The medical community’s comfort with deploying AI in clinical care is rapidly evolving — because it has to, according to health informatics leaders at NYU Langone Health.
They said that AI agents will likely be performing clinical tasks completely on their own — with no human in the loop — in the near future. Take blood pressure medication titration for example.
“We already have an AI assistant we built for our home blood pressure monitoring program — that right now still has a human in loop for doing the titrations of the meds. Five years from now, we’re not going to have a human doing those titrations,” said Dr. Devin Mann, senior director for informatics innovation at NYU’s Center for Healthcare Innovation and Delivery Science.
Dr. Paul Testa, NYU’s chief medical information officer, agreed, saying “there’s no reason to.”
In his eyes, hypertension management is a clear example of where full automation makes sense. Under current care models, getting a patient to their target blood pressure can take six to nine months, largely because of slow, incremental medication adjustments that require repeated interactions with the health system and its human clinicians.
But those steps, Dr. Testa said, follow well-established clinical guidelines and rely on objective home blood pressure data — making them well suited for AI-powered decision making.
Full automation could also significantly improve a patient’s “time to therapy,” Dr. Testa added. Patients typically experience a delay between diagnosis and effective treatment, and this period is often unnecessarily long — not because clinicians don’t know what to do, but because the healthcare system moves slowly, he explained.
AI could shrink that window by automating routine steps like data review, guideline-based decisions and patient follow-ups to reach the right treatment faster, Dr. Testa stated.
He also pointed out that there are some clinical workflows that no longer require human interpretation, such as diabetic retinopathy screening. The rate of screening for this disease remains low nationwide, hovering around 15% — but with full automation, Dr. Testa argued that those rates could approach 100%.
Screening rates remain low because the process still depends on a series of manual steps — ordering the test, interpreting results and placing referrals — each of which introduces friction and opportunities for delay. Fully automated screening and referral could eliminate those handoffs and ensure eligible patients are identified and routed to care consistently.
Dr. Mann emphasized that this push for full automation isn’t just about efficiency or speed — it’s about the fact that the workforce to deliver guideline-recommended care simply doesn’t exist.
Clinical guidelines often call for far more lifestyle counseling and ongoing support than health systems can realistically provide, he noted. In areas like nutrition and chronic disease management, the number of clinicians required would be orders of magnitude higher than the workforce that’s actually out there.
“There’s a missing workforce that [AI] will just step into. We’re never going to hire 50,000 dietitians. They don’t even exist, let alone the fact that the reimbursement is not really there for them. So [AI] will, I think, create roles that we always wanted to be in there with humans, but the humans just aren’t there,” Dr. Mann said.
He also pointed out that human effort should shift to relationship-based and complex care. As routine work is automated, clinicians could spend more time on patient education, shared decision making and edge cases — areas where persuasion, trust and nuance still matter and where AI struggles.
Taken together, Drs. Mann and Testa see a future in which fully autonomous AI is not a fringe experiment, but a practical response to the realities of modern healthcare.
Photo: ThongSam, Getty Images
