The health AI you want reads your wearable. Almost none of them do.

Hundreds of millions of people already ask ChatGPT and Claude for health advice. That is not the interesting number. The interesting number is zero: how many of those conversations can see your resting heart rate last night, your glucose curve after dinner, or your IGF-1 trend across a cycle.

You type symptoms into a text box. The model answers from the same generic prior it would hand a stranger. It has no idea what your body actually did this week. That is the gap. Demand for health AI is settled. The product that reads your own data is mostly still missing.

A generic chatbot is the wrong tool for a data-rich problem

The canonical health-AI use case is not "diagnose my rash." It is "my labs are normal and I still feel wrong." Data-rich person, clean clinical picture, unresolved functional problem. Whole communities run on this: someone posts full panels, everything in range, and dozens of strangers spend the thread troubleshooting what the labs cannot explain.

A generic chatbot cannot help there, because the answer is not in general medical knowledge. It is in the cross-variable pattern of that specific person over time, which the model has never seen.

And generic models are not just unhelpful in these cases, they can be dangerous. When Stanford tested AI chatbots on people in crisis, one answered a job-loss-plus-bridge-height query with a list of bridge heights. In clinical imaging, dropping a generic model into the workflow produced roughly four extra false positives per scan and pushed the review burden back onto staff instead of removing it. The pattern is consistent: a model with no grounding in your data and no clinician in the loop is worse than no model.

The demand is real, and it is specific

This is not a thesis waiting for validation. Builders keep saying it out loud. One developer with a nursing background put it plainly: the tools exist, the demand is clear, so where are the HIPAA-compliant health agents?

Meanwhile the data itself is going clinical. WHOOP started bridging continuous biometric data into Medicare reimbursement. An Oxford pilot validated consumer wearables (Oura, Whoop, Apple Watch) as inputs for a structured HRV study. Once wearable data enters clinical workflows, the layer that interprets it stops being a toy and starts being infrastructure.

The recurring insight underneath all of it: the market keeps selling information, and the real demand is delegation. People do not want another dashboard with sixty metrics. They want something to handle it.

Why wearable data is hard, and why most tools skip it

There are real reasons the generic chatbots avoid your wearable, and they are worth naming honestly.

• Devices disagree. A stress score from an HRV-only strap, a skin-conductance sensor, and a night-only ring are measuring structurally different things. Fuse them naively and you get confident nonsense. You need a normalization layer first.
• Access is partial. Pull wearable data through the wrong bridge and you can end up with a fraction of the real dataset. Full HRV and multiple daily readings usually require a direct integration, not a convenience API.
• Most metrics are noise. Apple Health exposes sixty-plus metrics, and only about eight have peer-reviewed mortality data behind them. An agent that treats all sixty as equally meaningful is worse than one that ignores fifty-two of them.

Reading a wearable well is not a plug-in. It is the hard part, which is exactly why the generic products skip it.

What a real wearable-data agent would do

Not diagnose. Not replace a clinician. It would sit in the execution and interpretation layer:

• Hold a baseline and watch for deviation. HRV, resting heart rate, and sleep, tracked continuously, catch shifts a single snapshot misses. Some compounds move resting heart rate by only two to three beats per minute, and continuous monitoring is the only thing that sees that at all.
• Ask the boring daily questions that make tracking actually stick. The sticking is the hard part, and a chat surface is genuinely good at it.
• Flag when subjective scales or resting heart rate drift off a person's own baseline, then route that to a human for any decision that carries real risk.
• Keep every read tied to a source and a rationale, so the reasoning is auditable instead of a black box.

The commercial shape that keeps surviving contact with reality is the same one: a model in front for context and continuity, a licensed clinician in the loop for anything that diagnoses or prescribes.

Where this touches peptides

This is not abstract for us. On a peptide cycle, attribution is the whole game. If you never measured a baseline, you cannot tell whether the compound did anything, whether a side effect is the peptide or something else, or whether the dose needs revisiting with your clinician. A wearable-data agent is the honest version of "did this actually work," and almost nothing on the market does it yet.

If you are running or planning a protocol, start with the unglamorous, correct version: set a baseline and measure attribution. Before you trust any vendor, product, or lab number attached to what you put in your body, learn how to vet it and understand what a peptide lab result actually proves. More of how we think about this is in the feed.