AI in Healthcare Statistics: Adoption, Uses, Attitudes and Results

This page collects current, citable statistics on AI in healthcare: how many hospitals and practices use it, what they use it for, market size, what clinicians and patients think, and what the evidence shows about results. Each figure lists its source, year, sample and method. Survey self-reports are kept separate from measured outcomes. Where the evidence is thin, especially on the "success rate" of AI, the page says so.

AI in healthcare statistics at a glance

Last updated and review cadence: we recheck each figure against its source on a regular review schedule. The "Year" column shows when each source collected its data. If you need 2025 or 2023 figures specifically, check that column rather than the page date.

Statistic	Source	Year	Sample or method	Type
Share of U.S. hospitals using AI or predictive models	AHA survey data or ONC data brief	Latest available release	Hospital survey; check the source's AI definition	Survey
Share of physicians using AI in their practice (latest wave)	AMA physician AI survey	Latest wave	Physician survey; check sample size	Survey
Change in physician AI use between survey waves	AMA physician AI surveys	Two most recent waves	Valid only if the question stayed the same	Survey
Share of medical practices using AI tools	MGMA or other practice-level source	Latest available	Practice poll; check sample and vendor sponsorship	Survey
Total FDA-authorized AI-enabled medical devices	FDA AI-enabled device list	Check the list's update date	Public list of authorized devices	Regulatory count
Share of those devices in radiology	FDA AI-enabled device list	Check the list's update date	Count by review panel	Regulatory count
Share of organizations using ambient clinical documentation	Independent surveys (data still sparse)	Latest available	Check respondent type and sponsorship	Survey
Global AI in healthcare market size, base year	Named market research firm	Firm's base year	Proprietary market model	Market forecast
Forecast market size and CAGR	Same firm as the base-year figure	Firm's forecast year	Proprietary market model	Market forecast
Share of digital health funding going to AI companies	Rock Health or equivalent	Latest annual report	Tracked U.S. venture deals	Funding report
Share of U.S. adults uncomfortable with their provider relying on AI	Pew Research Center or equivalent	Latest available	Panel survey of U.S. adults	Survey
Large healthcare data breaches reported to HHS OCR	HHS OCR breach portal	Period you select	Breaches affecting 500+ individuals (all causes, not AI-specific)	Regulatory count

These figures come from different sources that define "AI" differently. Some count any predictive model, and some count only generative AI. Do not compare them directly or add them together.

Adoption: how many hospitals, health systems and practices use AI

Hospitals and health systems. AHA survey data and ONC data briefs built on it are the usual sources for hospital adoption figures. Check how the source defines AI before you quote it. Hospital surveys often ask about "predictive models" integrated with the EHR, which can include rules-based risk scores that most people would not call AI. Record the source's definition alongside the figure.

Physicians. The AMA has surveyed physicians on AI in more than one wave, which makes it one of the few sources that shows a trend:

• Earlier wave: the share of physicians reporting any use of AI, with the survey year and sample size
• Later wave: the same measure from the follow-up survey, so the change can be shown

When you cite the change, confirm that both waves asked the same question. If the wording or definition changed, call the comparison indicative, not exact.

Independent and outpatient practices. Practice-level data is limited. MGMA and other practice-level polls publish occasional figures, but samples are smaller and definitions vary. Note whether the survey was sponsored by a vendor that sells AI tools. Sponsored surveys can still be useful, but readers should know who paid for them.

Caveat. Adoption surveys usually count any use. A practice that piloted an AI scheduling assistant for one month counts the same as a group that runs it across 12 locations. A high adoption rate does not mean broad deployment, daily use or measurable benefit.

Where AI is used: administrative versus clinical work

Administrative and operational uses. Most surveys report AI arriving first in administrative work, such as:

• Documentation support and note drafting
• Revenue cycle, coding and billing
• Prior authorization
• Scheduling and patient access
• Inbox and message triage

In a typical clinic, this is work like sorting front-desk requests, drafting replies to routine staff questions, or tracking prior authorization follow-ups that otherwise sit in a spreadsheet. For examples of these uses in day-to-day clinic operations, see how clinic teams use AI.

Clinical uses. Imaging and diagnostics account for most regulated clinical AI. The FDA publishes a list of AI-enabled medical devices it has authorized. The list is updated periodically, and radiology accounts for the largest share of devices on it. The list counts authorizations. It does not show how widely each device is used or how well it performs in practice.

Ambient and AI-assisted clinical documentation. This category covers tools that listen to a visit and draft a clinical note. It is one of the fastest-discussed categories, but published adoption data is still sparse. Check whether figures come from vendors or from independent surveys.

Why administrative AI moves faster. Software that meets the definition of a medical device goes through FDA review. Most administrative and operational tools, such as scheduling, billing support and internal coordination, do not, so organizations can adopt them on their own procurement and compliance timeline. Faster adoption does not mean lower risk. Administrative tools still handle PHI.

Market size and investment

Market size and forecasts. Several research firms publish AI in healthcare market estimates. When you cite one, record these fields:

Field to record	What to note
Research firm	The firm's name and the report title
Base year and value	The year the estimate starts from and its stated value
Forecast year and value	The end year of the projection and its stated value
CAGR	The growth rate the firm applies between the two years

These numbers are projections, not measured results. They often differ widely for three reasons:

• Definitions. Some include AI-enabled hardware such as imaging equipment, and others count only software.
• Regions. Some cover global spending, and others cover North America or the U.S. only.
• Segments. Pharma and drug discovery, payers and consumer health apps are included in some models and excluded in others.

If you use a market figure in a business case, cite one firm, state its scope and avoid averaging across firms.

Funding. Digital health funding trackers such as Rock Health report how much venture funding goes to AI-focused companies. Funding data tracks investor interest, not clinical adoption or results.

Clinician and patient attitudes toward AI

Physicians and nurses. Physician surveys, including the AMA's, ask whether respondents see more benefits or more concerns from AI. Commonly cited concerns include privacy, liability, accuracy and the effect on the patient relationship.

Burnout and administrative burden. Many clinicians say they are interested in AI because they hope it will reduce paperwork, not because they want it to make clinical decisions.

Patients. Pew Research Center and similar organizations have surveyed U.S. adults on how comfortable they would be with their provider relying on AI. Surveys also ask about concern over AI handling health data.

Transparency. Some surveys ask whether patients and clinicians want to be told when AI is used.

Attitude surveys measure perception at one point in time and are sensitive to question wording. For the reasoning behind these concerns, see benefits, risks and ethics of AI in healthcare.

What is the success rate of AI in healthcare?

There is no single success rate for AI in healthcare. Any page that gives one number is combining results that cannot be compared. Results depend on:

• The task. Detecting a finding on an image, drafting a note and predicting a no-show are different problems with different metrics.
• The dataset. A model tested on data from one academic center may perform differently on another population or at a community clinic.
• The comparison point. AI alone versus a clinician alone, a clinician with AI versus without it, or a new process versus the old one.
• The metric. Sensitivity, specificity, AUC, minutes saved, adoption rate and user satisfaction each measure something different.

Diagnostic accuracy. Peer-reviewed studies and meta-analyses report accuracy for specific tasks. When you cite one, include:

• The task, the metric and the comparator the study used
• The setting and the limitations the authors state

Note whether each study was retrospective (tested on past data) or prospective (tested in live care). Most published accuracy figures are retrospective.

Operational outcomes. For documentation time saved, use published studies rather than vendor marketing, and record the sample, setting and whether time saved was measured (e.g. EHR audit logs) or self-reported (survey). Measured and self-reported time savings can differ substantially, so label which kind you are citing.

Implementation and failure data. Not every AI project delivers. Some pilots never scale beyond a single site, and external validation studies have found deployed models performing worse in real-world use than in development. Results like these are underreported compared with positive findings, which skews the published evidence.

Checklist for judging any AI statistic:

1. Who funded it? A vendor, an association, a government agency or an independent researcher?
2. What was the sample size? How were respondents or patients selected?
3. Was it measured or self-reported? Time logs differ from "it feels faster."
4. Was it a real-world deployment or a lab test? Retrospective datasets are not live clinics.
5. What was the comparator? "Better" than what?
6. How is AI defined? Predictive models, machine learning, generative AI, or all three?

Risks, privacy and regulation in numbers

Data breaches. The HHS Office for Civil Rights posts breaches affecting 500 or more individuals. In recent years, most large breaches on the portal have been attributed to hacking or IT incidents. These figures cover all causes. They are not evidence that AI caused breaches unless a specific source says so. They show why any new tool that touches PHI needs the same scrutiny as the rest of the stack.

Federal regulation.

• ONC's HTI-1 final rule set transparency requirements for decision support interventions in certified health IT.
• FDA has issued guidance on AI-enabled device software, and some of it is in draft form, so check the FDA's current titles and status before citing it.

State legislation. A growing number of states have introduced or passed bills on AI in healthcare, covering topics such as disclosure to patients and limits on AI in utilization review.

HIPAA. HIPAA does not stop applying because a tool uses AI. If staff enter PHI into an AI tool, the vendor generally needs a Business Associate Agreement, and the organization still owns its safeguards and policies. For details on general-purpose tools, see HIPAA and AI tools like ChatGPT. Smaller behavioral health practices often ask whether HIPAA applies to therapists at all, and the answer shapes how they should vet these tools.

What these numbers mean for clinic and practice operations

The figures point in one direction. AI is reaching practices through administrative and coordination work, such as inbox triage, prior authorization, scheduling and documentation support, before most clinical uses. For a practice owner or administrator, the question is less "should we use AI" and more "which AI tools are already handling our data, and under what terms."

Evaluate AI the way you would evaluate any system that touches PHI: data handling, access control and auditability first, features second. Questions to ask every vendor:

• Is there a signed BAA? On which plan?
• Does the BAA cover the AI features specifically, including any third-party models the AI calls?
• Is our data used to train outside models?
• Can admins control what the AI can see and do?
• Can admins see a record of what the AI did, and revoke its access?
• Who in our practice is responsible for configuration, offboarding and staff training?

Compare answers against the vendor's published security documentation. For an example of what that documentation should include, see our security, HIPAA and BAA details.

Chanty is internal team chat and task coordination for clinic and practice staff, used alongside your EHR and other clinical systems, not in place of them. HIPAA support with a signed BAA is available from the Business plan, so practices can run HIPAA-compliant chat for staff. AI agents that work inside team threads are Enterprise-only. They come with permission controls, a real-time audit trail of agent actions, and no training of outside models on your conversations. If you evaluate Chanty, ask the same BAA question about agent processing and any models you connect. More on team communication software for clinics.

How we sourced these statistics

Selection criteria. We use primary sources first: government agencies (FDA, HHS OCR, ONC), professional associations (AMA, AHA, MGMA), and peer-reviewed journals. Reputable research organizations such as Pew come next. Market research firms are cited only for market forecasts and labeled as projections. Vendor-sponsored data is labeled as such.

Each figure includes the source name, publication year, sample and method, plus the URL and access date in the editor's source log.

Updates. We review every figure on a regular schedule. If you find a figure that is outdated or misattributed, contact us with the figure and the newer source, and we will review it.

No Chanty data. No figures on this page come from Chanty customer data or Chanty marketing claims.

For related reading, browse our healthcare communication and HIPAA guides.

FAQ

What percentage of healthcare organizations use AI?

It depends on the type of organization and how AI is defined. Hospital figures usually come from AHA survey data and ONC data briefs, physician figures from AMA surveys, and practice figures from MGMA or other practice-level polls, and each measures something different. Most adoption surveys count any use, including pilots.

What is the success rate of AI in healthcare?

There isn't one. Results vary by task, dataset, metric and comparator. A diagnostic accuracy figure for one imaging task says nothing about documentation time saved or scheduling outcomes. Look for peer-reviewed, real-world studies for the specific use you care about.

How big is the AI in healthcare market?

Estimates vary by firm, and each reports a base-year value, a forecast value and a CAGR under its own definition of the market. These are projections, and firms define the market differently, so do not compare them directly.

How many AI medical devices has the FDA authorized?

The FDA publishes and periodically updates a list of AI-enabled medical devices it has authorized, so check the current list for the total. Radiology accounts for the largest share. The list shows authorizations, not usage or real-world performance.

Do doctors trust AI?

Physician sentiment is mixed: surveys find interest in AI alongside reservations. Common concerns include privacy, liability, accuracy and the effect on the patient relationship.

Are patients comfortable with AI in healthcare?

Surveys, including Pew Research Center's, show that many U.S. adults have reservations about their provider relying on AI. Comfort tends to depend on what the AI is doing and whether patients are told about it.

Is it HIPAA compliant to use AI tools with patient information?

Only under the right conditions. HIPAA applies when PHI is entered into any tool, including AI. The vendor generally needs to sign a BAA that covers the AI features, and your practice remains responsible for access controls, policies and staff training. No tool makes a practice compliant by itself.

What are the most common uses of AI in medical practices?

Surveys point mainly to administrative work: documentation support, revenue cycle and coding, prior authorization, scheduling, and inbox triage. Clinical uses are concentrated in imaging and diagnostics.