The recent releases of generative artificial intelligence (AI) agents like ChatGPT show how AI technologies are entering the mainstream. In healthcare, clinicians are exploring how to use these chatbots for note-taking and medical consultations. For example, one study suggests that ChatGPT could work with electronic medical records to document patient visits, look up possible causes for a constellation of symptoms, draft clinical notes after a meeting, and recommend referral care (Lee et al. 2023). This is one promising type of AI deployment; there are many others.
A basic definition of AI is a machine or computing platform capable of making intelligent decisions. While the generative AI that powers ChatGPT is promising, there are already many traditional AI applications that can change healthcare now. In our experience, two types of AI are particularly important.
One is machine learning, which involves computational techniques that learn from examples rather than operating from predefined rules. An example would be programming a computer to detect cancer from an MRI image.
A second type is natural language processing, which is a computer’s ability to understand human language and transform unstructured text into machine-readable structured data. Translating the interaction between doctor and patient into a note is an example of natural language processing in action.
Together, machine learning and natural language processing can help with countless healthcare tasks, from back-office needs such as billing and referrals to diagnostic tasks such as making a list of possible treatment actions.
To date, AI adoption in healthcare has lagged behind its adoption in other industries (Cam et al. 2019). Innovations typically follow an S-curve: slow diffusion while technology is piloted, followed by rapid scaling and adoption, and finally maturity. Some industries have already reached the mature stage of the S-curve in AI – for example, financial services companies have deployed sophisticated AI algorithms for fraud detection, credit assessments, and customer acquisition. Healthcare has not yet reached this stage (Bughin 2017).
In a recent paper (Sahni et al. 2023), we asked two questions: can widespread use of AI in healthcare save money, and why is AI still used so sparingly in healthcare?
The potential for AI to save money in healthcare
Our bottom line is that widespread AI adoption within the next five years using the technology available today could result in savings of 5% to 10% of healthcare spending, or $200 to $360 billion annually. These savings would not sacrifice quality or access and, indeed, might enhance quality as the experience of healthcare improves. We term the combination of financial and non-financial factors as ‘total mission value’. AI could materially increase total mission value in healthcare.
How might this occur? Consider a few examples. For hospitals, AI-enabled use cases are emerging in nine domains: continuity of care, network and market insights, clinical operations, clinical analytics, quality and safety, value-based care, reimbursement, corporate functions, and consumer engagement. Take the clinical operations domain. Here, hospitals are already piloting use cases such as improving operating-room capacity, freeing up clinical staff time, and optimising the supply chain. Hospitals can develop an optimal operating-room schedule by better considering demand and time for procedures to create more open slots and access (Luo et al. 2020, Kilic et al. 2020). Across the nine domains, we estimate that hospitals could achieve annual net savings of $60 to $120 billion.
Physician groups can similarly benefit from AI. Consider the amount of time wasted because of missed visits. When a patient misses a visit, it not only creates more administrative work (to reschedule) but also underutilises physicians and a time slot that another patient could have used. A missed appointment also could lead to a patient’s condition worsening and a possible trip to the hospital. AI could help predict the likelihood of a missed appointment or a condition worsening so that the healthcare practice can perform proactive outreach. We estimate that physician groups could achieve annual net savings of $20 to $60 billion, roughly half from reducing administrative costs and half from simplifying existing processes.
Private insurance could be transformed as well. A central task for all insurers is claims management, including determining when to pay a provider and what prior authorisation to require. AI can help a payer identify and correct claims errors by learning from previous claims or pulling information from a physician note to approve a prior authorisation, reducing cost and administrative work. Across all domains, we estimate possible annual net savings of $80 to $110 billion for private insurers.
Why is AI not used more widely?
The magnitude of these potential savings raises several questions. A particularly important one is why AI is not more widely used.
There are two broad explanations for the slow adoption. Many economists believe that AI is underused because the healthcare payment system does not provide incentives for this type of innovation (Chernew and Heath 2020). Healthcare organisations today are still more commonly paid based on the volume of care they provide. Thus, use cases that would reduce the need for more care have the perverse effect of lowering revenue.
While the payment model may affect adoption, it is not the entire explanation. For example, everyone would benefit if physicians’ office websites were set up so that patients would not need to call the doctor for routine medication refills. Yet, that is rarely done.
The other view is managerial: large management barriers at both the organisational and industry levels are responsible for the slow adoption of AI in healthcare. Six factors are involved in successful AI adoption: designing a mission-led roadmap; gathering appropriate talent; having an agile delivery model; acquiring appropriate technology and tools; managing data; and changing the organisation’s operating model. Each is a challenge in today’s healthcare system.
Take just one of these: having the right data. While healthcare has terabytes of data, these data are largely unstructured, existing in many different systems and formats. Meanwhile, healthcare is not where skilled young data scientists are rushing to work. Further, this talent is scarce, resulting in rising salary premiums (Taska et al. 2020).
Even if the data were available, managing privacy in healthcare is complex (Gans et al. 2018). Inspiring confidence that the organisation effectively protects data, uses AI responsibly, and provides transparency is essential to winning the ‘digital trust’ of patients and physicians.
There are industry-level barriers, too. Factors that are out of any single organisation’s control can hinder widespread adoption, such as data heterogeneity across institutions, ongoing adaptability, and regulatory challenges. The healthcare system has difficulty standardising how to measure contribution margin per procedure, let alone integrating clinical notes from multiple locations.
These barriers are challenging but not insurmountable. Five years ago, who would have guessed that AI would be helping to diagnose cancer or to power chatbots that answer patients’ questions about COVID-19 or their symptoms? Indeed, we think the future is promising for AI adoption in healthcare, even beyond the use of generative AI applications such as ChatGPT. The COVID-19 pandemic, coupled with rising inflation and labour shortages, is straining the finances of healthcare organisations. Thus, any technology that can ease workforce burden and burnout and improve patients’ access to care will get a second look. Adoption of AI-enabled use cases – especially those that focus on administrative costs – could help the industry address these issues.
We expect that as healthcare organisations gain more experience with AI, they will begin to realise the full benefits of this technology, including improving labour productivity and flattening the curve of healthcare spending. More importantly, the application of AI in healthcare could improve the quality of care for patients and lead to greater satisfaction for doctors and patients alike.
Bughin, R J (2017), “The new spring of artificial intelligence: A few early economies”, VoxEU.org, 21 August.
Cam, A, M Chui, and B Hall (2019), Global AI survey: AI proves its worth, but few scale impact, McKinsey and Company.
Chernew, M E, and J Heath (2020), “How different payment models support (or undermine) a sustainable health care system: Rating the underlying incentives and building a better model”, NEJM Catalyst Innovations in Care Delivery 1(1).
Gans, J, A Goldfarb, and A Agrawal (2018), “Economic policy for artificial intelligence”, VoxEU.org, 8 August.
Kilic, A, A Goyal, J K Miller, E Gjekmarkaj, W L Tam, et al. (2020), “Predictive utility of a machine learning algorithm in estimating mortality risk in cardiac surgery”, Annals of Thoracic Surgery 109(6): 1811–19.
Lee, P, S Bubeck, and J Petro (2023), “Benefits, limits, and risks of GPT-4 as an AI chatbot for medicine”, New England Journal of Medicine 388: 1233–39.
Luo, L, F Zhang, Y Yao, R Gong, M Fu, et al. (2020), “Machine learning for identification of surgeries with high risks of cancellation”, Health Informatics Journal 26(1): 141–55.
Sahni, N R, G Stein, R Zemmel, and D M Cutler (2023), “The potential impact of artificial intelligence on healthcare spending”, in A Agrawal, J Gans, A Goldfarb, and C Tucker (eds.), The Economics of Artificial Intelligence: Health Care Challenges.
Taska, B, S Samila, J Azar, M Gine, and L Alekseeva (2020), “The demand for AI skills in the labour market”, VoxEU.org, 3 May.