The PCCP Inflection Point: What FDA Clearance Data Tells Us About AI/SaMD in 2025
By Sentir Health · 06/14/2026
Something changed in the FDA's clearance pipeline in 2023. If you work in AI-enabled medical devices, you've probably felt it without seeing the numbers clearly.
We built the Sentir PCCP Tracker to make those numbers visible. Here's what the data shows.
143 PCCP authorizations. 64 are AI/SaMD.
As of this snapshot in mid-June 2026, the FDA has authorized 143 devices with Predetermined Change Control Plans. Nearly half of them, 64, are AI/SaMD. In the preceding 90 days, 22 new PCCPs were added, more than a third of them AI/SaMD. (The live tracker recomputes these counts as the FDA posts new records, so current figures may differ.)
Twenty-two in ninety days is the one to pay attention to. The ramp people keep predicting already happened.
Why the timeline looks the way it does
If you look at PCCP authorizations by quarter, the chart is almost flat from 2019 through 2022, with a handful of blips in 2020 and 2021. Then 2023 happens and the bars start climbing.
The shape is no accident. It tracks the regulatory timeline.
The concept of a PCCP isn't new. FDA sketched the idea of pre-authorizing certain algorithm changes in its 2019 AI/ML discussion paper, and the statutory authority itself arrived in December 2022, when FDORA added Section 515C to the Food, Drug, and Cosmetic Act. And those early blips were not traditional devices at all. All of them were AI: Caption Guidance, the AI ultrasound guidance software, carried a change plan through its February 2020 De Novo and again through a follow-on 510(k) that September, and Medtronic's LINQ II cardiac monitor with its Zelda AI classification system followed in mid-2021. The pioneers were exactly the adaptive products the mechanism was invented for. What didn't exist yet was a template anyone else could follow.
That template arrived in April 2023, when FDA published its draft guidance on PCCPs for AI/ML-enabled device software functions. It was the first document that gave AI device makers a concrete picture of what a PCCP submission should look like, what it needed to include, and how FDA would evaluate it. Companies that had been watching the space started building PCCPs into their pipelines almost immediately after.
The final guidance dropped in December 2024, and it landed with more clarity than the draft. The title was broadened from ML-specific language to AI-enabled device software functions more broadly, reflecting that most of what FDA was actually reviewing was machine learning but that the framework needed to cover the full AI category. By that point the quarterly authorization numbers were already climbing, and they haven't slowed down since.
By 2025, the bars on the chart reached 28 authorizations in a single quarter across all device types, with AI/SaMD making up a growing share of that total.
Having a PCCP and being able to execute it are different things
Here's the part that doesn't get talked about enough.
A PCCP is not a section you write for your 510(k) submission and then forget about. Under the December 2024 final guidance, every authorized PCCP must include a Modification Protocol that covers four specific elements: data management practices, re-training practices, performance evaluation protocols, and update procedures. FDA expects pre-defined acceptance criteria for each planned modification type, and your quality system has to carry all of it, from data handling to the update itself.
That's a real operational commitment. And the moment you actually exercise your PCCP, it stops being theoretical.
When you retrain your model, or make any other change authorized under your PCCP, you need to demonstrate that the modification met your pre-specified performance criteria. FDA doesn't require you to monitor continuously or file interim reports along the way. But when you exercise the PCCP, the burden is on you to produce that demonstration on demand, and that raises an uncomfortable question: how do you show your model wasn't already degrading before the change? If performance drift is what prompted the retrain in the first place, your baseline matters enormously, and a baseline isn't something you can reconstruct after the fact. The evidence either exists by the time you need it, or it doesn't.
The guidance is clear that deviations from an authorized PCCP, including failure to meet pre-specified performance criteria, could significantly affect the safety or effectiveness of the device. It's also clear that all PCCP implementation must occur within the manufacturer's quality system, and with FDA's QMSR rule aligning 21 CFR Part 820 with ISO 13485 taking effect in February 2026, the documentation and traceability discipline around those records is under renewed scrutiny.
FDA never mandates continuous monitoring by name. What it requires is a Modification Protocol with pre-defined acceptance criteria and re-training practices, and without longitudinal performance data you have no systematic way to demonstrate you met those criteria before and after the change. You'd be reconstructing a history you don't have, at exactly the moment when documentation matters most.
What this means practically
Almost every company that cleared a device with a PCCP is doing some form of post-market monitoring already. They have scripts, dashboards, spreadsheets, a notebook someone runs each month. The problem is not the absence of monitoring. The problem is that ad hoc monitoring decays.
The engineer who built the monitoring pipeline leaves, and the institutional knowledge of how it works leaves with them. A quarter's worth of performance evidence never gets formally documented because everyone was heads-down shipping. The acceptance criteria in the running code drift out of sync with what the authorized PCCP actually specified. FDA guidance evolves, and the monitoring setup doesn't keep pace. Then multiply all of that across multiple cleared devices, multiple model versions, and multiple PCCPs, each with its own modification protocol and its own acceptance criteria.
That's the gap: not whether you're monitoring, but whether what you have today will still defend a modification submission two years from now, when the person who set it up is gone and FDA wants to see the evidence trail.
This is what we're working on at Sentir. We don't replace the monitoring your team already does. We make it durable and independent. Performance gets captured continuously against the acceptance criteria your PCCP actually specifies, in a record held apart from the team shipping the model. The audit trail is tamper-evident, it survives turnover, and it covers every device and version you ship.
Evidence your own team generates about its own model is self-reported, however careful the team is. A record maintained separately, one that can't be quietly backfilled or adjusted after the fact, is simply more defensible when an FDA reviewer asks how you know the device stayed within its authorized envelope.
About the PCCP Tracker
The Sentir PCCP Tracker is a free public tool that aggregates PCCP authorizations from FDA 510(k) and De Novo databases, with AI/SaMD classification based on FDA product codes and device description keyword analysis. It updates regularly and is built specifically for this regulatory mechanism. Browse it by category, clinical panel, PCCP type, or company.
If you work in AI-enabled medical devices as a regulatory affairs lead, QA director, or CTO, bookmark it. This market is moving faster than it looks from the outside.
Sentir keeps the performance record FDA-cleared AI vendors need on the day they exercise their PCCP. Learn more or book a call.
Methodology note: PCCP authorization counts sourced from FDA 510(k) Premarket Notification and De Novo databases. AI/SaMD classification based on FDA product codes associated with AI/ML-enabled software functions, supplemented by keyword analysis of device descriptions. All figures current as of 2026-06-14. Classification methodology available on request.