An AI health chat is wellness, CDS or a device depending on what the user asks — and what the model says back.
A conventional device can be scoped by its feature list. A health chat cannot. Its regulatory exposure is defined at runtime by two variables — what the user asks, and what the model says back — so the classification has to be done on the conversation, not the product.
Three regulatory positions are available to a consumer-facing or clinician-facing AI chat in the United States, and they are not a spectrum. General Wellness is an enforcement-discretion policy for low-risk products that promote a healthy lifestyle and make no disease claims. Non-device clinical decision support is a statutory carve-out, FD&C Act §520(o)(1)(E), available only when all four criteria are met — including that the software is for a health care professional and that the professional can independently review the basis for the recommendation. Everything else that meets the device definition is Software as a Medical Device, with a class, a pathway and an evidence burden.
The mistake teams make is to pick one position and write a disclaimer around it. FDA has said in warning letters and in its CDS guidance that it looks at what the software actually does — the output, the framing, the user — not at the terms of service. So the honest way to scope a chat is to map every plausible exchange onto the three positions and design the product to stay where you intend it to be.
The query × response matrix
Rows are what the user brings to the conversation. Columns are what the model does with it. Each cell is where FDA would most plausibly place that exchange; click a cell for the reasoning.
What the matrix says
The consumer-facing chat has no yellow. The CDS carve-out requires a health care professional as the intended user. For a consumer-facing product every yellow cell collapses to red, which means the only safe territory is the green band — general and personalized information that stays non-directive and makes no disease claim.
Personalization is not the line; directiveness is. A model can reference the user's own data and remain wellness so long as it educates rather than recommends. The moment the output tells a specific person to do a specific thing about a specific condition, it has crossed the activity axis regardless of how gently it is worded.
The user's question can move you. If a user describes a symptom or names a diagnosis, the same educational response that was green for a general question is now, at best, yellow. That is why a chat needs a query classifier in front of the model, not just a style guide behind it.
Design constraints that hold the position
| Constraint | Why it matters | Where it comes from |
|---|---|---|
| Classify every incoming query into a traffic-light tier before the model answers | The user, not the product team, sets the row. The tier decides the permitted column. | General Wellness policy · CDS guidance criterion 1 |
| Green tier: educate and contextualize; never recommend, dose, or triage | Keeps the exchange in the informational, non-directive cell FDA treats as low risk | Wellness policy · GenAI discussion paper activity axis |
| Yellow and red tiers: redirect to a clinician with the user's own data summarized, not interpreted | Summarizing is transfer; interpreting is analysis. The line is the one FDA draws for MDDS versus CDS. | §520(o)(1)(D)–(E) |
| No numeric risk, reference range, or in/out-of-range presentation for a consumer | FDA has treated a color-coded gauge itself as evidence of device intent in a 2025 warning letter | FDA warning letters · CDS guidance examples |
| Test the guardrails against the taxonomy, not against a prompt list | A hazard-and-test method per tier is what makes the position auditable when the model changes | ISO 14971 · GenAI paper testing methods |
Building a health chat?
We map your actual conversation logs onto this matrix and hand you the query taxonomy, the response constraints and the test method — signed.