AI in Healthcare: Real Examples by Specialty and Workflow

The FDA's public list of AI-enabled medical devices includes more than a thousand devices authorized for marketing in the United States, and radiology accounts for the largest share of them (FDA AI-Enabled Medical Devices list). For a healthtech team, the harder question is which of these examples actually work inside clinical workflows and what it took to get them there. This article walks through real, sourced examples by specialty and by workflow, then covers how to judge whether an example is production-ready.

Two kinds of healthcare AI

Healthcare AI falls into two broad groups, and the difference shapes how a product is built, sold and maintained.

• Regulated device software: AI that diagnoses, detects or treats a condition is usually a medical device. In the US it needs FDA marketing authorization through 510(k), De Novo or PMA, and it appears on the FDA's AI-enabled device list once authorized.

• Non-device software: tools for documentation, scheduling, billing and many clinical decision support functions sit outside device regulation if they meet the criteria in FDA's Clinical Decision Support software guidance, which FDA revised in January 2026 (FDA CDS guidance).

Both groups still need HIPAA safeguards when they handle protected health information, and both need validation before clinicians rely on them.

Examples by specialty

The examples below were among the first AI products the FDA authorized in their categories. Each one shows a different way AI fits into care.

Specialty	Example	What it does	Regulatory route
Ophthalmology	IDx-DR (2018)	Analyzes retinal images taken in primary care and returns either "refer to an eye care professional" or "rescreen in 12 months" for more than mild diabetic retinopathy	De Novo; described by FDA as the first device to use AI for this detection (FDA announcement)
Neurology (stroke)	Viz.ai ContaCT (2018)	Analyzes brain CT for signs of a suspected large vessel blockage and sends a text alert to a neurovascular specialist; notification only, the radiologist still reads the scan	De Novo (FDA announcement)
Pathology	Paige Prostate (2021)	Highlights the area of a prostate biopsy image most likely to contain cancer for the pathologist to review	De Novo; in FDA's study, pathologists' detection on individual slide images rose from 89.5% to 96.8% (FDA announcement)
Radiology (general)	Detection, triage and measurement tools	Flags suspected findings, reorders reading worklists by urgency, or measures structures automatically	Mostly 510(k); radiology is the largest category on the FDA list

Two patterns repeat across these examples. The AI narrows a clinician's attention or speeds up a handoff, and the clinician keeps the final decision. The product also plugs into an existing system, such as the PACS for imaging or a messaging channel for stroke alerts, instead of asking clinicians to open a new tool. We cover imaging in more depth in our guide to AI in medical imaging.

Examples by workflow

Many of the most widely used healthcare AI tools are not medical devices. They work on documentation, communication and operations. Ambient documentation is the fastest-moving of these; our article on the ambient AI medical scribe covers how those tools are built, and our guide to HIPAA-compliant AI covers the safeguards they need.

Workflow	What AI does	What has to be in place
Clinical documentation	Ambient tools capture the visit conversation and draft a note for the clinician to review and sign	Patient consent, a BAA with every vendor in the chain, and an approved way to write the note back to the EHR
Patient communication	Drafts replies to portal messages and patient education text	Clinician review before sending, and clear limits on clinical advice
Clinical decision support	Surfaces risk scores, reminders or guideline suggestions inside the EHR	Validation on local data and transparency about how the model works (see below)
Revenue cycle	Suggests codes, checks documentation and prepares prior authorization requests	Integration with payer APIs; CMS requires impacted payers to offer a Prior Authorization API generally from January 1, 2027 (CMS-0057-F fact sheet)
Operations	Forecasts bed demand, balances staffing and optimizes scheduling	Reliable, near real-time data feeds from the EHR and staffing systems

What separates a working example from a demo

A model that performs well in a paper or a sales demo can still fail on a hospital floor. The examples that last tend to share four things.

1. Integration: the AI reads from and writes to the systems clinicians already use, through FHIR APIs, HL7 v2 interfaces or vendor programs, with the hospital's security review completed.

2. Local validation: performance is checked on the organization's own patients and broken down by subgroup before go-live.

3. Transparency: under ONC's HTI-1 rule, certified health IT that supports predictive decision support interventions must make source attributes available to users, such as intended use, training data and performance measures (ONC HTI-1 final rule).

4. Monitoring: someone owns the model after launch, watches for drift and has a process to pause or update it.

How to evaluate an AI example before you adopt it

Question	Why it matters
Is it a medical device, and is it authorized for this exact use?	Using a device outside its authorized indication changes your risk and liability
What evidence supports it, and on which patients?	Results from one population do not always transfer to another
How does it connect to the EHR or PACS?	Integration effort often decides the real cost and timeline
Will the vendor sign a BAA, and where is data stored?	Required whenever the tool handles protected health information
Who monitors performance after go-live?	Models drift as practice patterns and data change

For teams building rather than buying, the same questions become requirements. Our healthcare engineering work starts from them so the product fits clinical workflows and compliance reviews from day one.

Conclusion

The strongest AI examples in healthcare do one narrow job well, sit inside an existing workflow and leave the final decision with a clinician. Regulatory status, local validation, integration and monitoring decide whether an example works in practice. Use those four checks when you read about the next AI breakthrough or plan your own product.

If you are planning an AI feature for a healthcare product, book a consultation with Ekipa's HealthTech engineers.

Frequently asked questions

What is a real example of AI in healthcare?

IDx-DR is one of the clearest examples. FDA authorized it in 2018 to analyze retinal images in primary care and tell the clinician whether to refer the patient to an eye specialist for diabetic retinopathy (FDA).

Which medical specialty uses AI the most?

Radiology. It accounts for the largest share of devices on the FDA's AI-enabled medical device list, mostly tools for detection, triage and measurement.

Is every healthcare AI tool regulated by the FDA?

No. Tools for documentation, scheduling and billing, and clinical decision support that meets FDA's non-device criteria, fall outside device regulation. They still need HIPAA safeguards when they handle patient data.

How do I know if an AI tool is FDA-authorized?

Search the FDA's AI-enabled medical device list or the FDA 510(k) and De Novo databases for the product name, and check that the authorized indication matches how you plan to use it.

Does AI replace doctors in these examples?

No. In each example above, the AI narrows attention or speeds up a handoff, and a clinician reviews the result and makes the decision.