In the past few years, AI adoption to tackle administrative work in healthcare has surged, with AI scribes and tools to handle billing and paperwork leading the charge.
These "high-frequency, low-stakes use cases" are a logical starting point for health systems, with a tangible return on value for those investments, Shiv Rao, M.D., CEO and co-founder of Abridge, said during the recent 2025 Forbes Healthcare Summit in New York City.
And as health systems learn how to implement AI workflows for those low-stakes use cases, leaders will be more comfortable to expand to other applications and workflows, industry experts say.
"I think the barriers to entry for this kind of innovation are coming down so quickly, it feels like to me, and we're also seeing higher-stakes workflows get innovated upon very, very quickly with this technology," said Rao, who leads Abridge, an AI scribe company that uses generative AI to document doctor-patient conversations.
As a practicing cardiologist, Rao says he uses AI for clinical decision support when seeing patients and rounding on the cardiology service at the hospital.
"I'm not asking for permission from anyone to do that, but I always feel like it takes two hands to clap. I'm going back and forth with it still to get to a better place. It's not always in the right sort of zip code with its answer at the start, but where I get to ends up being a better place than where I was initially," Rao asserted.
He added, "I think we're going to see a lot of innovation. We're going to see a lot of these different use cases across that entire 'two by two' blow up very, very quickly. I think what we've been able to demonstrate, and I think what has allowed for the dam to break, is that those lower-stakes workflows are working," Rao said.
Vineeta Agarwala, M.D., Ph.D., also a practicing physician and a general partner at Andreesen Horowitz, sees the potential for AI solutions to help address care access challenges and what she refers to as a "pervasive bad triage problem."
"We are triaging patients with the wrong care providers. We are triaging patients to the wrong sites. We are not triaging patients in time. I think a lot about triage, triage that's always on, triage that's always available. That is the gateway for me to think about the ways in which AI could better triage our entire healthcare system," she said during the Forbes panel.
At the same time, more and more patients are turning to AI chatbots to research medical issues, and that's driving AI into healthcare at an accelerated pace, noted Agarwala.
As of this year, 34% of U.S. consumers reported having used ChatGPT, a number that has roughly doubled since summer 2023. Among people under the age of 30, that number is much higher, with 58% having used ChatGPT. Since its launch, ChatGPT has been used by 10% of the global adult population, according to OpenAI.
A recent survey found that 35% of Americans say they’ve used AI to research a health concern, and nearly half of 16- to 34-year-olds have turned to AI for health advice.
On the physician side, AI-powered medical search platform OpenEvidence claims that more than 40% of providers have used its technology.
According to the Rogers adoption curve, a model that explains how new innovations gain traction, the 20% mark indicates when adoption starts accelerating from niche to mainstream, Agarwala noted. AI in healthcare has surpassed that mark, she said.
Patients and even many providers are willing to use AI tools to synthesize medical information. For patients, it means getting faster answers to a health-related question to use as a starting point for a conversation with a doctor. For physicians, AI tools like OpenEvidence provide access to a comprehensive trove of medical literature at the point of care.
There is now an opportunity for healthcare providers to meet patients in the middle by integrating AI tools and chatbots into the front-end of care delivery.
"Look at even a high-stakes situation, and let's say you've got a pregnant mom. It's Friday at 10 pm, and the hospital does not have labor and delivery staffing to answer her questions. That mom is primed to use an AI tool, and if the hospital could figure out how to integrate that AI tool with their staff who is stretched thin, now you actually have both sides ready to participate," Agarwala said. "The consumer is ready, the provider is ready. There's both cost savings, which our healthcare system urgently needs, and patient impact. For me, the core consumer behavior patterns is actually what's driving adoption of solutions."
The healthcare industry is rapidly approaching a point where there will only be "laggards," Agarwala asserted.
"The reason laggards typically exist is when there's no financial or other economic incentive to not become a laggard. Even though the market, by many standards, has recovered and is doing well, healthcare is not. Hospitals, hundreds, are going to shut down next year because of the expiration of ACA subsidies, because of Medicaid eligibility requirement changes, and because of tariffs on supplies that they need to buy. Hospitals are facing an unprecedented level of economic pressure, which I think will make that laggard pool diminishingly small," she added.
And as patients rapidly adopt AI tools, clinicians and provider systems will be pressured to keep pace so they're not on their heels, Rao said. It will soon become commonplace for patients to show up to their doctors' appointments armed with AI-generated information about their health issues, he noted. But, he stressed, rather than being apprehensive about this shift, physicians should get more prepared for those types of interactions.
"I do still believe that clinicians have an ability to leverage this technology in a differentiated way and really be that sherpa for the patient and the family member to help guide them to whatever their true north is. But, if we don't figure out how to leverage these tools in our workflows very, very quickly, I think it could get very uncomfortable," he said.
While back-office tasks, such as provider credentialing, are considered more "low stakes" use cases for AI technology, the bar for quality is still high, Derek Lo, CEO at health tech startup Medallion, noted.
Medallion automates back-office tasks in healthcare using AI and robotic process automation, with a focus on provider credentialing, privileging and tools to manage provider networks.
"If we put someone's last name wrong, we add a dash, or we miss one character, it can mean that then the provider isn't going to get credentialed. It can then mean that patients aren't going to see that provider on their insurance's directory. They're actually going to not be able to go and do an important surgery that needs to happen," he said.
"LLMs (large language models) are fundamentally non-deterministic programs. When the bar for quality is so high, it really warrants a deterministic solution. Where we've been really leveraging AI has been in speeding up our ability to essentially deliver deterministic automation that we know is going to be 100% correct every single time," Lo said during the Forbes panel. "It's definitely been very valuable as a tool for us to be able to speed up how fast we're delivering these efficiencies. We are also using it to do things that just weren't possible before, things like calling an insurance payer programmatically that, two years ago, would have been really difficult and required only human labor."
Abridge, a vertical AI company, also focuses on setting a high bar for ambient AI in healthcare. The company continuously refines its proprietary AI models to improve accuracy and performance and to reduce hallucinations in AI-generated clinical notes.
"That might mean that we're doing a lot of distillation into our own proprietary models. It might mean we're fine-tuning a lot with feedback, or post-training, as they call it, with feedback from end users across the country," Rao said.
Abridge is now deployed in more than 200 health systems and approaching 100 million clinical encounters a year, he noted.
"That's a lot of feedback that we're getting. A big part of our challenge, from a tech standpoint, is building the loops that allow our technology to continually improve," he said.
He added, "For any given piece of output that we might generate, structured or unstructured, we have dueling agents who are deciding whether any given sort of output is a false positive. Is it a confabulation? And if there is a high probability, per these agents, of it being a false positive, and if it's very severe, then they'll just programmatically remove that utterance or that output from whatever the output ends up being, back into the medical record system, for example," he explained.
An analysis found Abridge's technology outperforms GPT-4o by a wide margin, catching 97% of hallucinated content, while GPT-4o only catches 82%.
"Healthcare is one of those industries where 82 versus 97 is meaningful," Rao noted.
While it's important to focus on eliminating risks and potential harms in the use of AI, Agarwala stressed that approaching AI with too much caution and moving slowly can also have downsides given the industry's significant workforce shortages and care access challenges.
"Sometimes we're comparing AI to perfect, but we do need to compare it to the real world, and sometimes the real world is nothing," she said.
She added, "Twenty percent of scheduled colonoscopies are no-shows. Why? Because people forgot that they need to do a prep. Is there a risk in using AI tools to inform them about the prep? Of course, but the reality was that they didn't get a call. The alternative was nothing. There's so much low-hanging fruit in healthcare that I worry that sometimes we're almost over-fixated on the harms, when we forget that the reality might be nothing."