Artificial Intelligence in U.S. Health Care: What It Means for Patients and Providers

Table of Contents

• Key Points
• Background: Why This Research Matters
• Why Health Care Lags Behind Other Industries in AI Adoption
• Types of AI Used in Health Care
• AI in Reimbursement: Fixing the Billing and Insurance Maze
• AI in Clinical Operations: Smarter Hospitals and Operating Rooms
• AI for Quality and Safety: Protecting Patients and Improving Experience
• Clinical Implications: What This Means for Patients
• Limitations: What the Study Couldn't Prove
• Recommendations: What Patients Should Know and Do
• Frequently Asked Questions
• Source Information

Key Points

• AI in U.S. health care is still early, but it is being tested in billing, hospital workflows, and patient safety.
• AI can predict insurance claim denials, and one payer increased complex claims processed without denial from under 80% to over 90%.
• AI may improve operating-room use by 15–20%, based on simulations, potentially reducing surgical wait times.
• An early-warning AI system reportedly saved about 8,000 lives in one health network and reduced mortality across 21 hospitals.
• Most AI applications lack randomized controlled trials, so patients should ask about the evidence behind AI recommendations.

Background: Why This Research Matters

The adoption of artificial intelligence (AI)—machines or computing systems capable of making intelligent decisions—happens in phases across business sectors. The use of AI is already advanced in many areas, including reinventing how financial institutions provide investment advice, offering "recommendation engines" that suggest the next retail product to buy online, and developing driverless cars. But in health care delivery, AI remains in its early stages.

This gap matters because the need for AI to improve health care has become urgent. Consider the explosive growth of medical knowledge. In 1980, the collective body of medical knowledge required to treat a patient doubled every 7 years. By 2010, that doubling period had shrunk to fewer than 75 days. The numbers are striking: what medical students learn in their first 3 years represents only 6 percent of known medical information at the time of their graduation. Their knowledge may still be relevant but might not always be complete—and some of what they were taught will be outdated.

AI has the potential to supplement a clinical team's knowledge to ensure that patients everywhere receive the best care possible. Economic forces are accelerating this trend. The development of more robust AI algorithms from more data—an "economy of expertise"—has created a new subindustry called health care services and technology (including software and platforms, data analytics, and payment services), which could in the next few years be as large monetarily as the entire payer (insurance) subindustry is today.

The coronavirus disease 2019 (Covid-19) pandemic also served as a catalyst, prompting organizations to accelerate plans to digitalize and deploy AI. At management and board levels, recent public awareness of generative AI has increased conversations about AI. Additionally, AI adoption could have second-order effects, such as alleviating part of the ongoing shortage of physicians and nurses.

Why Health Care Lags Behind Other Industries in AI Adoption

AI adoption in health care delivery lags behind other business sectors for several specific reasons. Early AI took root in industries where large amounts of structured, quantitative data were available and computer algorithms could be trained on discrete outcomes—for example, whether a customer looked at a product and bought it or did not. Health care faces unique challenges:

• Qualitative information is harder to interpret. Clinical notes and patients' reports are generally more difficult for algorithms to process than structured numeric data.
• Multifactorial outcomes complicate training. Clinical decision-making involves many interwoven factors, making algorithm training more difficult than in simpler business contexts.
• Clinical workflow integration is challenging. Embedding AI output into an already complex clinical workflow is a major hurdle.
• Short-term financial focus. In the authors' experience, some health care organizations focus on near-term financial results at the cost of investing in longer-term, innovative technologies such as AI.

Health care organizations that prioritize innovation link investment decisions to "total mission value," which includes both financial and nonfinancial factors such as quality improvement, patient safety, patient experience, clinician satisfaction, and increased access to care. This broader view of value is essential for AI to flourish in health care settings.

Types of AI Used in Health Care

AI is broadly defined as a machine or computing platform capable of making intelligent decisions. In health care delivery, two types of AI have generally been pursued:

1. Machine learning: Computational techniques that learn from examples instead of operating from predefined rules. The algorithms identify patterns in data and improve their performance over time.
2. Natural language processing: The ability of a computer to transform human language and unstructured text into machine-readable structured data that reliably reflects the intent of the language. This is especially relevant for converting doctors' notes and clinical narratives into usable data.

The role of AI in improving clinical judgment has garnered the most attention, with a particular focus on prognosis (predicting disease course), diagnosis, treatment, clinician workflow, and expansion of clinical expertise. Specialties such as radiology, pathology, dermatology, and cardiology are already using AI in the process of image analysis. In radiologic screening, up to 30% of radiology practices that responded to a survey indicated they had adopted AI by 2020, and another 20% planned to begin using AI in the near future.

The potential of AI extends much further. The authors identified nine domains of health care delivery where AI use is emerging. Most organizations are still in the pilot phase of AI adoption and are attempting to validate the benefits. The three domains examined in detail are reimbursement, clinical operations, and quality and safety.

AI in Reimbursement: Fixing the Billing and Insurance Maze

Reimbursement—the system of checks and balances between payers (insurance companies) and providers (hospitals and doctors)—is key to the financial health of any health care organization. Uses of AI in this domain are both common and among the most advanced, with a higher-than-average total mission value. In what has been termed "the coding wars," AI has become an important tool for stakeholders to monitor one another, while also simplifying and reducing difficulty in the patient's experience with medical payments.

Understanding the Claims Process

Providers refer to processing insurance claims that payers should reimburse as revenue-cycle management. This task is often performed by staff who review a health care professional's billing and provide guidance on completing the bill in a way that aligns with the services provided. The goal is to ensure the amount actually paid to the organization is appropriate.

The claims process typically follows these steps:

1. Claims Creation: After a member receives a service, the provider fills out a detailed claim form, typically electronically.
2. Claims Submission: The provider sends the form to a payer; typically more than 90% of claims are submitted electronically.
3. Adjudication: The claim comes into the payer's inbound systems; typically more than 80% of claims are auto-adjudicated. About 20% of claims that must be manually adjudicated are generally resolved by claims processors or claims specialists.
4. Payment: A claims tool determines whether the claim will be paid, and reimbursement is initiated.
5. Claims Tracking: The provider and member can track claim statements online, and either party may dispute the claim.
6. Audits, Grievances, and Appeals: Teams audit claims monthly or quarterly to reduce errors and improve accuracy.

Here is a critical statistic: with the current system, more than 10% of claims are denied or delayed because of eligibility issues and missing data. But up to 85% of those denied claims could have been avoided.

Real-World Example: Hospital Uses AI to Predict Denied Claims

One large health system recognized that AI—specifically predictive analytics—could generate cost savings and improve not only cash collection and yield but also the patient's experience. The effort began with a large data set: 12 months of claims data representing millions of payer interactions, with a focus on more than 100 claims attributes.

The system ran the data through a regression model to find which attributes correlated most closely with a denied billing claim. With this new predictive model, unsubmitted claims were then prerun, increasing the number of claims identified as likely to be denied by 33%, as compared with a retrospective baseline. In addition, the model identified the most likely root causes of claims denial, such as National Drug Code denial or a specific payer's policy.

That health system now has a pilot program that uses the top 10 root causes to flag claims and address them before submission. Over time, the goal is to further develop the model to generate claims-specific root causes of denial and prevent billings from moving forward if they harbor these flaws. This could further improve the denied-claims record of the health system, reduce administrative spending needed for claims processing and reprocessing, and improve the patient's experience by reducing frustrating denials.

Real-World Example: Insurance Payer Streamlines Claims Processing

On the payer side, a large managed-care organization used AI to move from a traditional, labor-intensive model to an AI-based model, with the goal of eliminating upstream errors and reducing the need for manual claim adjudication. The organization trained a model that identified and weighted the factors leading to manual intervention, such as specific procedure codes. The model continuously generated output based on relative manual effort.

The results were substantial:

• The percentage of complex claims processed without denial increased from less than 80% to more than 90%.
• Administrative spending was reduced by 30%.
• Patient experience and clinician satisfaction improved.

AI in Prior Authorization

AI is also being used in prior authorization—the process where insurance companies review and approve treatments before they are provided. This process involves substantial manual labor, with only 21% of prior authorizations automated. The process is costly because it requires doctors and registered nurses to review authorization requests. From the payer's perspective, the objective is to ensure patients receive clinically appropriate treatment, making prior authorization a check on what the provider has ordered.

In an attempt to reduce friction, one payer created an integrated, clean database that included member eligibility and benefits information, historical medical and pharmacy claims, and historical prior authorization requests with clinical decisions, appeals, and outcomes. These data were fed into a triage engine that categorized requests into four levels of complexity based on factors such as the level of detail shared, the plan member's clinical history, and knowledge gained from processing similar requests.

AI has already begun to reduce the number of steps in the prior authorization process compared with traditional manual workflows. The majority of low- and medium-complexity prior authorization requests are now automatically approved. This has led to:

• Reduced turnaround times
• More consistent clinical outcomes
• Better overall experiences for patients and clinicians

The payer's long-term vision is to apply AI to further accelerate decision making.

AI in Clinical Operations: Smarter Hospitals and Operating Rooms

Clinical operations is another health care delivery domain with expanding AI use. Although AI adoption in clinical operations is not as advanced as in reimbursement, the total-mission-value potential is similar, and AI has been an area of intense research in this domain.

Optimizing the Operating Room

Consider the operating room—one of the most critical assets for clinical care in a health system. Demand for operating rooms is traditionally high, so a missed surgical slot can result in a substantial increase in wait time and loss of revenue. Scheduling delays due to surgeries running longer than anticipated can also have nonfinancial effects, such as a worse experience for patients and their families as they wait for an operation to end or a procedure to begin.

In U.S. health systems, more effective use of operating-room capacity can increase access to care—especially important today because of surgical backlogs and clinician shortages.

The authors describe operating-room optimization in three steps, each with increasing AI involvement:

1. Improving Operating-Room Management (Limited AI Use): This step uses descriptive analytics, such as a histogram showing the distribution of operating-room times over the previous 30 days, to identify variations in scheduling. Health systems have used this approach successfully for many years. Examples include measuring key operating-room statistics such as start time, surgical incision time, room turnover time, and patient turnover time (University Hospital, 1994; Medical University of South Carolina, 1998).
2. Predicting Operating-Room Use (Predominant AI Use): AI plays a central role by predicting operating-room use. Preoperative prediction analytics focus on reducing cancellations and estimating mortality risk. During surgery, predictions generally focus on the duration of the procedure and potential complications. Postoperative predictions aim to identify major complications.
3. Using Operating-Room Analytics in Real-Time (Predominant AI Use): This step turns prediction into action. AI would build prediction of procedure duration into precise scheduling, allow coordination of multiple operating rooms used simultaneously, and integrate predictions like likely surgery cancellations into operating-room optimization.

Organizations such as the Mayo Clinic and Lucile Packard Children's Hospital at Stanford have estimated that utilization would potentially be improved by 15 to 20% if AI were implemented. However, this step remains largely in the pilot phase, and whether the improvements will be realized is not yet known.

Notable examples from research cited in the article include:

• West China Hospital (2018): Predicted operations with high risk of cancellation.
• Unidentified academic institution (2020): Predicted risk of death during cardiac surgery.
• Gold Coast Hospital (2017): Predicted duration of elective surgery.
• University of Florida Health (2019): Predicted postoperative major complications and death.
• Mayo Clinic (2015): Simulated improvement in utilization by 19% and reduction of overtime by 10%.
• Lucile Packard Children's Hospital Stanford (2018): Simulated reduction in post-anesthesia care unit holds without decreasing operating-room utilization.

Tackling Clinician Burnout

Another important use of AI in clinical operations is addressing clinician burnout. Physicians now spend more than 50% of their time updating electronic health records (EHRs), and this use of time is a documented contributor to burnout. Multiple providers are piloting natural language processing to reduce this burden.

If these efforts are successful, natural language processing could turn unstructured data such as clinicians' notes into the structured data needed for the EHR, as well as for other uses such as documenting quality metrics or filling in appropriate Current Procedural Terminology (CPT) codes. This application of AI would give clinicians more time to spend with patients and on tasks that require human judgment.

AI for Quality and Safety: Protecting Patients and Improving Experience

Quality and safety constitutes a domain in which a substantial portion of value comes from nonfinancial factors. The current level of AI adoption in this domain is limited, as is the evidence on AI's broad effect on quality and safety. However, two uses of AI—focused on patient safety and patient experience—show real potential.

Reducing Major Adverse Events

The first use of AI focuses on reducing major adverse events—specifically, cases where current evidence-based methods are less useful in preventing and addressing complications, and where integration of complex, unstructured data with measurable metrics could help make predictions. Three problems have been identified as having the greatest potential for improvement with AI:

• Adverse drug events: Harmful reactions to medications that could be predicted and prevented.
• Decompensation: Sudden deterioration of a patient's condition, especially in the intensive care unit (ICU).
• Diagnostic errors: Mistakes in identifying a patient's condition that could be caught earlier with AI assistance.

Addressing these problems requires generating actionable information. This process uses data from sensing technology, including vital-sign monitors, to detect early warning signs of deterioration. The results have been impressive:

• One approach used a predictive algorithm for clinical deterioration in the ICU that reduced mortality in 21 hospitals.
• In another health system, an early-warning system reportedly saved approximately 8,000 lives across the network.
• A recent study showed that when a provider confirms an AI alert, mortality is further reduced—meaning the combination of AI detection and human clinical judgment is especially powerful.

Improving Patient Experience

The second use of AI—improving patient experience—can involve the Consumer Assessment of Healthcare Providers and Systems (CAHPS), a program of the Agency for Healthcare Research and Quality that measures patient satisfaction with health care experiences. AI can analyze patient feedback and identify patterns that help health systems respond to concerns more quickly and effectively, though the article notes this area is still developing.

Clinical Implications: What This Means for Patients

These emerging AI applications have direct implications for patients. In the reimbursement domain, AI that predicts denied claims before submission means fewer frustrating insurance denials. Patients could experience faster approvals, clearer billing, and fewer surprise claim rejections. The fact that up to 85% of denied claims could potentially be avoided—and that one payer increased complex claims processed without denial from under 80% to over 90%—suggests that AI could dramatically reduce the administrative headaches patients often face.

In the prior authorization process, AI-driven automation of low- and medium-complexity requests means patients could get faster approval for necessary treatments. Instead of waiting days or weeks for a manual review, the process could be nearly instantaneous in many cases.

In clinical operations, AI-optimized operating rooms mean shorter wait times for surgeries, fewer canceled procedures, and better coordination of care. For patients awaiting surgery, this is not merely a convenience—it is a matter of timely access to potentially life-saving treatment. A 15 to 20% improvement in operating-room utilization could significantly reduce surgical backlogs.

For clinicians, reducing EHR documentation burden by more than 50% through natural language processing could give doctors more face-to-face time with patients. This directly improves the patient experience and could help address the physician and nurse shortage crisis.

In quality and safety, AI systems that detect clinical deterioration before it becomes critical can save lives. The reported 8,000 lives saved in one health system and mortality reduction across 21 hospitals using ICU deterioration algorithms demonstrate that AI is not just a theoretical tool—it is already making a measurable difference in patient outcomes.

Limitations: What the Study Couldn't Prove

The authors are transparent about the limitations of the evidence. Most uses of AI in health care delivery have not been subject to randomized, controlled trials—the gold standard for medical evidence. Therefore, the usual level of evidence required for medical decision-making may be lacking for many AI applications.

The authors indicate where there is ample evidence and where it is absent. They acknowledge that their perspective is based on conversations with dozens of health care leaders and that this is not a substitute for randomized, controlled trials.

Additional limitations include:

• Pilot-phase uncertainty: Most organizations are still in the pilot phase of AI adoption and are attempting to validate the benefits. Whether improvements like the predicted 15 to 20% operating-room utilization gains will be realized is not known.
• Limited evidence in quality and safety: The evidence on the broad effect of AI on quality and safety is limited.
• Newer technologies unproven: Newer forms of technology, such as blockchain and generative AI, have not played a major role in health care delivery, and while some leaders argue they are essential to unlocking AI's potential, the authors' experience suggests otherwise.
• Information from experience: Much of the examples come from the authors' consulting experience rather than peer-reviewed published studies.

The authors also disclose that they are both employed by a company that provides consulting services for public and private organizations in this area, which is an important potential conflict of interest to consider when evaluating their examples.

Recommendations: What Patients Should Know and Do

For patients navigating a health care system that is increasingly incorporating AI, several practical recommendations emerge from this review:

1. Ask about AI in your care. If your radiologist, pathologist, or other specialist uses AI-assisted image analysis, ask what it means for your diagnosis. AI is being used in up to 30% of radiology practices, so this may be more common than you realize.
2. Be proactive about insurance approvals. Given that AI can only reduce friction in the prior authorization process when implemented, and that more than 10% of claims are denied or delayed, it still pays to verify that your provider has submitted all necessary documentation before procedures or treatments. The 85% of avoidable denials are often due to missing data or eligibility issues.
3. Report concerns about your health promptly. AI early-warning systems in hospitals are designed to catch patient decompensation, but they work best when providers have up-to-date vital signs and symptom information. Your reports of changes in how you feel are part of the "unstructured data" that makes these systems more effective.
4. Understand that AI supplements, not replaces, human judgment. Studies show that when providers confirm AI alerts, mortality is further reduced—the human-machine team works best. Your relationship with your clinical team remains the foundation of your care.
5. Be aware of evidence limits. Because most AI applications have not undergone randomized controlled trials, you should feel comfortable asking your provider about the evidence supporting any AI-assisted recommendation. It's a reasonable question, and the authors of this review would agree it's worth asking.
6. Recognize that AI may improve your patient experience. From fewer denied claims to shorter surgical wait times and less hurried doctors (who currently spend over 50% of their time on electronic health records), AI has the potential to make health care more patient-centered—but these benefits are still being validated in pilot programs.
7. Understand the financial stakes. AI's ability to reduce administrative spending by 30% (as demonstrated by one payer) and improve operating-room utilization by 15 to 20% could translate into cost savings and better access. These changes may take years to reach your local hospital, however, since most organizations are still in the pilot phase.

Frequently Asked Questions

What is artificial intelligence (AI) in health care?

AI uses computers to make intelligent decisions by learning from examples or processing human language. In hospitals, AI can help predict insurance denials, schedule operating rooms, and detect patient deterioration. It supplements doctors' knowledge but does not replace human judgment. Adoption is still early, so many uses are being tested in pilot programs.

Can AI reduce problems with insurance denials and prior authorization?

AI can predict which insurance claims are likely to be denied before submission, catching errors that cause up to 85% of avoidable denials. One payer raised complex claims processed without denial from under 80% to over 90%. AI also speeds up prior authorization for low- and medium-complexity requests, reducing turnaround times and frustration.

How could AI affect how long I wait for surgery?

AI can predict how long surgeries will take and identify procedures at high risk of cancellation. In simulations, hospitals like Mayo Clinic and Stanford estimated AI could improve operating-room use by 15–20%, potentially reducing backlogs and wait times. However, most hospitals are still testing these approaches, so real-world benefits are not yet guaranteed.

Can AI help keep me safe in the hospital?

AI can watch vital signs and alert staff to sudden patient deterioration, which may prevent deaths. One approach reduced mortality in 21 hospitals, and another early-warning system reportedly saved about 8,000 lives in one health network. When a provider confirms an AI alert, mortality is reduced further. This area still has limited evidence.

How strong is the evidence that AI works in health care?

Most AI uses in health care have not been tested in randomized controlled trials, which are the gold standard for medical evidence. Many examples come from pilot programs or expert experience, not peer-reviewed studies. This means the usual level of proof required for medical decisions is often lacking. Patients should ask about evidence behind any AI-assisted recommendation.

What should I ask my doctor about AI in my care?

If your doctor uses AI for image analysis, predictions, or alerts, ask what it means for your diagnosis or treatment. Ask what evidence supports the AI tool and how it was validated. Also verify that your insurance paperwork is complete, since avoidable denials often result from missing data. These questions can help you understand how AI affects your care.

If my doctor's diagnosis or treatment plan is partly based on artificial intelligence, should I get a second opinion?

AI is being used in many areas of health care delivery, including radiology, where up to 30% of practices have adopted it. However, most AI applications have not undergone randomized controlled trials, so the usual level of evidence required for medical decision-making is often lacking. AI is designed to supplement, not replace, human judgment. If you are uncertain about an AI-assisted recommendation, asking another expert to review your care can provide important reassurance. Diagnostic Detectives Network provides independent expert second opinions.

Source Information

Original Article: "Artificial Intelligence in U.S. Health Care Delivery"

Authors: Nikhil R. Sahni, M.B.A., M.P.A.–I.D., and Brandon Carrus, M.Sc.

Affiliations: Department of Economics, Harvard University, Cambridge, MA; Center for U.S. Healthcare Improvement, McKinsey and Company, Boston, MA

Journal: The New England Journal of Medicine (NEJM), 2023; Volume 389, pages 348-358

DOI: 10.1056/NEJMra2204673

Copyright: © 2023 Massachusetts Medical Society. All rights reserved.

This patient-friendly article is based on peer-reviewed research. It is intended for educational purposes and does not constitute medical advice. The original article was edited by Jeffrey M. Drazen, M.D., with guest editors Isaac S. Kohane, M.D., Ph.D., and Tze-Yun Leong, Ph.D., as part of the NEJM "AI in Medicine" review series.