Planning AI Updates for the US Market: An FDA PCCP Guide for UK Surgical-Device Companies

If your surgical device uses a machine-learning model you plan to retrain after a US launch, FDA's Predetermined Change Control Plan (PCCP) lets you have those changes authorised in advance, as part of the original 510(k), De Novo or PMA. Changes made in line with the authorised plan then need no new marketing submission. FDA's final guidance limits a PCCP to modifications that would otherwise require one. Its draft AI lifecycle guidance (January 2025) adds recommendations on bias and transparency. This article relies solely on the FDA texts in our FDA surgical robotics and SaMD base, which does not cover the MHRA, UKCA or CE-marking regimes.

Why a PCCP belongs in your US launch plan

Without a PCCP, a change that could significantly affect safety or effectiveness generally triggers a new 510(k) or, for PMA devices, a PMA supplement. Section 515C of the FD&C Act, added by FDORA (enacted 29 December 2022), removed that requirement for changes consistent with a PCCP that FDA has cleared or approved. FDA's guidance on PCCPs for AI-enabled device software functions is final: first issued on 4 December 2024, with the version in the base dated 18 August 2025.

For robotics firms the link is explicit. FDA's draft guidance on robotically-assisted surgical devices, issued on 25 September 2026 and not yet final, mentions built-in AI/ML features such as image segmentation, critical structure identification and instrument tracking. Where post-authorisation changes to such software are expected, it says you may propose a PCCP. It also recommends discussing AI/ML-enabled features with FDA through the Q-Submission Programme.

Scope: which changes qualify

FDA recommends including only a limited number of specific modifications that can be verified and validated, and intended to maintain or improve safety or effectiveness. The guidance groups suitable changes into three families:

• changes to quantitative performance specifications, for example retraining on new data from the intended use population with the same type and range of input signal;
• changes to inputs and compatibility, for example new makes or models of an acquisition system for the same signal type, limited new inputs, or updated compatible hardware, operating systems or cloud infrastructure;
• certain changes to use and performance, such as authorisation for a subpopulation within the original indication after retraining on a larger dataset.

Every modification must keep the device within its intended use and, for a 510(k) device, substantially equivalent to its predicate. FDA generally expects the indications for use to stay the same too. Plans may cover manual updates, automatic (“continuous learning”) updates or both. For automatic updates, FDA suggests defining guardrails and discussing them through a Q-Submission.

Building the three components

1. Description of Modifications

List each planned change with its rationale. State whether it is automatic or manual, whether it applies uniformly to all devices (“global”) or differs by site or patient (“local”), and how often updates are expected. Link each modification to a specific performance evaluation in the protocol.

2. Modification Protocol

For each modification, FDA's Appendix A describes four elements: data management practices, re-training practices, performance evaluation and update procedures, including how labelling will be updated and how users will be told. Traceability between the modifications and the protocol is expected.

3. Impact Assessment

Compare each modified version with the unmodified device and discuss benefits and risks, including risks of harm and unintended bias. Explain how verification and validation keep the device safe and effective, how modifications interact, and their cumulative impact, including on hardware and other software functions.

Data from UK sites: what the draft AI guidance implies

FDA's draft AI-enabled device software functions guidance (issued 7 January 2025; not final) recommends addressing transparency and bias across the product lifecycle. That includes “collecting evidence to evaluate whether a device benefits all relevant demographic groups (e.g., race, ethnicity, sex, and age) similarly”. For a company whose data comes mainly from UK hospitals, several recommendations deserve attention:

• address representativeness of data for development, testing and monitoring;
• explain performance across important subgroups; the draft lists patient characteristics, geographic sites and data-collection equipment;
• keep test data independent of training data, for example sampled from completely different clinical sites, and sequestered from developers;
• consider a performance monitoring plan to detect drift after deployment.

For clinical evidence on a robot as a whole, the RASD draft goes further. Data from investigations outside the US should be representative of the intended US patient population and of typical US operating-room use. Our inference is that a UK-only dataset will need a reasoned case for its relevance to US patients and sites. Confirm the expectations with FDA in a Pre-Submission.

Authorisation, labelling and life after clearance

1. File the PCCP as a standalone, versioned section titled “Predetermined Change Control Plan” in a submission that ends in an FDA decision: an original PMA or certain PMA supplements, a traditional or abbreviated 510(k), or an original De Novo request.
2. Remember that FDA does not authorise a PCCP in a Pre-Submission. Use the Pre-Sub for feedback only.
3. Include labelling stating that the device uses machine learning and has an authorised PCCP, and update it as modifications are implemented.
4. Implement only changes that are in the plan and carried out under the protocol, and document them in your quality system.
5. For anything else, apply the normal modification rules. Changing the PCCP itself generally needs a new marketing submission.

One housekeeping point: the PCCP guidance cites sections of the former Quality System regulation, and notes that the QMSR (incorporating ISO 13485:2016) takes effect on 2 February 2026. Under 21 CFR 820.1, the QMSR applies to finished devices imported or offered for import into the US, so your change-control procedures should reference it.

Common mistakes to avoid

• Writing an open-ended plan. FDA asks for a limited number of specific modifications, each linked to a performance evaluation. A broad promise to “improve the model” is not a plan.
• Using a PCCP for minor changes. Changes that would not need a new submission fall outside a PCCP; they are documented in the quality system instead.
• Assuming a later 510(k) can rely on the updated model. Where a predicate was authorised with a PCCP, a new device is compared with the version cleared before the PCCP changes, until a subsequent clearance covers the modified device.
• Forgetting the public summary. FDA recommends that the 510(k) summary, De Novo decision summary or PMA summary describe the PCCP: planned modifications, testing methods, validation and performance requirements, and how users will be informed.
• Treating a deviation as minor. A modification that departs from the protocol is no longer consistent with the PCCP and must be assessed under the normal modification rules.
• Overlooking cybersecurity. If the robot meets the FD&C Act definition of a cyber device under section 524B(c), its premarket submission must meet the section 524B(b) requirements. Updated cloud infrastructure in a PCCP will need to fit that analysis.

Sources: FDA final PCCP guidance and FDA draft AI-enabled device software functions guidance. For example, you can ask the base: “Does FDA's guidance on AI-enabled device software functions ask manufacturers to check for performance bias across demographic groups?” Nothing here is regulatory advice for a specific product.