FDA Has Already Told Clinical LLM Companies What It Expects: A PCCP and the Monitoring to Back It
By Sentir Health · 07/07/2026
If you are building a clinical product on a large language model, an ambient scribe, a patient-facing agent, a triage tool, a digital therapeutic, you could be forgiven for thinking FDA hasn't gotten to you yet. There is no final generative AI guidance. There is no cleared LLM chatbot to point to. The rules feel like they belong to a future version of the industry.
Then on November 6, 2025, FDA convened its Digital Health Advisory Committee for a full day on generative AI-enabled digital mental health devices. The committee worked through a hypothetical: a prescription generative AI chatbot for adults with major depressive disorder. FDA had authorized nothing like it. The point of the exercise was to work out what it would take.
We read the meeting materials the way we read clearance summaries for the PCCP Tracker: looking for the mechanism. It was named early and often.
The mechanism is a PCCP
Generative AI systems update, adapt, and shift in ways their makers don't fully control. A marketing authorization freezes a device at a moment in time. The tool FDA has for squaring those two facts is the Predetermined Change Control Plan, the same mechanism we track across 143 authorized devices. A PCCP lets a manufacturer pre-authorize defined categories of change so the model can keep moving without a new submission each time.
Look at what FDA said a generative AI PCCP would need to contain: defined ranges within which model behavior may shift, user notifications when it does, and postmarket monitoring of performance stability across institutions and populations, with drift and bias detection and rollback criteria.
Read that list again. Drift detection, rollback criteria, performance stability across sites and populations. A regulator just described a running system, not a document.
The committee went further: monitoring after authorization should look like the study that got you authorized, following trends over time and watching for the product's function shifting away from its intended use. That study doesn't end. It becomes the baseline for one that never stops.
The last generation mostly skipped this part
As of our June 17 snapshot, the FDA has authorized 143 devices with PCCPs, and 64 of them are AI/SaMD. Only 8 of those 64 state an explicit real-world performance-monitoring commitment in their cleared documentation. We wrote about that gap in our read of the 2026 landscape.
That 8-of-64 figure describes the predictive generation: classifiers and detectors with fixed inputs and outputs, most of them in radiology. For those devices, FDA accepted change plans that said little about monitoring, and most companies committed to none.
Nothing about the November meeting suggests that pattern carries forward. For a generative model, monitoring did not come up as an appendix to the change plan. It was the part FDA and its advisors kept returning to, in operational detail. The direction was already visible in January 2025, when FDA's draft lifecycle guidance for AI-enabled devices, still in draft as of this writing, laid out what a postmarket performance monitoring plan should include.
If you intend to hold a PCCP for a generative product, the monitoring section will not be the shortest section. It will be the reason the rest of the plan is credible.
Not planning to be a device? The questions follow you anyway
Most clinical LLM products today are not on a device pathway, by design. That's a reasonable call, often the right one. The questions don't go away, though. They just come from someone else first.
Hospitals ask first. An LLM product selling into a health system now passes through an AI governance review, and those committees want to know how you monitor the model in production, what your baseline is, and what happens when performance moves. And more and more, they want evidence, not intentions.
States have started asking too. In August 2025, Illinois enacted the WOPR Act, which bars AI systems from making independent therapeutic decisions or engaging in therapeutic communication without review by a licensed professional, with civil penalties of up to $10,000 per violation. Utah and Nevada set their own limits earlier in the year. The line between a wellness product and a regulated one is now being drawn by more than one authority, and the companies that can show what their model actually does in the real world are the ones in a position to argue about which side of the line they sit on.
What to do now
Write your monitoring plan as if it were a submission appendix, whether or not a submission is planned. FDA has already published the outline: performance metrics against a defined baseline, drift and bias detection with real thresholds, stratification across sites and populations, and rollback criteria for changes that fail their acceptance bar.
Then start keeping the evidence, because the baseline is the one piece you cannot produce retroactively. We made that argument for predictive devices in The PCCP Inflection Point, and it is doubly true for models that change under you. By the time anyone asks how you know the model is still performing, the record either exists or it doesn't.
Where this is heading
The November meeting was about mental health devices, but the framework discussed in that room, a change plan with defined boundaries and continuous evidence that the product stays inside them, is the likely shape of generative AI oversight in general. The Sentir PCCP Tracker holds the mechanism's whole history, from the first AI pioneers in 2020 to the current wave, browsable by category, clinical panel, PCCP type, or company. If FDA follows the path its advisors mapped, the next wave of clinical AI will be the first that has to bring the monitoring system, not just the plan.
Sentir Health is the independent system of record for the performance of clinical AI. We produce the audit-defensible evidence that regulators, hospitals, and your board ask for. Learn more or book a call.
Methodology note: PCCP counts come from FDA 510(k) and De Novo records via the Sentir PCCP Tracker, current as of 2026-06-17. Accounts of the November 6, 2025 Digital Health Advisory Committee meeting are based on FDA's public meeting materials and post-meeting summary. Monitoring-commitment labels are editorial classifications based on public documents, not official FDA designations.