinical AI companies will be built for Lucknow and Jakarta as much as for Boston. They will make a trusted doctor go further rather than try to replace one, and get paid for outcomes nobody on the other side of the table has reason to claw back. Their achievement will look small, and it will be enormous: a clinic, a few million times over, saying this is how we do things now.==============Dr. Amit Varma is the co-founder and managing partner of Quadria Capital, a healthcare-focused private equity firm investing across South and Southeast Asia.Abhinav Kejriwal is Cofounder & CEO of PreventiveHealth.ai, which, among other things, builds AI digital twins of physicians.Get the latest Business News and Live updates. Download the TOI app.