For years, one of the stranger constraints in health-tech product design came straight out of an FDA guidance document. If you wanted your clinical decision support software to stay outside FDA's definition of a medical device, and therefore outside premarket review, you had to make it offer the clinician more than one recommendation. A list of options, not a single answer. The logic was that a list invites the clinician to think, while a single directive invites them to just comply.

The result was a lot of products engineering artificial choice into tools where the clinical guideline pointed to one obvious path. You'd build software that knew the right answer and then dress it up with alternatives nobody would pick, purely to stay on the right side of a regulatory line.

In January 2026, FDA revised the Clinical Decision Support Software guidance and dropped that requirement. The change is narrow, the framing is dry, and the implications for anyone building clinical AI are larger than they first appear.

What actually changed

The guidance interprets Section 520(o)(1)(E) of the Food, Drug, and Cosmetic Act, the provision that carves certain clinical decision support out of device regulation entirely. To qualify as non-device CDS, software has to meet four criteria, and the relevant one here governs what the software outputs and whether the clinician can independently review the basis for it.

Under the 2026 revision, issued January 6, FDA will now exercise enforcement discretion for software that gives a single recommendation, provided only one recommendation is actually clinically appropriate. If multiple options are genuinely valid, you still have to present them. But you're no longer forced to invent alternatives when the evidence points one way.

FDA also added examples that lean directly into AI-shaped use cases. Software that analyzes a radiologist's findings to generate a proposed report summary can fall under enforcement discretion. So can tools that summarize patient data into something like discharge papers. The condition is consistent throughout: the clinician has to stay in the loop and be able to trace the recommendation back to "well-understood and accepted sources."

One more structural move worth noting. The 2022 guidance had a line excluding "time-critical decision-making" from non-device status. The 2026 version removed that phrase, but the restriction didn't disappear. It moved into the analysis of whether a clinician can independently review the basis for a recommendation, which is hard to do in the seconds before a time-critical decision. Same outcome, relocated. Don't read the deletion as a loosening for emergency-care tools.

Why the single-recommendation change matters more than it sounds

The old multiple-options rule wasn't just an annoyance. It actively distorted product design in a direction that was probably worse for patients.

From a behavioral standpoint, padding a clear recommendation with throwaway alternatives doesn't make a clinician deliberate more carefully. It adds noise to a decision that the evidence had already resolved, and it trains users to ignore the extra options as boilerplate. The regulatory requirement to show choice was producing fake choice, and fake choice erodes trust in the tool faster than a confident single answer does.

The flip side is the part teams need to be honest about. A single recommendation is more persuasive, and more persuasive means more automation bias. When software says "do X" instead of "consider X, Y, or Z," clinicians defer to it more readily, including when it's wrong. FDA's revised guidance leans hard on transparency precisely because of this: the price of being allowed to give one answer is that the clinician must be able to see and verify why. The reviewability requirement isn't legal boilerplate. It's the mechanism that's supposed to keep a single recommendation from becoming a black-box order.

So the practical design mandate shifts. The work isn't "generate plausible alternatives anymore." It's "make the basis for the one recommendation legible." Show the source. Show the patient-specific inputs that drove it. Make it fast to audit, because the clinician now has fewer cues prompting them to question the output.

The bigger story is what the guidance won't say

Here's the thing that struck me reading the analyses across the law firms covering this. FDA announced the update as an AI-focused modernization, and then the document is almost entirely silent on AI.

Multiple firms made the same observation. Covington noted the guidance offers no new policy for AI-enabled products. Orrick's read was that the qualifying line for AI is whether clinicians can understand and verify the underlying logic and data inputs, and that where models are opaque, time-critical, or directive, device oversight should still be expected.

That second clause is the one to sit with. The entire non-device carve-out hinges on the clinician being able to independently review the basis for a recommendation. A large language model that synthesizes a recommendation from a patient's chart and the medical literature is, almost by construction, hard to "independently review the basis" for in the way the statute imagines. The reasoning is a generated output, not a transparent chain a clinician can audit line by line.

So the guidance creates a real tension. It opened the door to single-recommendation AI summarization tools, and it conditioned that opening on a reviewability standard that generative models are structurally bad at meeting. Summarizing a radiologist's own findings into a report? That's reviewable, because the source is right there. A clinician-facing chatbot that produces a diagnostic recommendation from scratch? Much harder to argue the clinician can trace the basis, and that's exactly the category FDA declined to address.

What I'd do if I were building here

A few concrete things follow from this for product teams.

Treat reviewability as a product feature, not a compliance checkbox. The regulatory pathway and the right design choice happen to point the same way. Whatever your AI recommends, the interface needs to surface the inputs and sources behind it in a way a busy clinician can actually use in the moment. If you can't show the basis, you probably can't claim non-device status, and you also shouldn't be shipping a directive tool.

Know which side of the summarization line you're on. Tools that summarize or restructure information a clinician already has access to are sitting in the comfortable part of the new guidance. Tools that generate novel clinical recommendations from raw data are in the gray zone FDA pointedly left open. Those are different regulatory risk profiles and they should be different product decisions, made on purpose.

Don't assume the silence on LLMs is permanent. Several of the firms reading this guidance flagged that the omission likely signals more action coming. Building a generative clinical recommendation product on the bet that FDA will stay quiet is building on a timer. The defensible move is to design for the reviewability standard now, because whatever comes next is going to lean on it harder, not softer.

Resist the temptation to over-read the time-critical change. The phrase came out, the restriction stayed. If you're building anything that touches emergent or near-real-time clinical decisions, assume device oversight applies and plan accordingly.

The single-recommendation change is genuinely useful. It lets teams stop building fake choice into tools that knew the answer. But the more important takeaway is the shape of the gap FDA left. The agency told you how a transparent, source-grounded, clinician-reviewable recommendation engine can stay out of device regulation, and it conspicuously didn't tell you what to do with a model whose reasoning you can't fully show. For most teams shipping clinical AI, that gap is where the real product and regulatory work lives.

Primary source: FDA, Clinical Decision Support Software guidance (revised January 2026), summarized via the American College of Radiology. https://www.acr.org/News-and-Publications/2026/fda-updates-guidance-on-clinical-decision-support

Legal analysis: Covington & Burling, "5 Key Takeaways from FDA's Revised Clinical Decision Support (CDS) Software Guidance" (January 2026). https://www.cov.com/en/news-and-insights/insights/2026/01/5-key-takeaways-from-fdas-revised-clinical-decision-support-cds-software-guidance

Legal analysis: Orrick, "FDA Eases Oversight for AI-Enabled Clinical Decision Support Software and Wearables" (January 2026). https://www.orrick.com/en/Insights/2026/01/FDA-Eases-Oversight-for-AI-Enabled-Clinical-Decision-Support-Software-and-Wearables

Technical analysis: Hardian Health, "The FDA's 2026 Clinical Decision Support (CDS) Guidance Update" (January 2026). https://www.hardianhealth.com/insights/fda-2026-clinical-decision-support-c-guidance-update