'AI is never used to deny care' — on tech, trust, accuracy and prior auths

Artificial intelligence (AI) could make patient care faster, but doctors feel that’s not happening.

A full 61% of physicians believe AI is clogging the process of patient care by increasing prior authorization denials, according to results from an American Medical Association survey this year. The U.S. Centers for Medicare & Medicaid Services have introduced more prior auths in its new WISeR model. Short for Wasteful and Inappropriate Service Reduction, the added use of prior authorization has sparked criticism from Congress on down.

Cohere Health Chief Medical Officer Brian Covino, M.D., said physicians’ fear is understandable, but backward: Used correctly, he argues, AI should lower denial rates, not raise them. It might actually happen in 2027 when the U.S. health care system starts operating under CMS-0057-F, the CMS Interoperability and Prior Authorization Final Rule, which requires health plans to run four standardized electronic systems, including one built specifically for prior authorization requests.

Medical Economics spoke with Covino about how AI actually gets built into a health plan's decision-making, why he believes it can be more accurate than a human reviewer, and how much faster prior authorization can move once AI is involved.

This transcript has been edited for length and clarity.

What role does artificial intelligence play once it's built into the prior authorization system?

Brian Covino, M.D.: Artificial intelligence is an interesting term, because a lot of people don't truly understand what it is, and it has a broad meaning. What I would tell people is that if AI is used for a specific use case, it's going to be more accurate. In our platform, for example, we use it, but it's trained by our clinicians, and this is the most important thing for health care uses of any AI: Clinicians have to be in the loop, sitting side by side, arm in arm, with the technologists who build the systems. That's what makes it more accurate and more efficient. What we use isn't off the shelf, it has to be specifically trained for health care uses. Machine learning models can read documents much faster than humans, and if clinicians train those models to read medical documents, the models can not only do it much faster, they can actually become more accurate at pulling information out of health care documents than humans can. But I can't stress enough that you have to have clinicians in the loop when you're developing this for health care uses.

AI hallucinations and mistakes get a lot of attention, but humans make mistakes too. Which makes fewer: AI or humans?

Brian Covino, M.D.: From what we've seen in our models, and we do audit them on a regular basis, once they're fully developed and trained, the AI actually is more accurate than humans, because humans do make mistakes. If we can read medical documents in real time with improved accuracy over humans, using clinicians to develop it, that's the best of both worlds. If you use a combination of models trained for a specific use case by clinical staff, you don't see the hallucinations that get talked about with some of the off-the-shelf, standard AI tools that are available.

Related coverage: What CMS’ 2027 AI rule means for prior authorization (and maybe for fax machines)

An AMA survey found 61% of physicians worry AI could increase denial rates. Will insurers simply train AI to deny more prior authorization requests?

Brian Covino, M.D.: First and foremost, this is a key point everybody needs to understand: Artificial intelligence is never used to deny care. That's non-negotiable. What we've actually seen in our use of AI is that we've been able to lower denial rates, and I'll tell you why. A significant portion of denials in prior authorization is due to a lack of information needed to make the decision. For some reason there's not enough submitted with the request, the process goes back and forth, and you don't get it, and you finally have to deny it. Part of what AI can do, since it's reading the information submitted in real time, is identify that certain elements are missing. Say an MRI is needed prior to a procedure. The AI reads the submission, sees the MRI isn't there, and in real time can send that message back to the submitter — we appreciate your information, we've got everything else you need, but we don't see an MRI, could you please submit that now? That solves some of this missing-information problem. We've seen decreases of 30% to 40% in missing-information issues from these prompts. If it goes beyond that, we can prompt manually, or it can go to a physician-to-physician phone call. So, getting back to the point, using AI to read these health care documents and solve some of the missing-information problem, we've actually seen denial rates go down. That reduces delays for care, and it reduces the whole appeals process, which is onerous for everybody. It's also important because when a denial is sent out, the patient gets a letter, and they can interpret that as, my doctor is a bad doctor, because my care got denied, and that's often not really the case. So avoiding a lot of that back and forth, and some of those unpleasant experiences for patients, is important.

Compared with a near instant transaction like an Amazon purchase, how fast is prior authorization once AI is involved? Weeks, days, minutes?

Brian Covino, M.D.: There are a couple of elements of timing. First is the time it takes staff and the physician's office to submit. We've seen a 50% decrease in the amount of time it takes to submit information. They enter information in the electronic portal, upload documents. In the past, they used to have to answer a long series of very clinical questions, and these are often people who aren't clinically trained, so they'd have to parse through the physician's and clinicians' notes to find the answers. That was a very burdensome, long process. So up front, we've reduced the time it takes for staff and doctors' offices to submit.

Second, since AI can read the information in real time and give feedback — if we have everything we need and it's an appropriate case, which is often 85% to 90% of the time — physicians can get a response in close to real time: 30 seconds to a minute after they hit submit, they get an answer that the case has been approved. When we survey our physicians, they tell us that when they use Cohere, they can start scheduling patients more quickly for procedures; they're not waiting one or two weeks for an answer.

There have also been guidelines on turnaround times: Medicare used to have a 14-day turnaround requirement, which has been shortened to seven days this year. Many commercial plans are down to a three-day turnaround. So it's obvious why health plans are going to have to move to an electronic format to meet those time frames. Even for cases that aren't approved in real time and go to review, we can usually manage those within hours to a couple of days at most. Using electronic format and AI responsibly can shorten the time frames of this entire process significantly.

Could greater standardization from CMS and private insurers help streamline prior authorization?

Brian Covino, M.D.: I believe so, and that's also why, getting back to the AI issue, generic AI formats fall short in this area. They can't pick out the nuances. When we train our clinicians to train the AI, it can pick out the nuances and requirements based on what health plan it is, what line of business it is, because those often change fairly frequently. I think there's a move, I think AHIP, America's Health Insurance Plans, has helped with this, toward a mandate, so to speak, to determine what types of information are needed for certain cases and standardize that. That would certainly help, because it's very difficult for staff to remember what's needed across multiple health plans, multiple lines of business and multiple benefit plans, whether private insurance or government-sponsored insurance. The final area is medical policy: Even within Medicare, medical policies aren't always standard from one Medicare administrative contractor to another. As we move forward, standardizing some of those as well will help solve some of the problems with this whole process.