Why Healthcare AI Is Moving From Automation to Physician Augmentation

Pilots in healthcare tend to die when they ask doctors to work around the software. The tools that last are the ones built to fit the clinic day as it already runs, and then proven there over months rather than in a demo.

Healthcare AI usually doesn't fail in the lab. It typically fails a few weeks into a pilot, when a physician closes the extra tab and goes back to the inbox they already had. The model was fine. The day was not.

The reason the category is shifting, slowly, from automation to augmentation is boring. Automation promises to take work off a doctor's plate. Augmentation promises to stay on the plate, in the same tools, leaving the last decision where it already lives. One of those pitches survives a medical director. The other fills conference slides.

The numbers bear out the pattern. The Healthcare AI Adoption Index, published in April 2025, came from Bessemer Venture Partners, AWS, and Bain & Company. They surveyed more than 400 healthcare executives.

Only about 30 percent of completed generative AI proofs of concept had made it into production. Michael Meucci, chief executive of the health-data company Arcadia, made a related point in a June 2026 MedCity News commentary: AI "isn't consistently reaching the moments where decisions are made," he wrote. The bigger problem, he added, is workflow integration, not model performance. When workflow integration fails, the pilot ends, usually without a postmortem.

Why the Graveyard Is So Full

Independent practices rarely have the slack to nurse a pilot along. They cannot absorb a tool that adds two minutes to an already late clinic. If the software does not remove work from the day, it becomes work.

Adoption, not accuracy, turns out to be the harder test. A model can score well on a benchmark and still lose to a habit. The measure that matters is duller: whether a clinician is still using the tool a quarter after the vendor stops calling.

Abhinav Kejriwal came to the problem from business and technology, and he is happy to admit he wasn’t a researcher. He co-founded PreventiveHealth.ai, which builds physician-specific software for U.S. outpatient care: "The computer-science training and the years in newsrooms are useful here for opposite reasons: one for systems, the other for how people actually take in a message," he says.

His test for a pilot is blunt. If nobody uses it, it is not a product: "Customers care whether a problem is gone, not how clever the stack is. The gap between what people say they want and what they will actually use has changed how I think. The thesis most healthcare AI companies start from is that AI will replace doctors. That is the wrong frame," he says.

"A doctor without AI does not scale: hours, attention, memory are finite. An AI without a doctor is not credible. It can be fluent and wrong."

The line is becoming familiar. Kejriwal treats adoption as a matter of discipline: "Action is cheap now. Anyone can prompt a model. The scarce skill is asking the right question and sitting with the boring work of defining the problem."

What Changes When You Design Around the Day

The companies still standing in clinics, rather than in decks, tend to share a few unfashionable choices. They sit inside the browser and the electronic record the doctor already uses. They draft. They do not send.

The last click has to stay with the person whose license is on the line: "No answer goes to the patient directly," Kejriwal says. "The doctor always has a supervising layer."

The company applies the same rule outside the clinic. Its own case study of a multi-site functional-medicine client puts it this way: "AI was an analysis accelerant, not an autonomous decision-maker; every change was reviewed and confirmed with the client."

This is why Kejriwal believes DrKai, the company's clinician tool, has succeeded, because it is not an all-purpose assistant: "The idea is a personalized AI twin for each physician," he says.

In practice, a patient's question arrives through the portal or inbox, the software proposes an answer in that doctor's manner, and nothing leaves until the physician has reviewed it: "The doctor looks at it, they correct it if they want, and then they click yes."

He adds: "That click matters more than anything the model writes."

In his telling, the corrections are the adoption story: "The first versions of those drafts, doctors were rewriting almost all of them. About 95 percent," he explains. "That is now closer to 5 percent."

The workflow point is the other half: "If it is not in the workflow, it will not get used," Kejriwal says.

"The design does not ask a doctor to log into one more thing and abandon what they already use." He puts it more bluntly: "Every extra login is a reason to stop."

That design targets the mid-pilot drop-off. A tool that appears inside the message a doctor was already going to answer does not have to compete for attention. It only has to make the answer faster: "The draft is the easy part," Kejriwal says.

"The hard part is making sure the doctor never has to leave the place they already work to use it."

Where It Is Running

Kejriwal says that workflow is no longer a lab exercise: "This is running in real clinics, with real patients, with the doctor still responsible."

DrKai has been deployed in live U.S. healthcare settings. One of those is Forum Health, an integrative and functional-medicine network that lists more than 30 clinics on its website: "Forum Health has worked with PreventiveHealth.ai since October 2025," Dr. Shilpa Saxena, Chief Medical Officer of Forum Health, reveals.

"Our providers and staff use its physician and coach avatar platforms and its DrKai browser extension to answer patient questions across our community. In our own use, the early drafts needed real editing," she says.

"But after the first dozen or so questions, the clinicians were approving most of them with little or no change, and a patient question that used to take clinicians around 20 minutes now takes about two minutes. That is a major improvement in operations."

Kejriwal’s phrase for the standard is "customer ecstasy, not customer service.”

He says: "The test is whether a clinic is glad it adopted the tool, not whether a ticket got closed. A clinic that lets a tool into its patient conversations is taking a real risk. It is better for a doctor to tell you how it went than for the company to.”

Functional medicine makes sense as a starting point. A visit is rarely one procedure. It turns into a string of questions, follow-ups, and judgment calls: "Those are the questions that pile up in a doctor's inbox, and they are where a good draft saves the most time," Kejriwal says.

The relationship with that group also reaches past the inbox: "With one partner, the company built a business-intelligence layer that showed them their own operation at a granularity they had not had in 15 or 20 years of running the business," Kejriwal reveals.

"They have treated the company as part of the leadership conversation, not only as a vendor."

He adds: "The fact that this group let an outside team into strategy meetings is the strongest evidence so far."

Rami El Assal is Managing Partner and Founder of Boutique Venture Partners. He is also a Stanford Leadership Fellow who mentored Kejriwal. He describes the footprint this way: "It is already in U.S. clinics, a large functional-medicine network, and groups that include physicians affiliated with UT Health Houston."

He adds: "Abhinav sees the second setting as a harder adoption test because it sits outside the functional-medicine group where the workflow was first proven."

El Assal contrasts it with competitors: "Most of this field is trying to get the physician out of the way," he says. Of Kejriwal's system, he adds: "The system drafts in one doctor's manner. The doctor still has to approve what the patient sees."

El Assal mentored Kejriwal. By his own account, he has no financial stake in PreventiveHealth.ai. He does not use the software as a clinician. Forum Health now speaks for the first setting he describes. The second has been described only by the company and by El Assal so far, and neither deployment has yet been the subject of an independent outcomes study.

What Isn't Proven Yet

PreventiveHealth.ai is a young company, and while adoption can fade, the real test of a clinical tool is not the first quarter of use but the fourth.

By its own dating, Forum Health is about a year in. Two principles Kejriwal has articulated apply here: "Half-finished work is not a strategy," and "Own the outcome. Do not explain it away."

The numbers can't settle that question by themselves. A falling rewrite rate could mean the drafts are getting better. It could also mean physicians are checking less carefully: "If doctors stop reading the drafts, the rewrite number improves for the wrong reason," Kejriwal says.

"That is why the number alone cannot be the proof." Telling those apart takes independent chart review and outcome tracking, which is the evidence that would move this story from adoption to results.

What "Augmentation" Has to Mean

"Physician augmentation" is increasingly the polite way to say: we will not send this unsupervised. Augmentation that still lives in a side app will die the same death as automation that promised to replace the doctor.

"Nobody in a clinic asks for augmentation by name," Kejriwal says. "They ask for less time in the inbox and their own judgment still on the answer."

"The United States spends a very large share of GDP on healthcare and does not get matching outcomes. That is a systems failure, not a slogan," Kejriwal says.

"I do not imagine one company repairs it. I do think physician-centered tools and cleaner economics for

practices are a more honest place to push than tools that try to take the physician out."

That standard is measurable, and it cuts both ways. If the usage PreventiveHealth.ai reports holds up across more clinics and more quarters, it will be evidence that augmentation can scale beyond a pilot.

Kejriwal states the finish line in plain terms: "The shift from automation to augmentation will be won by the software a physician still has open on a Thursday after it was installed at the beginning of the week."

"Thursday is the test because nobody is watching," he says. "The launch is over, the vendor is gone, and the inbox is full."

For Kejriwal, the payoff shows up on physician day: "It helps them be a better version of themselves, because they are able to deliver better care with less stress," he says.

About the Author

Sashindra Suresh is an experienced writer specializing in artificial intelligence, software development, and emerging technologies. With a strong ability to translate complex technical concepts into clear, engaging insights, she has contributed to a wide range of publications and platforms. Her work focuses on making cutting-edge innovations accessible to both industry professionals and curious readers alike.