I'm a senior AI and full-stack engineer. For the past year I have been building a health AI, and this week I read up on what everyone else ships. The first thing I noticed is that "health AI" is used for four different products. They fail in different ways, so each needs a different kind of engineering.
1. The scribe
It listens to a visit and writes the note. Ambient scribes such as Heidi sit here.
The failure to design for: a line in the record that was never said, written in a calm clinical voice. My rule would be that nothing reaches the chart without a clinician reading it, and the transcript stays next to the note so that reading is real and not a click-through.
2. The answer engine for clinicians
A doctor asks a question and gets an answer with citations. OpenEvidence and Vera Health are examples.
According to STAT (January 2026), OpenEvidence is free to US clinicians and advertising is its main business model; the same report covers a $250 million raise. Vera Health says on its site that it searches more than 60 million peer-reviewed papers and links every key statement to its original source.
The failure to design for: a citation that exists but does not support the sentence next to it. A link is not evidence. The check I would build is a test that takes each claim and its cited passage and asks whether the passage actually says it.
3. The chart copilot
It reads the whole record so the doctor does not have to. Navina describes itself on its site as pulling records from the EHR, health exchanges and claims data into one view, and surfacing evidence for what it flags.
The failure to design for: noise. A copilot that flags everything trains people to ignore it. The check is boring and measurable: of the things it flagged, how many mattered, and what did it miss?
4. The patient assistant
This is the one I build, so I'll say it plainly. Tabibu is a free health-information assistant that started in Kenya. The person on the other end is not a clinician and cannot judge the answer.
The failures to design for are false reassurance and dosing. So the rules are written as tests, not as hopes: it does not diagnose, does not call a result normal, and does not tell you to start, stop or change a medicine. It has no clinician review yet. That limit is listed on the site, because a limit you hide is worse than one you publish.
What this changed for me
Before this I picked the model first and added safety later. Now I pick the failure first. Ask "what is the worst believable wrong answer for this kind of product?", then build the check that would catch it.
It also matters that safety work is real but not magic. In a 2026 Mount Sinai study of 20 AI models, published in Communications Medicine and described in a Mount Sinai press release, a short safety reminder cut potentially harmful clinical choices from 16.6% to 10.1%. That is progress. It is also not zero.
Which of the four do you think is hardest to get right? I would like to hear where you disagree.
Descriptions of other companies come from their own sites or from the press named above; I have not tested those products. General information, not medical advice.
I post short videos in this series on LinkedIn and TikTok; come say hi.
