Static Categories, Adaptive Systems: The Real Fault Line in AI Device Review

The pattern beneath the pilots

FDA’s clearance of ThinkSono Guidance, an AI tool that steers clinicians through vascular ultrasound for DVT evaluation, is textbook 510(k). Bounded task, fixed intended use, decades-old pathway. That pathway still works and will keep working for products like it.

The draft MDUFA VI commitment letter and the EU’s MDCG 2025-9 Breakthrough Devices framework have already been read, correctly, as parallel regulatory responses to generative and agentic AI. That reading is accurate but incomplete. The sandbox and the accelerated pathway are not isolated administrative fixes for a new product category. They are the visible edge of a wider fracture running through several regulatory definitions at once, not just review structure.

The FDA’s wellness-versus-medical-device line already shows the same weakness. Identical hardware falls under different rules depending on how a company frames its marketing claim, because the classification depends on stated intended use rather than device behavior. Separately, routine health data is now outperforming trial data for AI model training, and FDA’s evidentiary framework, built around controlled trial data, has not caught up to that shift. Intended use, evidentiary basis, and hardware classification are three static categories built for fixed products. Agentic AI stresses all three simultaneously, not just the submission pathway.

Hype check: what “agentic” actually has to prove

“Agentic” has become a marketing label as much as a technical one. A recent governance framework analysis still treats AI risk largely through existing classification tiers, which exposes a real gap between vendor language and regulatory readiness. A tool that adapts its output based on new inputs after deployment is a different regulatory animal than one that merely automates a fixed decision tree with a conversational interface. Buyers should test the claim against actual post-deployment behavior, not the vendor’s chosen adjective. The label does not change the review tier. Demonstrated adaptive behavior does.

Early engagement cuts both ways

Sponsors weighing sandbox or Breakthrough pathway participation should treat it as a negotiated commitment, not a free option. Early entrants help shape the evidentiary bar, human oversight expectations, and post-market monitoring standards the rest of the market will inherit. That is real leverage. It also means absorbing the cost of uncertainty first, and it means whatever commitments a sponsor accepts to get through the pilot may become the template regulators reference for everyone who follows, whether or not that template fits a later product’s actual risk profile. Shaping the bar and being bound by it are the same transaction.

The decision in front of you

The near-term decision is not whether to chase the agentic label. It is whether a given product’s behavior actually breaks the static assumptions built into intended-use language, trial-based evidence, and hardware classification. If it does, sandbox and Breakthrough participation deserve board-level attention now, with full awareness that early terms tend to outlive their pilots.

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. Core thesis that static regulatory categories are stressed by adaptive AI is coherent and defensible, but the claim that ‘routine health data is outperforming trial data for AI training’ is presented
Source & Claim Verification	Qwen · local	cleared. All factual claims are traced to citations, but some sources are redundant and could be consolidated for clarity.
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately identifies regulatory gaps for adaptive AI systems but lacks explicit mapping to ISO 42001, EU AI Act, or MDR/IVDR requirements.
Technical Accuracy	Llama	cleared. The article accurately identifies key technical and regulatory challenges posed by adaptive AI systems in medical devices, but could be strengthened with more precise technical definitions and example
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively identifies and challenges vendor hype around ‘agentic AI’ by demanding proof of adaptive behavior over mere labeling, and by highlighting the regulatory implications of such b
Novelty & Non-Duplication	Grok	held. Synthesis of already-wired items (ThinkSono 510(k), MDUFA VI, MDCG 2025-9, wellness/device line, RWD-vs-trials) under a familiar static-categories-vs-adaptive-AI frame adds only modest original angle
Validation	DeepSeek	cleared. The briefing’s central claim that static categories are breaking under agentic AI is plausible but relies on speculative evidence about adaptive behavior, which is not factually refuted by the provide

Sources cited: 11. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.