The Algorithm Will Change. File Like You Know It.
Regulatory strategy for AI/ML-enabled devices — from locked models to adaptive algorithms, under FDA’s change-control framework and the EU’s twin MDR + AI Act regime.
A Model That Can’t Improve Is a Liability. A Change You Can’t Control Is a Recall.
Most cleared AI devices run locked models — the weights that finished review are the weights in the field. That satisfies the reviewer and starves the product: the data your device generates is exactly what would make it better. The traditional answer was a new submission per meaningful update, which turns a learning system into a filing cadence.
The Predetermined Change Control Plan is the way out: describe the modifications you intend, the protocol you’ll follow, and the impact you’ve assessed — and get the future cleared with the product. FDA finalized the framework for AI-enabled device software functions in 2025. We treat the PCCP as a product decision, not a submission appendix, because it defines how fast you can learn.
The question in every AI review: what exactly changes after clearance, and who agreed to it?
Anatomy of a Change Plan FDA Will Actually Authorize.
A PCCP is three documents wearing one cover page. Each has a job, and each has a way of failing review.
Description of Modifications
The bounded list of what you intend to change — no more, no less.
- Specific, verifiable changes — “retrain on site-local data” is a plan; “improve performance” is a wish
- Stays inside the cleared intended use, or it isn’t a PCCP change
- Scoped tightly enough that a reviewer can say yes
Modification Protocol
How each change is developed, validated, and released.
- Data management: provenance, curation, reference standards
- Re-training and re-tuning practice, with acceptance criteria set in advance
- Performance evaluation — including the subgroups in your intended-use population
- Update procedures: labeling, user communication, rollback
Impact Assessment
The argument that benefits and risks stay in balance across every planned change.
- Each modification traced to its risks and mitigations
- Interactions between changes considered together, not one at a time
- The case that the device stays safe and effective at the edge of the plan
The 2025 final guidance raised the bar in two places US teams keep missing: characterize performance across the demographic and clinical subgroups of your intended-use population, and keep labeling and public-facing documents current as modifications actually ship — including telling users what changed and when.
Provenance, consent, curation, reference standards — your training data has a paper trail, or it has a problem.
Your Training Data Is a Regulated Asset. Treat It Like One.
The GMLP principles FDA co-authored with Health Canada and the UK’s MHRA read like common sense until an auditor asks for evidence: representative data, independence of training and test sets, reference standards a clinician would defend, performance monitored in deployment. Each principle implies records — and most data science teams don’t keep them by default.
Subgroup performance is where reviews concentrate. A sensitivity number that holds across sites, scanners, skin tones, and disease severities is a different engineering problem from a headline AUC. We build the data governance and the subgroup evidence before the submission needs them — because retrofitting representativeness is somewhere between expensive and impossible.
In Europe, Your AI Device Now Answers Twice.
The AI Act layers onto MDR and IVDR rather than replacing them — AI-enabled medical devices are high-risk AI systems, with obligations arriving on their own clock.
AI Act in force
The regulation enters into force; prohibitions and AI-literacy duties phase in first.
Two conformity routes, one product
Data governance, transparency, human oversight, and logging obligations stack on top of your notified-body file.
High-risk obligations bite
For AI embedded in MDR/IVDR-regulated devices, full high-risk compliance lands — after the EU’s one-year deferral from 2027.
The planning window
Notified-body capacity and harmonized standards are the bottleneck. Programs that wait for the deadline will queue behind everyone who did.
What an AI Program Is Actually Committing To.
Three facts that should be in the board deck, not discovered in review.
FDA’s PCCP guidance for AI-enabled device software functions — the mechanism that lets a cleared model keep learning inside an authorized envelope.
When the EU AI Act’s high-risk obligations reach AI embedded in medical devices — layered on MDR, assessed alongside it.
The unit of performance evidence. A model that works “overall” and fails a demographic is, in review terms, a model that fails.
Six Failure Modes We Are Brought In to Prevent.
Every one of these is cheaper to fix in architecture than in a deficiency response.
The PCCP written after clearance
Shipping v1 locked and discovering the change plan needed the training pipeline designed differently — a new submission to fix a paperwork decision.
Subgroup gaps found by the reviewer
Performance characterized on the development population, questioned on the intended-use population — the additional-information letter that costs a year.
A “bug fix” that moved the weights
Model updates shipped through the software-maintenance door without a change-control decision anyone can defend later.
Training data with no chain of custody
Datasets assembled from wherever, consented for something else, deduplicated by memory — unusable the day provenance is questioned.
Transparency treated as marketing
No model card, no meaningful labeling about inputs and limits — then the final guidance’s user-communication expectations arrive as a surprise.
An EU strategy that stops at MDR
CE marking scheduled without AI Act data-governance and human-oversight evidence — two conformity assessments discovering each other late.
AI Regulatory Leadership That Has Sat With the Data Scientists.
Our AI leads have scoped change plans, defended subgroup evidence, and built data governance that survives an audit.
“The question isn’t whether your model will change — it’s whether the change lands inside a protocol the agency already agreed to.”
The discipline we bring to imaging AI, monitoring algorithms, and adaptive software.
Bringing an AI-Enabled Device to Market? Negotiate the Change Plan Up Front.
Bring senior AI regulatory leadership in while the training pipeline is still being designed — the PCCP you can file is the one you built for.
Senior-led. Embedded in your team. No junior hand-offs.