Microsoft’s Copilot Health Wants Your Medical Records, And That’s Raising Big Privacy Questions

Microsoft is pushing deeper into health care with a new consumer feature called Copilot Health, an AI chat experience designed to pull together your medical history, lab results, and data from wearables to deliver personalized guidance.

The tool is rolling out in the U.S. by invitation only, a cautious debut for a product operating in one of the most sensitive arenas for AI: people’s bodies and their private health data. Microsoft says it’s not trying to replace doctors. Critics and security experts are focused on a different question: what happens when a tech giant becomes the hub for your most intimate information, and gets something wrong?

Copilot Health is pitched as a “concierge medicine”-style helper without the concierge price tag, more like a prep coach for your next appointment than a digital physician. But the promise of tailored answers depends on something Americans are increasingly wary of handing over: their health records.

Invitation-only launch in the U.S., inside Microsoft’s consumer Copilot

Copilot Health appears as a dedicated health portal and chat tab inside Microsoft’s consumer Copilot, separate from the Copilot tools Microsoft sells to businesses. The invitation-only rollout limits early traffic and gives Microsoft time to see how people actually use it.

Users are encouraged to upload structured health information, medical history, documents, and data from connected devices, so the chatbot can respond with more context than a general-purpose AI. The idea is straightforward: “persistent fatigue” means something different depending on age, medications, past diagnoses, and activity levels.

In a demo using fictional data, the system flagged a high-risk scenario, jaw pain in someone with a prior heart attack, and urged an in-person evaluation the same day. That’s the kind of moment where an AI assistant has to steer people toward urgent care, not reassurance.

Mustafa Suleyman, who leads Microsoft’s consumer AI efforts and previously co-founded DeepMind, has framed the product as a way to help people prepare for doctor visits: describe symptoms more clearly, understand test results, and show up with better questions.

The most realistic day-to-day use may be appointment prep. Someone tracking fluctuating blood pressure on a smartwatch, for example, could ask what details to log, sleep, caffeine, stress, and what to ask their clinician. The line Microsoft is trying to walk is thin: help without diagnosing, guide without prescribing.

Wearables plus medical records: the bet on HealthEx and data aggregation

The engine behind Copilot Health is data ingestion. Microsoft says it can integrate with more than 50 devices and tracking services, including Apple Health, Oura, and Fitbit. For many users, that means daily metrics like heart rate, sleep, and activity, and, on some devices, blood oxygen levels.

On the medical-records side, Microsoft says Copilot Health connects through HealthEx to more than 50,000 hospitals and health care organizations across the U.S. That matters because without access to clinician notes, diagnoses, and medication lists, AI health advice tends to collapse into generic internet-style guidance.

Microsoft also points to lab-result integration through a service called Function, opening the door for users to ask what a test measures, what can influence a number, and what questions to bring to a doctor. The challenge is obvious: explaining results without crossing into clinical interpretation that should come from a licensed professional.

Microsoft says Copilot Health provides citations and links to sources it considers credible, an attempt to address a core problem with general chatbots, which can deliver confident-sounding misinformation. But even well-sourced information can be misapplied if key details are missing, pregnancy, a medication change, or a past condition a user forgets to mention.

Microsoft promises encryption and separation of health data, but HIPAA questions linger

Microsoft is emphasizing security from the start. The company says data entered into Copilot Health is encrypted and protected with additional internal security controls. It also says health data is kept separate from other Copilot conversations, meant to reduce the risk that a chat about travel plans or bills bleeds into medical context through memory or suggestions.

Microsoft also says Copilot Health data won’t be used to train its AI models, and that users can delete their information at any time. Those assurances are increasingly standard in consumer AI, but they still require users to understand settings, deletion mechanics, and the difference between storage, logging, and product improvement.

Company executives have been careful about positioning. Dominic King, a Microsoft vice president who has worked closely with Suleyman, has stressed the importance of avoiding triage mistakes, cases where the system falsely reassures someone who needs care, or panics someone who doesn’t. In medicine, those errors don’t land evenly: a false negative can delay treatment; a false positive can send people flooding into ERs and urgent care.

One notable gap: Microsoft has described Copilot Health as secure, but public messaging reported so far has not clearly stated HIPAA compliance. HIPAA, the U.S. health privacy law that governs many providers and insurers, can be a make-or-break signal for hospitals and clinics deciding whether a tool fits inside regulated workflows.

Microsoft says 230 doctors helped shape guardrails, but real-world use is the test

To bolster credibility, Microsoft says it built an internal clinical team and consulted an external panel of more than 230 physicians across 24 countries. The goal: pressure-test answers, safety guardrails, and high-risk symptom scenarios.

Microsoft also cites content partnerships with Harvard Health and UpToDate, a widely used clinical reference published by Wolters Kluwer. High-quality sources can reduce wild advice, but they don’t solve the hardest part: context. The same symptom can be harmless or life-threatening depending on the patient’s history and what they’re not saying.

Microsoft’s own demo example, jaw pain after a prior heart attack, shows what’s at stake. It’s easy to get that right in a controlled setting. The harder challenge is consistency across millions of messy, incomplete, emotionally charged conversations.

A fictional user example in the original reporting captured the demand: a 43-year-old with Type 2 diabetes who wants help sorting questions before calling a primary care doctor, not a bot that tells him to panic. That’s the sweet spot Microsoft is aiming for, and also where overconfident AI can do real damage.

Big Tech is racing to own the “front door” to health advice

Microsoft is entering a crowded field. OpenAI announced ChatGPT Health in January, including tools to upload records and sync health-app data. Amazon, through its primary-care company One Medical, launched a Health AI assistant around the same time, aimed at health questions, scheduling, and medication management.

Microsoft argues the demand is already massive: the company says its AI products, including Copilot and Bing, handle more than 50 million health-related questions every day. Microsoft also says that in January, 41% of health conversations were general information, 11% were about symptoms, 9% focused on fitness coaching, and 8% involved conditions and care.

Even more telling, Microsoft says roughly 1 in 5 of those health conversations include personal data, symptoms, test results, and other sensitive details. In other words: Americans are already using chatbots as a first stop, often because getting a quick appointment can be hard.

One of Copilot Health’s most practical features may be the least controversial: it can search real-time directories to help users find clinicians by specialty, location, languages spoken, and insurance coverage. That’s actionable without pretending to be a doctor. But it doesn’t solve the underlying bottleneck, available appointments. If wait times stay long, tools like this risk becoming a permanent substitute, increasing the temptation to treat AI guidance as definitive medical advice.