People now ask ChatGPT about 300 million health-related questions every week. Meanwhile, the largest multisite study of AI scribes to date found that the technology health systems have spent the most money and attention on saves the average clinician about 16 minutes of documentation a day.
Those two numbers describe a sector where adoption and impact are moving at very different speeds. What remains unsettled is which AI tools are changing how care gets delivered and which are mostly changing how it gets marketed.
AI analyst and author Hassan Taher ranks five of them below. The order answers a narrower question than market size or funding: how much of a real problem does the tool solve today, and how reliably can an organization confirm that it is solving it?
Taher scores each tool on three criteria.
Evidence of use. Sustained, voluntary adoption by the people the tool is meant to serve, rather than announced customer logos.
Verifiability. Whether a wrong output gets caught before it causes harm. A tool that writes a structured update to a system of record is easier to govern than one producing free text that a busy clinician is supposed to read closely.
Bottleneck fit. Whether the tool addresses the constraint that actually limits capacity, or simply the one that is most visible.
"Healthcare has a habit of buying the AI that demos well in front of physicians," Taher says. "The demo is a doctor talking and a note appearing on the screen. It's impressive. But the question a health system should ask is where its patients are actually getting stuck, and very often the answer is before they ever see a doctor. They're on hold, waiting for a callback, trying to reschedule, or trying to find out whether their insurance covers the visit."
OpenEvidence takes the top spot on adoption. The clinical search tool answers physician questions with responses grounded in peer-reviewed literature and clinical guidelines, and more than half of U.S. physicians now use it regularly, according to Becker's Hospital Review. In September the company raised $250 million at a $15 billion valuation, up from $12 billion in January. Mount Sinai and Sutter Health have embedded it in their Epic workflows.
What distinguishes OpenEvidence is that physicians largely adopted it on their own, without a procurement process. In healthcare, that is about as strong a signal of usefulness as exists. It also scores well on verifiability, because answers cite the underlying studies and a clinician can check the source in seconds.
The limitation: a citation only works as a check if someone reads it. A tool that is usually right gradually trains its users to stop verifying, and that erosion happens slowly enough that nobody notices it.
"OpenEvidence is a faster path to the literature, not a substitute for clinical judgment," Taher says. "The risk isn't the tool. It's the habit the tool builds."
The second-ranked tool targets the bottleneck that physician-facing AI ignores. RingCentral launched AIR Pro for Healthcare at HIMSS in March. It is an agentic voice AI platform that handles patient access work across phone, SMS, video, and messaging. Its pre-built skills cover patient identity verification, intelligent call routing, appointment scheduling and rescheduling, check-in, insurance verification, billing inquiries, and intake with basic triage. When a call needs a person, the agent transfers it to staff with the full interaction history attached.
Integration is the key. The platform launched with more than 80 integrations into EHR and healthcare management systems, including Epic, Oracle Health, athenahealth, and eClinicalWorks. That lets the agent complete a task rather than take a message. On a rescheduling call, it verifies the patient's identity, checks provider availability, updates the system of record, and sends an SMS confirmation, all in one interaction.
That design scores high on Taher's verifiability criterion. A rescheduled appointment either appears correctly in the scheduling system or it does not. The platform also keeps full transcripts and conversational analytics, giving compliance teams an audit trail, and it runs on what RingCentral describes as a HIPAA-ready architecture with role-based permissions. PM Pediatrics, a pediatric urgent care provider with 80 locations, already uses RingCentral's AI portfolio and has reported a measurable impact on how it manages patient calls.
The limitation: AIR Pro for Healthcare launched in controlled availability for selected U.S. organizations, with broader availability planned for later in 2026, so independent outcome data is still limited. Pricing requires a quote, and organizations get the most value when they already run their communications on RingCentral.
"Nobody writes a JAMA paper about the patient who hung up after eleven minutes on hold, but that patient didn't get care either," Taher says. "Patient access is high-volume, repetitive, rules-based work with a clear right answer, which is exactly where agentic AI is strongest. The tasks are bounded. Did the appointment move? Was the insurance verified? Did the patient reach a human with the context intact? You can measure every one of those. That makes the front door one of the few places in healthcare where an organization can deploy AI agents and actually know whether they're working."
Radiology accounts for 1,039 of the more than 1,300 AI-enabled devices on the FDA's list, nearly 80% of the total, as of the agency's December 2025 update. Vendors including Aidoc, Qure.ai, GE HealthCare, and Siemens Healthineers offer tools that flag suspected strokes, pulmonary embolisms, and brain bleeds, then move urgent studies to the top of the reading queue.
No category has had more regulatory review, and triage tools address a genuine bottleneck: the gap between a scan and a radiologist's read.
The limitation: FDA clearance establishes that a device performs a defined task to a defined standard. It does not establish that patient outcomes improve once the device is installed. Many hospitals now run several narrow algorithms from different vendors, and each one needs its own validation and ongoing monitoring.
"Imaging AI is the most mature category and the least glamorous, which is usually a good sign," Taher says. "The open question isn't whether these algorithms work in isolation. It's whether a hospital running fifteen of them has anyone responsible for watching all fifteen drift."
Ambient scribes, led by Abridge, Microsoft Dragon Copilot, and the AI Charting feature Epic rolled out in February, have been the most heavily marketed category in healthcare AI. The best evidence suggests the benefit is real and modest.
The JAMA study published in April tracked more than 1,800 clinicians using AI scribes at five U.S. health systems against 6,770 controls at the same institutions. Scribe users spent 16 fewer minutes per day on documentation, a 10% reduction, and saw about half an additional patient visit per week. After-hours EHR time did not change. Only 32% of users relied on the tool in more than half of their visits, and that group saw roughly three times the documentation savings.
"The modest reductions in documentation time we observed are unlikely to fully account for changes in burnout," said Rebecca Mishuris of Mass General Brigham, the study's senior author.
The limitation: verifiability. Scribes produce free-text notes that become the official record, and errors such as dropped negations or substituted drug names are easy to miss in a quick review. Healthwatch England documented exactly those failures in NHS records this summer. It is the same challenge of hallucination in AI systems that Hassan Taher has described, arriving in a document clinicians rely on.
"I rank scribes fourth, not last, because sixteen minutes a day matters to a clinician who is underwater," Taher says. "But the industry sold them as the cure for burnout, and the evidence says they're a partial treatment with a side effect that needs monitoring. Health systems should deploy them for the clinicians who will use them in most visits, and they should fund someone to audit the notes."
OpenAI's ChatGPT Health, first tested in January and opened to all U.S. adult users in July, lets people connect data from Apple Health and MyFitnessPal along with medical records from systems such as Epic and Oracle Health. OpenAI says weekly health-related queries on ChatGPT have grown from 230 million to 300 million.
No healthcare tool comes close to that reach.
The limitation: OpenAI states that the service is not intended to diagnose or treat health conditions, which places it outside the clinical oversight that governs everything else on this list. There is no clinician in the loop to catch a confident wrong answer before a patient acts on it.
"It ranks last on my criteria and first on raw impact, and that tension is the point," Taher says. "It's the only tool here where the person checking the output is also the person least equipped to check it. That isn't an argument against it. It's an argument for health systems to meet patients where they already are, instead of pretending those conversations aren't happening."
The tools near the top share a trait: their outputs can be checked. Further down the list, outputs become harder to verify and the burden of catching errors shifts to whoever has the least time to do it.
Health systems focused on clinician workload will find ambient scribes useful if they target heavy users and budget for auditing. Organizations struggling with access, call volume, and no-shows will see faster, more measurable returns from front-door automation like AIR Pro for Healthcare. Every buyer should expect the gap between pilot and production that Hassan Taher has examined in the MIT finding that most AI projects fail to deliver measurable returns.
"The next phase of healthcare AI is agents that don't just answer questions but take actions inside the EHR, the scheduling system, and the billing platform," Taher says. "Once AI is acting rather than advising, verifiability stops being one criterion among several. It becomes the whole evaluation."
This article discusses healthcare technology for informational purposes. It is general commentary and not medical, legal, or investment advice.
[NGpg/VP]
Suggested Reading:
Subscribe to our channels on YouTube and WhatsApp
Download our app on Play Store