Hospitals now buy AI the way they once bought imaging equipment: carefully, with committees, and with a long list of questions. The FDA reports more than 1,600 AI enabled medical devices authorized for U.S. marketing as of September 2026, though the agency notes its public list is not comprehensive. Most are imaging tools. A growing share of new spending, however, goes to software that never touches a scanner, such as ambient scribes, revenue cycle automation, and research assistants.
AI tools in healthcare fall into three different groups, and mixing them up causes confusion. Some are regulated medical devices that make or support clinical claims. Others are enterprise software that automates documentation or operations. A third group is general purpose AI, such as large language model chat assistants, that clinicians and patients use informally. Regulatory status must be verified for any clinical claim.
How Healthcare AI Tools Differ
The workflow, not the algorithm, determines risk and value. A tool that drafts a clinic note carries different stakes than one that flags a possible stroke on a CT scan.
Clinical decision support tools surface diagnoses, drug interactions, or guideline recommendations. Documentation tools convert conversations to notes. Imaging AI detects findings on scans. Patient communication tools answer questions and route messages. Drug discovery and research platforms model proteins and mine literature. Operations tools predict demand, schedule staff, and automate billing. Remote monitoring analytics analyze streaming data from wearables and home devices.
Regulatory status follows intended use. Software that diagnoses, treats, or prevents disease can be a regulated medical device. Administrative tools generally are not. Marketing language does not equal clinical validation.
20 Tools by Healthcare Function
| Function | Tools covered |
|---|---|
| Clinical documentation | Microsoft Dragon Copilot, Abridge, Ambience Healthcare, Suki |
| Imaging and diagnostics | Aidoc, Viz.ai, Paige, Lunit |
| Clinical decision support and knowledge | OpenEvidence, UpToDate Expert AI, Epic AI features |
| Predictive and operational | Epic Sepsis and deterioration models, Qventus, Olive style automation tools |
| Patient engagement | Hippocratic AI, Ada Health, Sword Health |
| Drug discovery and research | AlphaFold, Insilico Medicine, Recursion |
| Remote monitoring | Current Health (Best Buy Health), BioIntelliSense |
AI Tools for Clinical Documentation and Administrative Work
Documentation is the clearest early success for generative AI. Physicians spend hours each day in electronic health records, and burnout is well documented. Ambient scribes listen to visits, with patient consent, and draft notes for clinician review.
1. Microsoft Dragon Copilot
Microsoft combined Nuance’s Dragon Medical dictation with its DAX ambient technology into Dragon Copilot. It drafts notes, orders, and summaries inside major electronic health records. Large health systems use it at scale. Evidence includes health system reports of reduced documentation time and clinician satisfaction gains, though rigorous randomized data are still emerging. It is an administrative tool, not a diagnostic device. Limitations include occasional note errors and the need for clinician review.
2. Abridge
Abridge, founded in Pittsburgh, provides ambient documentation for large systems such as Kaiser Permanente and Mayo Clinic. Its notes link back to the source conversation to help clinicians verify content. It integrates deeply with Epic. Pricing is enterprise quote based. Limits include accuracy in noisy rooms and variable adoption across specialties.
3. Ambience Healthcare
Ambience offers AI scribe and coding support that aims to improve documentation completeness for billing. It targets health systems and large groups. A coding focus appeals to finance leaders, yet the tool must be monitored for overcoding risks.
4. Suki
Suki provides a voice enabled assistant for note creation, orders, and coding suggestions. It serves clinics of varying sizes and integrates with several record systems. Smaller practices often find it accessible. Performance varies with accents, specialty vocabulary, and workflow.
AI for Diagnosis, Imaging and Clinical Decision Support
5. Aidoc
Aidoc analyzes CT scans and other images to flag urgent findings such as brain hemorrhage or pulmonary embolism and prioritize radiologist worklists. It has multiple FDA clearances. Evidence includes studies showing faster turnaround for flagged cases. It supports, rather than replaces, radiologists, and false alerts can cause alert fatigue.
6. Viz.ai
Viz.ai pioneered AI stroke detection, alerting care teams to large vessel occlusion and coordinating transfer. It has multiple FDA clearances and has expanded into cardiology and other areas. Studies suggest reduced time to treatment in some hospital networks. Success depends on workflow integration among referring hospitals and comprehensive centers.
7. Paige
Paige develops AI for digital pathology, including FDA authorized prostate cancer detection support, and was acquired by Tempus AI in August 2025. Pathology digitization is a prerequisite, which is still incomplete in many labs. Evidence supports improved sensitivity as a second reader. Costs and infrastructure are barriers.
8. Lunit
Lunit, based in South Korea, offers AI for chest X-rays and mammography. Its breast cancer tools have been evaluated in screening studies, including large European trials of AI supported mammography reading that suggested similar cancer detection with reduced workload. Regulatory clearances vary by region. Findings must be confirmed by radiologists.
9. OpenEvidence
OpenEvidence is an AI search and answer tool for clinicians that cites medical literature. It has seen rapid adoption among U.S. physicians. It operates as a reference tool, not a regulated diagnostic device. Responses should be checked against primary sources, since language models can still err.
10. UpToDate Expert AI
Wolters Kluwer added generative AI to UpToDate, a trusted evidence summary resource. Answers draw from curated content rather than the open web. That curation reduces hallucination risk. It is subscription based, and institutions often license it.
11. Epic AI Features
Epic Systems, the largest U.S. electronic health record vendor, embeds AI for message drafting, summarization, and predictive models. Its reach is large because many hospitals already use Epic. Studies of its sepsis model raised concerns about performance outside development settings, showing why local validation matters.
AI for Drug Discovery and Biomedical Research
12. AlphaFold
Google DeepMind’s AlphaFold predicts protein structures with high accuracy and earned a share of the 2024 Nobel Prize in Chemistry for its creators. The AlphaFold database provides predicted structures for hundreds of millions of proteins. It accelerates basic research and target identification. It does not directly design approved drugs and has limits for protein dynamics and complexes.
13. Insilico Medicine
Insilico uses generative AI to identify targets and design molecules. Its drug candidate for idiopathic pulmonary fibrosis has advanced through clinical trials, a milestone for AI discovered drugs. Clinical success remains unproven at scale. Company disclosures should be separated from peer reviewed evidence.
14. Recursion
Recursion combines automated lab experiments with machine learning to map biology and screen compounds. It merged with Exscientia in 2024 to broaden its pipeline. Results have been mixed, and several programs have been narrowed. It illustrates how long drug development still takes even with AI.
AI for Patient Engagement and Digital Health
15. Hippocratic AI
Hippocratic AI builds safety focused generative agents for non diagnostic patient communication, such as follow up calls and pre visit outreach. It tests agents with clinicians and partners with health systems. It is not a diagnostic tool. Limits include language nuance and escalation reliability.
16. Ada Health
Ada offers an AI symptom assessment app and clinical platform. It is used by health systems and consumers in many languages. Studies of symptom checkers show wide variation in accuracy, so Ada is positioned as an assessment aid, not a diagnosis. Regulatory status differs by market.
17. Sword Health
Sword uses AI guided digital physical therapy with sensors, targeting musculoskeletal pain. It sells mainly to employers and health plans. Published outcome data come largely from company studies and observational analyses, so independent trials would strengthen evidence.
AI for Operations and Remote Monitoring
18. Qventus
Qventus uses AI to optimize surgical scheduling, patient flow, and discharge planning. Hospitals use it to reduce delays and manage capacity. Value is typically measured in operational metrics. Limits include dependence on data quality.
19. Current Health
Current Health, now part of Best Buy Health, provides remote monitoring kits with wearables and AI based alerts for hospital at home programs. It has FDA clearances for some components. Evidence includes program level reductions in readmissions in some settings. Patient adherence and clinician workload remain challenges.
20. BioIntelliSense
BioIntelliSense offers wearable patch sensors with FDA cleared monitoring of vital signs and algorithms for early deterioration. It serves hospitals and home care programs. Its value depends on integrating alerts with staffing.
Regulatory and Evidence Status
| Tool | Regulatory status (verify) | Evidence type | Key limitation |
|---|---|---|---|
| Dragon Copilot | Administrative software | Health system reports, early studies | Note accuracy |
| Abridge | Administrative software | Health system pilots | Noise, specialty variation |
| Ambience | Administrative and coding support | Vendor and customer reports | Coding oversight |
| Suki | Administrative software | Customer studies | Variable accuracy |
| Aidoc | FDA cleared modules | Retrospective and prospective studies | Alert fatigue |
| Viz.ai | FDA cleared modules | Time to treatment studies | Workflow dependence |
| Paige | FDA authorized prostate tool | Clinical validation studies | Digital pathology prerequisite |
| Lunit | Cleared in multiple regions | Screening trials | Regional differences |
| OpenEvidence | Reference tool | Usage data, early evaluations | Verify answers |
| UpToDate Expert AI | Reference tool | Curated content | Subscription cost |
| Epic AI features | Mixed, some models are regulated or not | Internal and external studies | Local validation needed |
| AlphaFold | Research tool | Peer reviewed (CASP, Nobel) | Not a drug designer |
| Insilico | Drug candidates under trial | Clinical trials phase 2 | Unproven at scale |
| Recursion | Pipeline in trials | Early phase data | Mixed results |
| Hippocratic AI | Non diagnostic | Safety testing, pilots | Escalation reliability |
| Ada Health | Varies by market | Validation studies | Symptom checker accuracy |
| Sword Health | Digital therapeutic category | Company and observational | Independent trials limited |
| Qventus | Operations software | Hospital operational data | Data dependent |
| Current Health | FDA cleared components | Program outcomes | Adherence |
| BioIntelliSense | FDA cleared devices | Validation studies | Alert workload |
Hospital Implementation Considerations
Buying an AI tool is the easy part. Making it work requires governance, validation, and training. Hospitals increasingly build AI committees that review safety, bias, privacy, and workflow fit before deployment.
Key questions include whether the model was trained on populations similar to local patients, how performance is monitored after deployment, and who is accountable for errors. The WHO has issued guidance stressing safety, transparency, and equity in health AI. The FDA’s approach to AI enabled devices follows a risk based pathway, including predetermined change control plans for software that updates over time.
Hospital Implementation Considerations
| Consideration | Questions to ask | Why it matters |
|---|---|---|
| Clinical validation | Peer reviewed or independent evidence? | Marketing is not validation |
| Local performance | Tested on local data? | Models can degrade in new settings |
| Integration | Works with the EHR? | Adoption depends on workflow |
| Privacy and security | HIPAA agreement, data use? | Patient trust and legal risk |
| Governance | Who monitors performance? | Drift and bias need oversight |
| Liability | Who is responsible for errors? | Clinicians remain accountable |
| Cost model | Subscription, per use, or enterprise? | Return on investment varies |
| Training | Staff education included? | Poor use erodes benefits |
Is ChatGPT a Medical AI Tool?
General purpose chatbots are not regulated medical devices. Clinicians and patients use them for explanations, drafting, and summarization, and studies show they can perform well on exam style questions. They can also produce confident errors. Health systems that allow their use typically require approved versions with privacy protections. Anyone using a chatbot for medical decisions should verify with a clinician.
From Experimentation to Evidence Based Implementation
The early hype cycle is giving way to harder questions. Do ambient scribes reduce burnout for the long term? Do imaging tools improve outcomes, not just speed? Do drug discovery platforms yield approved medicines? Early data are encouraging for documentation and some imaging applications, while others lack outcome evidence.
Expect consolidation. Large electronic health record vendors are embedding AI directly, which squeezes standalone point solutions. Regulators are refining oversight, and payers are starting to ask for proof of value. The tools that endure will be those that show measurable gains in safety, access, or efficiency and that earn clinician trust.
Questions to Ask Any Healthcare AI Vendor
Procurement teams that ask the same disciplined questions of every vendor tend to avoid expensive disappointments. Seven questions cover most of the risk.
What exactly is the intended use? A tool cleared to flag suspected intracranial hemorrhage on non contrast head CT is not cleared to read every brain scan. Intended use defines both the regulatory status and the scope of evidence.
Where is the evidence, and who produced it? Peer reviewed studies by independent groups carry more weight than white papers. Prospective studies carry more weight than retrospective ones. Evidence that a tool improved a clinical outcome, such as time to treatment or survival, outranks evidence that it matched a clinician on a test set.
How was the model validated on populations like ours? A model trained at large academic centers may behave differently in a rural hospital or in a population with different demographics. Vendors should be able to report performance by age, sex, race, and scanner or documentation type.
What happens when the model is wrong? Every tool produces false positives and false negatives. The team should know the expected rates, how alerts are routed, and who is accountable when an error reaches a patient.
How will performance be monitored after go live? Models drift as practice patterns, equipment, and patient mix change. A monitoring plan with named owners and thresholds for action is a sign of maturity.
What does it cost in total? Pricing models include per clinician subscriptions, per study fees, and enterprise licenses. Integration work, training, and clinician time spent reviewing outputs are real costs that rarely appear in a quote.
What happens to our data? Contracts should state whether patient data can be used to train the vendor’s models, how long it is retained, and who can access it. A business associate agreement under HIPAA is the minimum.
Return on investment also deserves a plain definition. For documentation tools, the measurable outcomes are minutes of after hours charting saved, visit volume, and clinician burnout scores. For imaging triage, they are turnaround time for critical findings and downstream length of stay. For operations tools, they are throughput, cancellations, and staffing efficiency. A pilot with predefined metrics, a control group when possible, and a firm decision date turns a vendor claim into a local fact.
FAQ
Q: What are the most useful AI tools in healthcare?
A: Ambient documentation tools, imaging triage software, and clinical reference assistants have shown the clearest practical value so far. Usefulness depends on the setting and workflow.
Q: Is ChatGPT a medical AI tool?
A: ChatGPT is a general purpose tool and not a regulated medical device. It can help explain information or draft text, but it can make errors. Clinical decisions need professional judgment.
Q: Which AI tools are FDA authorized?
A: Many imaging tools, including products from Aidoc, Viz.ai, Paige, and Lunit, hold FDA clearances or authorizations for specific uses. The FDA maintains a list of AI enabled devices, though it is not comprehensive.
Q: How is AI used in hospitals?
A: Hospitals use AI for documentation, imaging prioritization, sepsis and deterioration alerts, scheduling, billing, and remote monitoring. Many deployments are administrative.
Q: What are the risks of healthcare AI?
A: Risks include inaccurate outputs, bias, privacy concerns, workflow disruption, and overreliance. Local validation and human oversight reduce these risks.
Q: Do AI scribes reduce physician burnout?
A: Early studies and health system reports show less documentation time and improved satisfaction. Long term, controlled evidence is still developing.
Q: How much do healthcare AI tools cost?
A: Many enterprise tools use quote based pricing per clinician, per scan, or per enterprise. Public prices are limited.
Q: Can AI replace radiologists?
A: Current tools assist with detection and prioritization. Radiologists remain responsible for interpretation and communication.
Q: How can a hospital evaluate an AI vendor?
A: Ask for peer reviewed evidence, regulatory status, local validation plans, integration details, and a monitoring plan. Involve clinicians and compliance staff early.