What is the most widely deployed artificial intelligence in U.S. healthcare today?

Not a diagnostic tool. Not a chatbot helping patients find a doctor. It's billing software. Revenue cycle management AI that automates coding, disputes claim denials, and captures reimbursement for every service delivered. Emergency room claims coded at the highest complexity levels jumped 7 percentage points in two years.

This is AI working exactly as designed — built to align with the current incentives of the system.

I spoke about this at the Business Group on Health Annual Conference during a plenary session on AI in healthcare. Alongside Gijo Mathew of Spring Health and Ozgun Ataman of Well, we addressed the HR and benefits leaders who guide healthcare decisions for some of the largest employers in America.

One message was clear. The question is no longer, Should we adopt AI? That decision has already been made. The real question — the one benefits leaders are uniquely positioned to address — is, Which AI is already operating within your benefits program, and whose interests is it designed to serve?

Healthcare AI serves two very different purposes with vastly different economic impacts, but these are often conflated in discussions. Here's what you need to know:

• Administrative AI handles billing, coding, prior authorizations, and claims processing. It automates manual tasks, reduces errors, and accelerates provider revenue capture — potentially increasing costs.
• Clinical AI focuses on care delivery. This includes triage, risk stratification, clinical decision support, proactive outreach, and diagnostic assistance. It aims to improve care quality and reduce overall costs by preventing costly events, steering members to high-value care, and closing care gaps early.

As a self-insured employer, your purchasing decisions directly impact care delivery. Recognize which AI solutions are in your benefits ecosystem and ensure they align with your members' needs.

AI adoption is already happening. Your members are finding their own way.

In 1961, a landmark study highlighted a lasting pattern. Of every 1,000 adults in a given month, roughly 750 experienced symptoms, but only 250 saw a physician. The remaining 500 managed illness entirely outside the formal system, and not always by choice. Often because access was too slow, too expensive, or too complicated.

Researchers replicated the study decades later. Same ratio. And it held through managed care, the Affordable Care Act, the proliferation of urgent care clinics, the telehealth boom. Whatever structural reforms the system attempted, the pattern didn't move. Most illness is still managed outside formal healthcare.

Unmet healthcare needs are now going to AI. Last December, Rock Health reported that 32% of consumers used AI chatbots for health information, double the percentage just one year before. OpenAI has reported 580,000 weekly healthcare messages from underserved regions, with 70% occurring outside clinic hours. People turn to AI when and where the traditional system isn’t available. Having insurance doesn’t guarantee access — and when access falls short, people improvise.

But their use of AI isn’t driven by confidence. In fact, public sentiment toward AI is poor. A Quinnipiac poll found 55% of Americans believe AI will do more harm than good. Only 44% think it will positively impact medical care, compared to 84% of AI experts.

This isn’t a contradiction; it’s a signal. Members seek AI because navigating a fragmented system alone is worse, even if they distrust the technology.

For employers, this highlights a key point: Excluding AI-forward vendors doesn’t shield your members from AI. Instead, it redirects them toward unmanaged consumer tools with no clinical oversight, benefits integration, or accountability.

During the panel, someone raised an important question about the AI arms race between providers and payers: providers using AI to maximize reimbursement, payers using AI to deny claims. The Peterson Health Technology Institute has documented these "bot wars," where automated systems amplify prior authorization volumes instead of streamlining them.

Despite this friction, I'm hopeful that over time this dynamic could evolve into something far more efficient.

When you swipe a credit card, your bank communicates with the merchant's bank in milliseconds, in a transaction entirely invisible to you. Suspicious charges get flagged, legitimate purchases clear instantly. No one builds a career around fighting the other side's algorithm.

Imagine healthcare payment systems functioning the same way: provider- and payer-side AI working together to swiftly and accurately resolve claims, eliminating the costly adversarial friction we see today.

The technology to achieve this already exists. What's missing is structural will. And that's where self-funded employers have a unique role to play by demanding innovation that minimizes waste and improves outcomes.

This brings me to the point I feel most strongly about: Healthcare should embrace outcomes-based pricing. While "value-based care" has become a loaded term with varying definitions, the concept is quite simple. You get paid based on outcomes.

In other industries, AI-native companies are already following this model. For example, some customer service firms only charge when their AI resolves an issue, not per interaction or per user. If the technology fails to deliver results, the vendor doesn't get paid.

Healthcare can adopt this same principle, if buyers demand it. Employers, as the largest healthcare purchasers in the country, can drive change by demanding AI that prioritizes members over hospital revenue. Their influence will force the market to deliver member-focused outcomes.