AI in Healthcare Examples: 13 Real Uses in Practices

By Sohel Siddique, who builds Lydoh and VoiceboxMD. Updated November 4, 2026.

The short answer

Where AI shows up in a practice today

Most AI in a practice is administrative: answering phones, predicting no-shows, drafting visit notes and patient replies, and suggesting billing codes. The clinical examples are FDA-authorized devices that read images or tests, such as diabetic eye exams in primary care, polyp detection in colonoscopy and dental X-ray reading. Every example below that works has a person checking the output before it reaches a chart, a claim or a patient.

We picked examples with a verifiable source: a peer-reviewed study, an FDA decision, a CMS program page, or our own published data. Two are cautionary: AI that got billing codes wrong more often than right, and a risk score that shortchanged Black patients. One is ours, and marked as such.

Sources checked November 4, 2026. General information, not medical or legal advice.

Our take

The best evidence-to-risk ratio is administrative

For a small practice, the AI worth adopting first is administrative, not diagnostic. Clinical devices have the most rigorous accuracy studies, but an error lands on a diagnosis, and their FDA labels limit them to narrow uses: DermaSensor, for example, flagged most benign lesions as positive in its own study (FDA DEN230008). Administrative AI has thinner evidence, but its mistakes are cheap and a person catches them: a no-show call that wasn't needed, a note the clinician corrects before signing.

Expect relief, not hours back. Of the three time measurements on this page, one scribe saved 9.5% of note time, the other scribe saved nothing measurable, and AI-drafted inbox replies saved no time at all. Burnout and task load scores moved the right way in all three groups, though as secondary measures still to be confirmed (Lukac 2025; Garcia 2024).

What would change our mind:a randomized trial showing a front-office AI tool harmed patients through missed or mishandled requests, or a trial in small community practices showing a diagnostic aid improved patient outcomes with fewer false alarms than DermaSensor's. Either would move the line. Disclosure: we sell administrative AI (Lydoh, and VoiceboxMD for notes), so weigh our view accordingly.

13 examples

Examples of AI in healthcare, by where they sit in a practice

Each one: what it does, the evidence (one source you can open), the safety requirement, and who it fits.

Front desk and phones

01

An AI receptionist on the practice line

Ours: we make Lydoh

What it does.: Answers every call, takes appointment and refill requests for staff to confirm, transfers callers who need a person, and puts urgent after-hours calls through to the doctor.
The evidence.: In our own case study of a one-doctor practice, Lydoh answered 2,151 calls in its first four and a half weeks; 18% came outside office hours, a third were passed to the front desk by design, and staff marked 89% of the 843 requests it wrote up as handled. On day one, our own fraud limit turned away 132 patient calls before we fixed it (the case study). Vendor data, not an independent study.
Safety requirement.: A BAA before any patient call reaches the vendor, no medical advice, emergencies sent to 911, a person one transfer away, and a visible flag on every call the AI failed to answer.
Who it fits.: Practices that miss calls at lunch, at peak hours or after hours. Start with after hours and overflow.

Scheduling and reminders

02

No-show prediction with a targeted reminder call

What it does.: A prediction model scores each upcoming appointment for no-show risk, and staff call only the highest-risk patients a week ahead, instead of calling everyone.
The evidence.: A randomized trial in a Massachusetts General Hospital primary care clinic enrolled 2,247 patients the model rated at 15% or higher risk. With a coordinator call seven days ahead, 22.8% no-showed, against 29.2% with the usual automated call alone (Shah et al., J Gen Intern Med 2016).
Safety requirement.: Check who the model flags. If it leans on insurance type or past attendance, make sure the extra help adds access for those patients rather than double-booking over them.
Who it fits.: Any practice with a meaningful no-show rate and someone who can make a short list of calls each day.

Documentation

03

Ambient AI scribes that draft the visit note

What it does.: Listens to the visit, with the patient's consent, and drafts the note for the clinician to edit and sign.
The evidence.: In a randomized trial of 238 outpatient physicians across 14 specialties, one scribe (Nabla) cut time in notes by 9.5% against usual care; the other (Microsoft DAX) showed no significant change. Both groups reported less burnout and task load, and clinicians said clinically significant errors appeared "occasionally" (Lukac et al., NEJM AI 2025).
Safety requirement.: A BAA, patient consent to the recording, and the clinician reads and signs every note. The trial's own authors say the occasional inaccuracies "require ongoing vigilance". Disclosure: we make VoiceboxMD, an AI clinical notes tool (voiceboxmd.com).
Who it fits.: Clinicians who finish notes at night. Expect relief more than hours back.

Coding, billing and prior authorization

04

AI-suggested billing codes

What it does.: A general chatbot, or a coding feature built on one, proposes ICD-10 or CPT codes from a description.
The evidence.: Given the official description of real codes from a large health system, GPT-4, the best model tested, returned the exact ICD-10-CM code 33.9% of the time and the exact CPT code 49.8% of the time. The authors concluded these models are "not appropriate for use on medical coding tasks without additional research" (Soroush et al., NEJM AI 2024).
Safety requirement.: A certified coder or the billing clinician owns every code. Treat a suggested code as a search hint, never as the claim.
Who it fits.: Coders who want a faster lookup, with the code book still open. Not a reason to drop coding review.

05

AI reviewing your prior authorization requests (the payer side)

What it does.: Payers are putting AI on their side of prior authorization. Medicare's WISeR model uses technology companies with AI and machine learning to review requests for selected services.
The evidence.: WISeR runs from January 1, 2026 to December 31, 2031 in traditional Medicare in New Jersey, Ohio, Oklahoma, Texas, Arizona and Washington, for services such as skin substitutes, nerve stimulator implants and knee arthroscopy for osteoarthritis. CMS says "all recommendations for non-payment are determined by appropriately licensed clinicians" (CMS WISeR model).
Safety requirement.: On your side: document medical necessity fully, track every denial and its stated reason, and appeal. A clinician at the payer must make any non-payment decision.
Who it fits.: Practices in the six WISeR states that bill the listed services to traditional Medicare.

Patient messages

06

AI-drafted replies to patient portal messages

What it does.: The EHR drafts a reply to each patient message; a clinician or nurse edits it or starts over.
The evidence.: In a five-week Stanford pilot with 162 primary care and gastroenterology clinicians, drafts were used for 20% of replies. Reading, writing and reply times did not change, but task load and work exhaustion scores fell (Garcia et al., JAMA Netw Open 2024).
Safety requirement.: A BAA (it reads patient messages), and a clinician approves anything clinical before it goes out. Never auto-send.
Who it fits.: Practices with a heavy portal inbox, if the EHR offers it under its existing agreement.

Imaging and test reading

07

Autonomous diabetic eye exams in primary care

What it does.: Staff take retinal photos during a diabetes visit and the software decides, without an eye doctor reading the images, whether the patient needs referral to eye care.
The evidence.: FDA granted IDx-DR in 2018. In its pivotal study of 900 patients at 10 primary care sites, it found more-than-mild diabetic retinopathy with 87% sensitivity and 90% specificity, run by operators who had never done eye imaging before a four-hour training (FDA DEN180001).
Safety requirement.: Use only for the labeled patients (adults with diabetes, no prior retinopathy diagnosis) and camera. FDA's label: refer positive results to eye care immediately, retest or refer when there is no result, and remember it detects diabetic retinopathy only.
Who it fits.: Primary care and endocrinology practices whose diabetic patients skip the yearly eye exam.

08

Polyp detection during colonoscopy

What it does.: Marks possible polyps on the screen in real time so the endoscopist looks again.
The evidence.: In the randomized study behind FDA's 2021 decision, the adenoma detection rate was 55.1% with GI Genius and 42.0% without, across 263 patients and six endoscopists (FDA DEN200055).
Safety requirement.: FDA's label: "not intended to replace clinical decision making" and not a stand-alone diagnostic. The endoscopist decides what to remove.
Who it fits.: GI practices and endoscopy centers that track adenoma detection rates.

09

A skin-lesion second read for primary care

What it does.: A handheld device scans a lesion the physician already finds suspicious and suggests whether to refer to a dermatologist.
The evidence.: FDA granted DermaSensor in January 2024 for physicians who are not dermatologists. In its pivotal study, sensitivity for skin cancer was 95.5% against 83.0% for the primary care physicians, but specificity was 20.7% against their 54.2%, so most benign lesions still came back positive (FDA DEN230008).
Safety requirement.: Labeled for patients 40 and older, on lesions already judged suspicious, never as a screening tool, never as the sole basis for a diagnosis. Expect more referrals.
Who it fits.: Primary care practices far from a dermatologist that want help deciding whom to refer.

10

AI second reader on dental X-rays

What it does.: Marks suspected caries, calculus, margin problems and periapical lesions on bitewing and periapical X-rays.
The evidence.: FDA cleared Pearl's Second Opinion in 2022. In its reader study, 25 dental readers each read 2,010 images with and without the software, and aided detection improved significantly for all four findings (FDA K210365).
Safety requirement.: A second reader only; FDA's summary says the clinician "needs to make final determinations". It doesn't grade depth or severity, so it can't justify a filling on its own.
Who it fits.: General dental practices, and as a chairside picture for patients. More in AI for dentists.

11

AI reading routine ECGs for a weak heart

What it does.: Software reads an ECG the practice already ordered and flags patients likely to have a low ejection fraction, prompting an echocardiogram.
The evidence.: In a pragmatic trial across 120 primary care teams and 22,641 adults, clinicians who saw the AI result made a new diagnosis of low ejection fraction in 2.1% of patients, against 1.6% with usual care (Yao et al., Nature Medicine 2021).
Safety requirement.: A flag, not a diagnosis: the echo decides. Check the product's FDA status before using it in your practice.
Who it fits.: Primary care practices that do ECGs routinely and can order echocardiograms.

Population outreach

12

Prompts for serious illness conversations

What it does.: A model flags patients at high risk of dying within six months, and clinicians get a weekly list and a text before the visit suggesting a goals-of-care conversation.
The evidence.: In a randomized trial at nine oncology clinics with 78 clinicians, these conversations happened in 15.2% of high-risk patient visits with the prompts, against 3.6% without (Manz et al., JAMA Oncology 2020).
Safety requirement.: The list goes to clinicians, never to patients. A clinician decides whether and how to start the conversation.
Who it fits.: Oncology, geriatrics and primary care practices caring for seriously ill patients.

13

Risk scores that choose who gets extra care

What it does.: An algorithm ranks patients so a care manager reaches out to the highest-need ones first.
The evidence.: A widely used commercial algorithm predicted health costs rather than illness. Because less is spent on Black patients' care, it rated them healthier than equally sick White patients; fixing it would raise the share of Black patients getting extra help from 17.7% to 46.5% (Obermeyer et al., Science 2019).
Safety requirement.: Ask what the score predicts (cost, visits or illness), and compare who is flagged across groups before you use it to ration outreach.
Who it fits.: Practices in value-based contracts that receive risk lists from a payer or ACO.

At a glance

The 13 examples compared

Risk level is our judgment of what an uncaught error costs a patient or the practice. The human check is what each example needs before its output is used.

Example	Area	Evidence	Risk	Human check required
AI receptionist (Lydoh, ours)	Front desk and phones	Vendor case study (ours)	Medium	Staff confirm every request; urgent calls go to a clinician
No-show prediction	Scheduling and reminders	Randomized trial, primary care	Low	A person makes the call; review who gets flagged
Ambient AI scribe	Documentation	Randomized trial, 238 physicians	Medium	Clinician edits and signs every note
AI code suggestions	Coding, billing and prior authorization	Benchmark study: under 50% exact	Medium	A coder confirms every code
Payer AI on prior auth (WISeR)	Coding, billing and prior authorization	CMS model (no outcome data yet)	Medium	Payer clinician on denials; your staff track and appeal
AI-drafted portal replies	Patient messages	Pilot study: less burden, no time saved	Medium	Clinician approves every reply
Diabetic retinopathy AI (IDx-DR)	Imaging and test reading	FDA De Novo, 900-patient study	High	Positive results referred to eye care; follow-up tracked
Colonoscopy polyp detection (GI Genius)	Imaging and test reading	FDA De Novo, randomized study	Medium	Endoscopist decides on every mark
Skin lesion device (DermaSensor)	Imaging and test reading	FDA De Novo, low specificity	High	Physician decides referral with the full exam
Dental X-ray AI (Pearl)	Imaging and test reading	FDA 510(k), 25-reader study	Medium	Dentist confirms every mark before treatment
AI-ECG for low ejection fraction	Imaging and test reading	Randomized trial, primary care	Medium	Clinician orders and reads the echo
Mortality-risk nudges	Population outreach	Randomized trial, 9 clinics	Medium	Clinician decides every conversation
Care-management risk scores	Population outreach	Peer-reviewed audit: biased	High	Review flagged lists by group; clinician overrides

Before you adopt any of these

Four questions for every AI tool

Does it touch patient information? Then the vendor signs a Business Associate Agreement first. Our guide on whether ChatGPT is HIPAA compliant shows which general AI plans sign one.
What is the evidence, and for whom? A study at an academic center or a device's FDA label tells you the setting and patients it was tested on. Outside those, you are the experiment.
Who checks the output, and when? Name the person: the clinician who signs, the coder who confirms, the staff member who calls back.
What happens when it fails? Ask the vendor how you find out. Our own case study shows a failure caught by a guardrail we had to fix.

Put the answers in writing with our free AI policy template. Specialty guides: AI for doctors, AI for dentists and AI in veterinary practice.

FAQ

Questions about AI in healthcare

What are some examples of AI in healthcare?

In a doctor's office today: an AI receptionist that answers the phones, models that predict no-shows so staff call the right patients, ambient scribes that draft visit notes, AI-drafted replies to portal messages, code suggestions for billing, autonomous diabetic eye exams in primary care, polyp detection during colonoscopy, an AI second reader on dental X-rays, and AI that reads routine ECGs for a weak heart. Payers also now use AI to review prior authorization requests.

What are examples of AI in a medical office?

The common ones are administrative: answering calls and taking appointment requests, reminder and no-show outreach, drafting visit notes, drafting replies to patient messages, and suggesting billing codes. Each works best with a person checking the output: staff confirm requests, the clinician signs the note, a coder confirms the code.

Which AI in healthcare has the best evidence?

The clinical devices FDA has authorized have the most rigorous studies behind their accuracy: IDx-DR for diabetic eye disease, GI Genius for colonoscopy and Pearl Second Opinion for dental X-rays, among others. But their mistakes land on a diagnosis, and their labels limit how they can be used. For administrative AI, the best evidence is randomized trials of no-show prediction and ambient scribes, which measured attendance, note time and burnout rather than patient outcomes.

Does AI save doctors time?

Less than the marketing says. In a randomized trial of ambient scribes, one product cut time in notes by 9.5% and the other made no significant difference. In a Stanford pilot of AI-drafted portal replies, time did not change at all. Both studies did find lower burnout and task load, so expect relief more than hours back.

Is AI in healthcare safe?

It depends on the use and the check around it. Every example on this page that works has a person reviewing the output before it reaches a chart, a claim or a patient. The risks are real: general chatbots got under half of billing codes exactly right in one study, and a widely used care-management algorithm was biased against Black patients. Use tools with a Business Associate Agreement for anything involving patient information, and keep a human in the loop.

Do I need a BAA to use AI in my practice?

If the tool will receive patient information, yes. HIPAA requires a Business Associate Agreement with any vendor that creates, receives, maintains or transmits protected health information for your practice. That covers phone answering, scribes, inbox drafting and coding tools. Consumer chatbots generally don't sign one; our guide on whether ChatGPT is HIPAA compliant lists which plans do.

Is AI being used for prior authorization?

On the payer side, yes. Medicare's WISeR model, running from 2026 to 2031 in six states including New Jersey and Texas, uses AI and machine learning to review prior authorization requests for selected services, with licensed clinicians making any non-payment decision. Separately, a CMS rule requires Medicare Advantage, Medicaid and CHIP plans to decide prior authorizations within 72 hours for urgent and seven calendar days for standard requests, and to give a specific reason for denials, starting in 2026.

What AI should a small practice start with?

Start where a mistake is cheap and a person catches it: after-hours and overflow phone answering, no-show outreach, and note drafting with the clinician signing. Leave diagnostic AI to FDA-authorized devices used exactly as labeled. Write a short AI use policy first so everyone knows which tools are approved.

Prior authorization timeframes: CMS Interoperability and Prior Authorization Final Rule (CMS-0057-F).

Keep reading

More on bringing AI into your practice

Guides, a case study and checklists for using AI in a medical, dental or veterinary practice, safely.

AI in your practice Read the case study

AI in healthcare examples: 13 real uses in medical practices