AI Knocked Over Medicine’s First Domino: Two Are Still Standing

Alexander Sheppert, DO, PhD, MBA, is an internal medicine resident physician and AI researcher and the founder of Matic.

​I came to medicine from software. Before medical school, I spent years writing code, and when AI got better at reading and writing language, the first obvious clinical use for it was scribing: A doctor talks with a patient for 10 minutes, the AI listens and a draft of the note appears. I built Matic, one of the first AI scribes, sure doctors would adopt it once it worked.

That bet was right. AI scribing is now one of the first mainstream AI tools in healthcare, used by nearly a third of physician practices, as reported by KFF Health News.

The FDA has authorized a growing list of AI-enabled medical devices. About $60 billion has gone into medical AI startups in the last decade, according to research from Flare Capital Partners cited by Healthcare Dive. The AMA reports that more than 80% of doctors now use AI of some kind.

Across the industry, I've heard venture firms and health systems asking the same question: Which AI products come next? Here is my answer.

Now that I practice medicine instead of only programming it, I see something I missed as an engineer. Scribing is one-third of the workflow every visit runs through. It was the first domino to fall, but the next two are ready to be addressed.

First, before a doctor enters a room, they review their patient's chart. Second, during and after the visit, they write their notes. This is the part of the process that has been automated. Third, the visit is coded and billed. Those three steps are the spine of an encounter, and AI scribing addresses only the middle step.

The first step is reading the chart. Most patients now carry years of records, and a doctor managing a complex case has to spend real time digging through those records. Much of what matters for treatment sits in earlier notes and old results.

I have seen how a single wrong entry can outlive everyone who was involved in that entry: A result recorded incorrectly once can then be copied forward, note after note, until it reads as settled fact. An AI briefing tool can read the whole record, not just the latest note, and surface that original entry and flag that everything written since contradicts it.

The third step is coding and billing, and the cost of getting it wrong is large. The U.S. diagnosis code set, ICD-10-CM, encompasses about 70,000 entries. No doctor can keep all of those entries in memory. Doctors may pick the first code that looks close to their case, even when a better one exists. They may also assign codes that charge less than they should, a pattern called underbilling, because they lack the time and training to find the most accurate code.

Some large organizations answer this problem with dedicated coding teams that audit charts and file bills. This is one of the most expensive bureaucratic layers in U.S. healthcare, costing an estimated $40 billion a year in billing and collections alone, according to a McKinsey report cited by the American Hospital Association.

A language model has no such limit. It can work across all 70,000 diagnosis codes and the thousands of procedure codes at once and propose the best-supported one in seconds. Insurers, of course, can use the same tools to audit the same notes.

Believing all three of these steps should be addressed together does not guarantee success across the board. A briefing that misses something in a patient's history can create liability, and unlike a scribe’s note, which a clinician signs line by line, a briefing tends to be trusted without that scrutiny. Automated coding also invites payer audits.

Chaining all three processes together is harder than addressing any one individually: A mistake early travels across the whole chain. If records are long and messy, or if the chart sits inside systems outside of your control, a new tool may have a difficult time securing the right sort of access. The bar for success is high here, and most attempts will not clear it.

Twenty years ago, none of these jobs looked the same. A doctor trained in the 1990s might remember a thin paper folder, a few handwritten lines for a note and a short slip for a bill. Three laws changed all of this.

The 2009 HITECH Act paid clinicians to adopt electronic records, and within a decade, adoption grew to more than four in five doctors. The 2010 Affordable Care Act tied reimbursement to the diagnoses captured in the chart. The 2015 ICD-10 transition expanded the diagnosis codes from roughly 14,000 to about 70,000.

Together, these laws produced the workflow doctors now work within.

Right now, these three steps are split among people who do not share what they each know, and errors live in the gaps between them. You might see an incorrect code because a note was insufficient, and that note was insufficient because a chart went unread. These steps draw on the same records, so a tool that does one well is already part of the way to accomplishing the other steps.

So here is my next bet: Scribing came first because it was the easy domino. The machine only had to listen. The other two steps ask the AI to make judgments based on history, which is more difficult. But within a few years, the AI scribe solution practices just adopted will stop being just a scribe. That scribing function will become one feature in the solution's arsenal.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?