Inside the Microsoft Lab Teaching AI to Read Medical Charts in Twelve Languages

The lab itself doesn’t look particularly impressive. Rooms in a Redmond building with glass walls, whiteboards with partially erased Mandarin characters, and a coffee maker with the words “please descale me” written on it in three different languages. It’s more difficult to identify what’s going on inside. A group of engineers, medical professionals, and linguists are working to train a model to read a discharge summary in Tagalog fast, accurately, and without frightening the patient, just like an exhausted emergency room nurse might at three in the morning.

This is the peculiar and cautious part of Microsoft’s AI aspirations. For its U.S. launch, the company’s new medical assistant, Copilot Health, has received the majority of the media attention. However, the deeper wager is the multilingual chart-reading work, which the team describes with almost reluctance. It turns out that the most frequently mentioned topic on the Copilot mobile app is health. Furthermore, the majority of those individuals do not ask questions in textbook English.

Speaking with team members gives me the impression that they have witnessed the consumer AI race become noisy and do not want to be involved. The head of Microsoft’s AI division, Mustafa Suleyman, has described the strategy as “deliberate, slightly slower, more meticulous.” It sounded more like a confession than a catchphrase when he said it to The Wall Street Journal. The typical product-launch logic is broken by medical data. When someone’s potassium reading is the cause of your failure, you cannot fail quickly.

The technical work is truly peculiar. The PDF that a clinic in São Paulo emailed to a patient last Tuesday is not the same as a blood panel that was formatted by a Tokyo lab and printed in Lahore. Units are different. There are differences in reference ranges.

A value that is marked red in one country may be comfortably within the normal range in another, and the model must be able to recognize the similarities between “€����a,” “glucose,” and “血糖.” The engineers’ disagreement over whether to standardize units before or after translation serves as a reminder that, among other issues, medicine is a localization challenge that has never been fully resolved.

The Copilot Health product itself is a distinct tab within the app that is encrypted, removable, and isolated from standard chats. With authorization, it can retrieve data from over 50,000 U.S. hospitals via HealthEx, combine it with wearable data from gadgets like Fitbit and Apple Health, and create what the team persistently refers to as a “coherent story.” It’s a telling phrase. Diagnoses are not being sold by them. They are marketing a story that a patient could use and carry into a fifteen-minute appointment.

The contrast with the more vocal sectors of the AI industry is difficult to ignore. In terms of consumer chatbot usage, Microsoft lags behind OpenAI and Google; a quicker company would have released this a year ago. Rather, the system was examined by more than 230 doctors. Harvard Health co-wrote the answer cards. Citations can be seen. The National Academy of Medicine’s standards are used by the clinical team to assess credibility. It’s still unclear if that fosters trust or just slows adoption.

None of this is visible from outside the lab. It doesn’t matter which corporate division created the translator when a patient in Karachi is staring at an unknown lipid panel. They are concerned about the accuracy of the little green text beneath the number. One chart at a time, the lab is silently getting ready for that test in twelve different languages.