FDA just published the map to a fully autonomous AI doctor. It didn't draw the route.
Ery Anguiano, Principal·5 min read·Docket FDA-2026-N-7874 · comments close 19 October 2026
FDA's discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical Devices (18 August 2026), does more than solicit comment. It is the beginning of a framework.
There are plenty of tactical details still to bake out, but the paper shows the agency's current thinking on how a model's behavior drives risk, and how risk drives the evidence plan. That alone makes it required reading for anyone building or funding a generative-AI health product.
More importantly, it shows FDA is not breaking any fundamental principle in its mission to assure safety and effectiveness. There are some novel testing methods proposed — standardized-patient testing, essentially licensure exams for the model — and novel monitoring methods, including machine-based supervisory agents: a swarm of models watching the model. But the fundamental risk analysis, the potential clinical consequences of the device's behavior, is unchanged.
That should be music to the ears of anyone who worried generative AI would get a wild-west pass: output anything, claim anything, as long as you disclaim it. It will not.
What this means for a product teamThe regulatory question for a generative-AI feature is not "is it an LLM?" It is the same question FDA has always asked: what does the output cause a person to do, and what happens if the output is wrong? Scope the feature by the clinical consequence of its behavior and the rest of the framework follows.
→ Move right · activityM1 The directiveness ladder D1→D4 — "sometimes increased" becomes "increase to 20 mg."
↑ Move up · consequenceM3 Measurement, signal or IVD output the user cannot evaluate. M4 Patient-facing rather than HCP-facing. M5 Generalist rather than specialist HCP. M8 Agentic, multi-step, tool-using.
⊘ No movementM2 Disclaimers do not de-escalate — "talk to your doctor," "I am not a medical professional."
↔ Change the unit of analysisM6 Multi-turn migration: assess trajectories, not single functions. M7 Bidirectional escalation error: under- and over-escalation both count.
? Raised, not adoptedM9 Reversibility · downstream safeguards · time pressure · output traceability.
One mechanismM3, M4 and M5 all reduce to one question FDA ties to the CDS exclusion, FD&C Act §520(o)(1)(E): can the user independently evaluate the output?
Figure 1 · The twelve-cell scenario spaceFDA's two axes — activity and consequence — with the agency's center of gravity in the action-directing, severe-consequence cell, and the nine modifiers FDA names that relocate a function without changing what it does. The activity axis is a cliff: crossing from information into action is the CDS exclusion line. The consequence axis is a slope: it scales evidence. Cell risk labels are interpretation.
The fourteen worked examples in the paper, plotted across FDA's risk matrix, are real benchmarks for any regulatory professional working with a product team. They tell you where the agency currently places a given combination of autonomy, directiveness and clinical stakes — and therefore what evidence it will expect.
Non-directiveInformation that does not point at an action — "a risk score for a future cardiovascular event."Action-directingA function that, "while nominally informational, directs an action" — e.g. "a strongly worded patient-facing recommendation about whether to seek emergency care."Not a binary"The distinction … may not be binary but rather would likely involve a continuum." Four rungs: general information → connect practice to the user → endorse an action → give an instruction.Action-taking"The affirmative assignment of a clinical diagnosis, the prescription of a medication, or the initiation of a clinical order set." Those are the only three examples FDA names.A3 versus A4A function that "acts with continuous HCP supervision" versus one that "acts in fully autonomous fashion." FDA never defines "continuous."
Figure 2 · FDA's worked examples on FDA's gridAll fourteen illustrations from the discussion paper, quoted verbatim and placed on the agency's own framework. Dashed card: the one example FDA declines to locate on a single side of the line. Note the empty cells — no non-directive severe example, nothing placed at limited consequence for non-directive information, and nothing yet for HCP-supervised action below severe consequence.
As an exercise, I plotted some of the most visible generative-AI health features on the same grid to see where they land. The result is a decent snapshot of where the market actually sits relative to where FDA has drawn its lines.
But if a fully autonomous AI doctor is the goal, the best use of this matrix is to build a coherent roadmap from it — one that sequences the indications, the model's permitted actions, and the evidence plan required to move from one cell to the next. Whoever builds that roadmap first, and helps FDA develop the test methods along the way, is the one who ends up building the AI doctor.
Figure 3 · Four moves to a fully autonomous AI doctorEach move is one cell, and none is a leap. Prove action-directing at severe consequence; cross into action-taking with a human ratifying every output; remove the human by narrowing consequence rather than scope; restore consequence at full autonomy. Every terminal cell already holds an FDA worked example. Stage sequence and pathways are planning assumptions, not commitments.
Primary source. U.S. Food and Drug Administration. Considerations for the Regulation of Generative AI-Enabled Medical Devices. Discussion paper, 18 August 2026. Docket FDA-2026-N-7874. Comment period closes 19 October 2026.
This analysis reflects the author's reading of a public document and is not legal advice. It does not describe any client engagement.
Building a generative-AI health feature?
We map the feature against FDA's matrix and tell you which cell you are in, which cell you want, and what it takes to get there.