What is Predetermined Change Control Plan (PCCP)?

Why It Matters

An approved model is not frozen. Retraining, a prompt change, a new tool, or a shift in the data all move behavior, and the traditional device pathway treats each substantive change as a new submission. That collides with how AI systems are actually maintained: the changes arrive continuously, and the paperwork cannot keep pace with them.

The PCCP is the FDA’s resolution. Describe the modifications you plan to make, and the method you will use to develop, validate, and assess them, and have that plan reviewed once as part of the marketing submission. Implementations inside the described envelope then do not need their own submission.

What the Plan Has to Describe

Three things, per the FDA’s guidance for AI-enabled device software functions: the planned modifications, the associated methodology to develop, validate, and implement them, and an assessment of the impact each modification could have.

The FDA, Health Canada, and the UK’s MHRA also published joint guiding principles for PCCPs, built on the ten Good Machine Learning Practice principles. The tenth of those is the load-bearing one here: deployed models are monitored for performance, and retraining risks are managed.

Where It Breaks

The plan is an envelope. A change that falls outside the described modifications still needs a submission, and teams that treat the PCCP as permission to ship anything discover the boundary at the worst moment.

The second break is the monitoring the plan assumes. A PCCP rests on the premise that the deployed model is watched and that retraining risk is managed. Without drift monitoring and a revalidation trigger, the conditions the plan was granted under are not being met, whatever the deployment looks like day to day.

The third is evidence produced retrospectively. If the validation record is assembled at the next submission rather than generated as the system runs, the team is reconstructing history instead of reporting it.

How Flytebit Handles It

The controls a PCCP depends on are the ones we build for any governed production system: versioned prompts and models, a regression gate that blocks a change when scores drop, drift monitoring with a defined revalidation cadence, and a decision record per run that holds the evidence as it happens. The industry application is on our Healthcare & Life Sciences page, and the evaluation method behind the gate is in Evaluating Agentic AI Systems.