by Kaushal Kulkarni, M.D.

“Will AI replace doctors?”
“Will algorithms take over clinical decision-making?”

These are the questions I keep hearing on panels, in the hallways, and at lunch at conferences. But they’re also common on podcasts, on blog posts, in investor research memos, and across LinkedIn threads and comments with thousands of impressions.

I get why it’s a seductive question - there are very few topics that are as hot right now as the promise and potential of AI, and flaws and failures of healthcare. It’s only natural that putting the two of them together is sure to grab attention.

It's a reasonable question too: the U.S. healthcare system is predicted to have a shortage of approximately 100,000 physicians within the next decade (AAMC, 2023), and 63% of doctors report burnout (Shanafelt et al., Mayo Clinic Proceedings, 2022). [Sidenote - a better question is, who are the 37% of physicians who are not feeling burnt out? I would like to meet one of them one day].

I understand the hype around the question as well, especially given some of AIs accomplishments in diagnosis and clinical decision making. IBM Watson Health’s early claims of outperforming oncologists in treatment recommendations and Google’s 95% accuracy in detecting diabetic retinopathy only scratch the surface.

But as a physician (and even as an ophthalmologist), I still think we are asking the wrong question when we are asking whether AI will replace doctors or clinical decision making.

Why?

Physician Burnout Isn't Driven by Patients

This may sound crazy or novel, but it’s usually the non-clinicians who are the ones getting excited about AI’s power to replace doctors. Most physicians will tell you that, for the well trained, diagnosing patients is the easy part; AI’s biggest opportunity is handling everything before, during, and after the diagnosis.

For physicians, the frustrating part that leads to loads of provider burnout is all of the complex administrative, burdensome, mind-numbing, time-consuming pre-doctor and post-doctor work that takes place in hospitals and offices around the country surrounding the actual diagnosis of the patient.

Everything from fielding patient calls, triaging, checking benefits, scheduling, new patient onboarding, medical record retrieval, review, and chart prep, medication reconciliation, scribing, care coordination and referrals, prior authorization, to patient education - the list goes on. It is said that these administrative tasks consume, on average, 2 hours of admin time for every 1 hour of patient care (Annals of Internal Medicine, 2016). But the real truth is that even within that 1 hour of “patient care”, at least 80-90% of that is actually still administrative work. In fact, if you examine it, most of what the public calls “healthcare” and most of the time that patients spend with their doctor is actually administrative, not clinical (hence why your doctor spends most of your visit time facing the computer instead of you). The end result of this is a system that leaves doctors frustrated and rushed, and leaves patients angry and waiting.

If all of this pre-doctor and post-doctor work were automated away with AI, I suspect there would be a material impact on physician burnout, most physicians would love their jobs, and the patient experience would be much better. Automating the administrative parts frees up the few doctors we have left to do what they originally wanted to do, which is simply to diagnose, treat, and care for people. Those people, in turn, would continue to trust their doctors (not a robot), and recognize that the doctor is well-informed and armed with the full picture of their medical history. Does it really instill much confidence when you have to regurgitate medical history to the provider you were referred to? Shouldn’t they be able to access it and see it easily?

But what’s happening behind the scenes is truly astounding – medical records today are irreparably fragmented (more on this in the future). Healthcare generates approximately 30% of global data—growing faster than any industry (IDC). And yet, there is and will never be a magic API that connects all healthcare data. On the contrary, much of this workflow is still manual. Nobody under the age of 40 has ever seen a fax machine - unless they work in healthcare. The healthcare industry is probably single handedly keeping the fax machine industry alive as records are faxed from one provider’s office to the next.

New providers receive faxes, which sometimes contain 200-page PDFs, and before making a decision, someone has to make sense of all that information. The sense-making, whether in the context of clinical care or clinical trials, is still a manual, inefficient, burdensome, time-consuming, error-prone task.

But AI won’t scale in healthcare without trust. 75% of physicians surveyed by the AMA believe that AI tools can help with efficiency, but only 35% feel more excited than concerned about increased usage of AI. We have to be able to close this gap. The answer is using models trained on clinically relevant data, validated in real workflows, and designed to augment—not replace—clinical judgment.

The future of AI in healthcare is not clinical decision making, it’s amplifying clinical decision making by automating the manual tasks required before and after those decisions are made. A future where AI aggregates  records instantly, cutting retrieval time by 80% (weeks to days). Then, natural language processing extracts insights from unstructured notes, flagging critical details in <10 minutes vs. the 60+ minutes it takes to do so manually, saving 70% of prep time and 50%+ on costs (AMA data).

A Better Conversation

It’s time to move beyond the “AI vs. doctors” narrative. It’s reductive, distracting, and doesn’t serve the people who matter most: patients. What if instead, we could automate the manual tasks that take the vast majority of provider time, and providers could spend 80% of their time on patients, not data.

The real opportunity is this: AI that helps clinicians be faster, better informed, and more present with their patients.

That’s not science fiction. It’s happening now—and we should be talking about it.

Sources: AAMC (2023), Mayo Clinic Proceedings (2022), Annals of Internal Medicine (2016), Statista (2023), Health Affairs, IDC, AMA surveys

‍