The FDA loosened its AI health rules. That's not the same as no rules.

For most of the last decade, the regulatory current in healthtech ran one way: toward more oversight. Every new AI feature came with the same quiet question. Does this turn our software into a medical device the FDA has to clear before it can ship? In early 2026 the agency started rowing the other direction.

On January 6, the FDA released revised guidance that pulls a large category of clinical software, and a long list of consumer wearables, out from under its review. It finalized the clinical-decision-support piece a few weeks later. Commissioner Marty Makary framed it as cutting red tape. The substance is that products which used to live in a regulatory gray zone now have a clearer route to market without a premarket submission.

If you're building something that reads patient data and offers a suggestion, this is good news with a sharp edge. Looser federal review doesn't mean looser standards. It means the FDA is stepping back from the gate and leaving more of the judgment, and more of the liability, with you. The teams that come out ahead are the ones who know exactly which line they're standing on, because the line moved. It didn't disappear.

What actually changed

The headline change is about clinical decision support: software that takes patient information and hands a clinician a recommendation. Since the 21st Century Cures Act, this kind of software has been able to escape "device" status if it clears four conditions. It doesn't analyze a medical image or signal directly. It works from existing medical information. It supports a clinician's judgment instead of replacing it. And it shows its reasoning clearly enough that the clinician can review the basis rather than take the output on faith.

What's new sits on top of that test. The old reading was that a tool offering a single recommendation looked like it was directing the decision, which nudged it toward device territory. The 2026 guidance says the FDA will generally leave a single-output tool alone when only one option is clinically appropriate anyway. Helpful. The problem is that the FDA never defined "clinically appropriate," so the boundary you're being handed is real but fuzzy, and you're the one who has to decide which side of it your feature lands on.

The second change is for consumer hardware. The revised general wellness policy gives much wider berth to wearables that report things like heart rate, blood pressure, and blood glucose, as long as they're pitched for general wellness rather than diagnosing or treating a condition. The whole exemption lives in the claim. "Track your heart rate trends" is wellness. "Detect atrial fibrillation" is a medical claim, and a medical claim drags you straight back under FDA oversight however casual the marketing copy sounds.

What didn't change, and where the line still is

Plenty of software is still squarely a device, and the guidance is explicit about it. If your tool analyzes a medical image or a raw signal like an ECG waveform to produce its recommendation, you're regulated. If it predicts a time-critical event, say the risk of a cardiac event in the next 24 hours, you're regulated even if it only outputs a single number. Anything built to diagnose from an image is regulated. The deregulation is real, but it carved out the lower-risk middle and left the genuinely high-stakes work exactly where it was.

Here's the part the celebratory coverage tends to skip. When the FDA reviews a device, that review is also a shield. Clearance is the agency on record saying your evidence was good enough. Take the review away and you don't lose the obligation to be safe and effective, you lose the paperwork that proved you were. If a clinician leans on your unreviewed recommendation and a patient is harmed, "the FDA decided we didn't need clearance" is a thinner thing to stand on than "the FDA cleared us." Less regulation can quietly mean more exposure.

So the work doesn't vanish. It changes shape. Instead of assembling a submission for the FDA, you're assembling the evidence for yourself: where your training data came from, how you validated the model, how you show a clinician the reasoning behind a recommendation so they can actually exercise the independent judgment the law assumes they will. That fourth Cures Act criterion, explainability, stops being a regulatory checkbox and becomes the thing standing between you and a product liability claim.

TEMPO: a new on-ramp, not a free pass

Alongside the loosening, the FDA opened a door that didn't exist before. The TEMPO pilot, short for Technology-Enabled Meaningful Patient Outcomes, lets manufacturers put certain digital health devices in front of real patients before full authorization, under the agency's enforcement discretion on premarket and investigational-device requirements. It's paired with a CMS payment model called ACCESS, a ten-year program starting in mid-2026 that pays for managing chronic conditions based on outcomes instead of visits.

The pilot is narrow on purpose. The FDA plans to take roughly ten manufacturers in each of four tracks: hypertension, diabetes, chronic musculoskeletal pain, and depression. In return for the early-access flexibility, you collect and hand back real-world performance data. That's the trade. TEMPO isn't a way to skip proving your product works. It's a structured way to prove it in the field rather than in a submission first, with a Medicare payment pathway waiting if the outcomes hold up.

Even if you never apply, TEMPO tells you which way the wind is blowing. The FDA's stated direction, in Makary's words, is a smarter and more forward-thinking framework for AI, which in practice means more enforcement discretion and more weight on real-world evidence over upfront review. Build for that world. Instrument your product to capture outcomes from the first release, because the regulatory and payment systems are both starting to reward the companies that can show what actually happened to patients.

What to actually do with this

If you're scoping a healthtech build in this environment, a few decisions are worth making before you write the feature, not after:

• Pin down which side of the device line you're on, in writing. Map your feature against the four Cures Act criteria and the still-regulated exceptions. If you analyze an image or signal, or predict a time-critical event, plan for clearance and don't talk yourself out of it.
• Treat your marketing claims as a regulatory surface. The wellness exemption lives entirely in how you describe the product. One sentence promising to detect or diagnose can move you back under FDA oversight, so your copywriter and whoever owns regulatory need to be in the same room.
• Build explainability in, not on. Showing a clinician the basis for every recommendation is both the law's fourth criterion and your best defense if an output is ever challenged. It's far cheaper as an architecture decision than a retrofit.
• Instrument for real-world evidence from the start. Whether or not you ever join TEMPO, the FDA and CMS are both shifting toward rewarding outcome data. A product that can't show what happened to patients can't take advantage of either.

The reflex read on deregulation is that it makes building easier. Sometimes it genuinely does. But pulling back the FDA's review mostly moves the work rather than deleting it: from a submission you hand to a regulator to a body of evidence you keep for yourself, and from the agency's judgment to your own. For teams that were going to do the safety work anyway, that's a real acceleration. For teams hoping looser rules let them skip it, the bill didn't disappear. It just moved to a later, more expensive date.