The authors of JAMA Summit Report on Artificial Intelligence divide health AI into 4 buckets:
Clinical tools - sepsis alerts, diabetic retinopathy screening, echo AI. >1,200 FDA-cleared, majority imaging. Adopted by 90% of US health systems.
Direct-to-consumer (DTC) - >350,000 mobile health apps, $70B market. Self-diagnosis dermatology apps, mental health chatbots, smartwatch AFib/fall detection. Often labeled "wellness" to avoid FDA oversight.
Business operations tools - bed management, revenue cycle, scheduling, prior auth. Rapid adoption by payers and providers, but health effects rarely studied.
Hybrid tools - both clinical + business, e.g. AI scribes that listen to visits, generate notes/bills AND suggest diagnoses. "Every patient-clinician conversation in the US may soon be accompanied by a live interactive AI agent."
In more detail:
1. Clinical Tools
"Most AI tools in the medical literature fall under the category of clinical tools"
This is the most studied group, but also the most fragmented in adoption.
What they are: Software directly supporting clinicians — automated screening for diabetic retinopathy, portable echo with automated EF, EHR-scanning sepsis alerts with treatment recommendations, radiology/pathology image interpretation, EHR-embedded clinical decision support.
Scale: >1,200 FDA-cleared AI devices, majority in medical imaging. The paper notes AI has "transformed medical imaging, augmenting image interpretation and dramatically changing how radiologists and pathologists work, with adoption of some type of AI by 90% of US health care systems."
Adoption paradox: Imaging AI and native EHR AI are widespread partly because some EHR tools have been outside FDA oversight and the marginal cost to "turn on" is perceived as low. Yet clinicians remain concerned about accuracy, value, and utility.
No reimbursement for most AI use, even when there is, it doesn't offset full costs. Add skepticism about algorithmic bias, automation bias (favoring AI over contradictory info), lack of generalizability, and lack of clinician/patient trust.
2. Direct-to-Consumer (DTC) Tools
Tools used by patients or individuals with health or wellness concerns
Examples: App using smartphone camera to self-diagnose skin lesion, chatbot for mental health support, algorithm using smartwatch biosensor to detect falls or arrhythmias.
Market: "There are currently more than 350,000 mobile health apps, with AI frequently embedded" — 3 of 10 adults worldwide have used one, $70B+ annual market.
Regulatory loophole: Although products like AFib monitors are FDA regulated, companies can label tools as low-risk general wellness products and avoid reporting requirements. They fall under FTC, which focuses on privacy/security and false advertising, not clinical safety.
Why they matter: Most don't connect to health systems, limiting clinician oversight. When they do, integration costs are high.
3. Health Care Business Operations Tools
The least discussed in literature, but fastest growing revenue side.
Examples: AI to optimize bed capacity, OR scheduling, revenue cycle management, patient/staff scheduling, supply chain, reporting. A popular use case: AI generating prior authorization requests — and payers using AI to evaluate those same requests.
Impact blind spot: Consequences for patients not well understood.
4. Hybrid Tools - the fastest adoption curve
Tools supporting both operations and clinical care/patient experience.
Flagship: AI scribes — listen to patient-clinician conversation, generate notes and bills (reducing documentation burden) while also offering diagnoses/treatment recommendations for verification. Initially mixed results, but now well-received when integrated with EHR.
Patient navigation: Web-based AI that brings more patients into a system while directing them to right clinicians — generating revenue and improving access.
Future: Authors note "Every patient-clinician conversation in the US may soon be accompanied by a live interactive AI agent."
These are being adopted very rapidly because ROI is obvious to systems (less burnout, more billing capture), even though evaluation largely focuses on satisfaction and workflow rather than health outcomes or quality.
References:
https://jamanetwork.com/journals/jama/fullarticle/2840175