Consider a single sensor stream: the optical signal a wrist-worn device uses to estimate blood pressure. In July 2025, the FDA sent a major wearable maker a warning letter over precisely this, on the ground that estimating blood pressure is "inherently associated with the diagnosis" of hypertension — a clinical claim, and therefore a regulated one.[1] Eighteen months earlier the same feature would have drawn the same objection. Six months later, it did not: in January 2026 the FDA revised its General Wellness guidance to permit non-invasive wearables that estimate physiologic parameters like blood pressure as wellness products, provided they make no disease claims, and by June 2026 the agency had closed the warning letter after the company adjusted the feature.[2][3] Same signal, same device category, three different regulatory answers in eighteen months. Nothing about the physics changed. What changed was the line.
This is the first thing to understand about the boundary between wellness and clinical AI: it does not run where the technology is, but where the claim is. The FDA's framework turns on intended use. A product that helps someone manage general fitness sits in the wellness lane; the moment it claims to diagnose, treat, or drive a specific clinical action, it crosses into device territory and the regulatory obligations that come with it.[2][4] The same heart-rate signal is wellness when it informs a workout and a medical device when it triggers a diagnostic alert. The boundary is drawn in language and purpose, which is why two products with identical sensors can sit on opposite sides of it.
The second thing to understand is where the line becomes bright rather than blurry. FDA's clinical-decision-support guidance sets out criteria that keep software on the non-device side, and one of them is categorical: software that analyzes a signal from a signal-acquisition system — a continuous ECG waveform, a glucose stream, an imaging pattern — is a regulated device, regardless of how carefully it frames its claims.[5] Discrete, point-in-time measurements generally do not cross this line; continuous streams analyzed for a pattern do. This is the distinction between surfacing information for a clinician and making the clinical determination — and it is the one that a platform running AI on biosignals has to respect in its architecture, not merely in its marketing. Software that surfaces and informs, leaving the consequential call to a clinician, can stay on the non-device side. Software that analyzes raw waveform to reach a time-critical conclusion cannot.
The third thing — and the reason this matters for how a product is built — is that the line moves, and it moves in both directions. The past year is the proof. FDA tightened enforcement against a wellness product making a blood-pressure claim, then loosened the guidance that governed it, then closed the case.[1][2][3] Meanwhile the deregulatory turn at the FDA did not retire the risk; it relocated it. A consumer class action followed the warning letter, using the FDA's own finding as its foundation.[6] States wrote their own rules: Illinois barred AI from delivering therapy without a licensed professional, and by early 2026 several states had enacted disclosure and practice constraints on AI in mental health.[7] A product that hard-codes today's version of "the line" into its logic is building on ground that has visibly shifted three times in eighteen months and is being redrawn by regulators the FDA does not control.
This is the case for treating the boundary as a governed policy rather than a fixed feature. If the line between wellness and clinical use is enforced, consequential, and mobile, then the defensible architecture is one where an organization defines the boundary as an explicit policy and the platform enforces it on every output — checking each AI output against the lane it is permitted to operate in, allowing what is in bounds, rewriting what drifts toward a clinical claim it is not cleared to make, and escalating what needs a human. When the line moves, as it moved twice in the past year, a policy can be updated. Logic compiled into a product cannot, short of a release cycle that regulation will not wait for. The distinction between wellness and clinical AI is not a one-time classification a product passes at launch. It is a boundary a product has to keep observing as its own claims evolve and as the regulators — plural — redraw it.
We take this seriously because we operate exactly on the frontier where the line is contested: AI running on biosignals, near patients, inside institutions that answer to the FDA, to state regulators, and to their own counsel. Our position is that the governance of that line is not a compliance footnote but the durable part of the work. Models will get more capable, and the wellness-clinical boundary will keep moving as they do. The layer that keeps every output on the right side of a line that will not hold still — that enforces the policy, preserves the audit, and adapts when the rule changes — is the one that lets an institution adopt AI in care without betting that today's version of the line is the last one. It will not be.
Sources
- FDA, Warning Letter to WHOOP, Inc., July 14, 2025 — fda.gov
- FDA, "General Wellness: Policy for Low Risk Devices" (reissued Jan. 6, 2026) — fda.gov
- STAT, "FDA drops enforcement against wearable maker Whoop," June 23, 2026 — statnews.com
- Ropes & Gray, "FDA Adapts with the Times on Digital Health: Updated Guidances on General Wellness Products and Clinical Decision Support Software," Jan. 2026 — ropesgray.com
- FDA, "Clinical Decision Support Software" guidance (current version, Jan. 2026) — fda.gov
- ArentFox Schiff, "Whoop There It Is: FDA Warning Letter Now Anchors a Class Action Against Whoop" (Rowe v. Whoop, N.D. Cal.), 2025 — afslaw.com
- Holland & Knight, "New Illinois Law Restricts Use of AI in Mental Health Therapy," Aug. 2025 — hklaw.com
