Skip to content
Entourage
Article19 min read

AI and software, October–November 2026: PCCP examples, generative-AI signals and function-based regulation from FDA, TGA, PMDA and the UK

Between early October and the start of November 2026, FDA, TGA, PMDA and the UK government published material on AI-enabled and software-based medical products. Most of it clarifies existing approaches or announces policy work; the one statutory change is Australia's amended CDSS exemption, in effect since November 1, 2026.

Diana Hohage

Diana Hohage

Principal Consultant

In brief

The regulatory status of AI software depends on the function it performs and the claims made for it, not on the label “AI.” FDA's PCCP examples are illustrations, not templates, TGA applies a function test to AI scribes, PMDA's principles cover AI in regulatory work, and the UK commitments still depend on consultation, guidance and secondary legislation.

Status: November 12, 2026

Between early October and the start of November 2026, four regulators published material on AI-enabled and software-based medical products. Most of it clarifies existing regulatory approaches or announces future policy work rather than introducing new premarket requirements. FDA published a Predetermined Change Control Plan (PCCP) web page with a list of authorized devices and two hypothetical case examples for AI-enabled software. Australia’s TGA clarified when AI scribes are medical devices and said it will take compliance action. Japan’s PMDA issued basic principles for the use of AI in regulatory activities. The UK government accepted all 44 recommendations of its National Commission into the Regulation of AI in Healthcare. The one amendment of statutory wording, to Australia’s existing CDSS exemption, took effect on November 1, 2026; TGA describes it as providing greater clarity about when the exemption applies. [1] [2] [8] [12] [16] [27]

Read together, these documents point in the same direction. The regulatory status of AI software depends on the function it performs and the claims made for it, not on the label “AI,” and change management, monitoring and transparency are moving from the end of the development process to its center. This article describes each development, separates what is binding from what is a signal, and ends with planning implications. We also incorporate FDA’s statements on foundation-model and large-language-model (LLM) functionality, because they belong to the same lifecycle theme. [3] [4]

Dates and status

  • October 20, 2026: FDA updated its PCCP page with a list of authorized devices with PCCPs and published two hypothetical case examples. They are non-binding illustrations. [1] [2]
  • October 19, 2026: the comment period on FDA’s generative-AI discussion paper closed. [4]
  • November 1, 2026: the amended wording of Schedule 4, Part 2, item 2.15 of the Therapeutic Goods (Medical Devices) Regulations 2002 (the clinical decision support system, CDSS, exemption) took effect through instrument F2026L01167. TGA describes the amendment as providing greater clarity about when the exemption applies. [25] [26] [27]
  • November 2, 2026: PMDA posted its Basic Principles for the Use of AI in Regulatory Activities as an Early Consideration. [12] [13]
  • November 2026: the first wave of innovators for phase 3 of the MHRA’s AI Airlock is to be selected. [17]
  • December 2026: MHRA intends to issue draft guidance on predetermined change control plans. [16]
  • Spring 2027: MHRA intends to launch a consultation on qualification and classification of software and AI-enabled devices, and the UK implementation roadmap is due. [16]

Key takeaways

  • FDA’s examples are illustrations, not templates. FDA states that the examples do not include or address all information expected in a PCCP. [2]
  • A PCCP has limits that FDA derives from section 515C of the FD&C Act. It can cover modifications that would otherwise require a new marketing submission, keep the device within its intended use and, for 510(k) devices, preserve substantial equivalence. [1] [2]
  • FDA describes a PCCP for automatic global retraining without human review. The example surrounds it with automated acceptance gates, blocked deployment, staggered roll-out, version display and an audit trail, and it names the absence of a human in the loop as a risk. [2]
  • TGA applies a function test. Pure transcription and formatting are unlikely to be devices; selecting, prioritizing, interpreting or acting on clinical information can make a scribe a device, and a disclaimer does not change that. [8] [10]
  • PMDA addresses AI used in regulatory work, not AI-based devices. The document sets out principles of human responsibility, governance and lifecycle management and creates no new requirement. [12]
  • The UK moved from advice to a government program. The concrete obligations still depend on a consultation, guidance and secondary legislation. [16]

FDA: PCCP page, authorized-device list and two hypothetical cases

On October 20, 2026, FDA’s Center for Devices and Radiological Health updated its page on Predetermined Change Control Plans and published Novel Medical Devices with Predetermined Change Control Plans: Hypothetical Case Examples. A PCCP describes planned modifications, how the manufacturer will develop, test and implement them, and how it will assess their impact. If FDA authorizes a PCCP as part of a marketing submission, the manufacturer may make the planned changes without a new marketing submission for each one, as long as they are included in and implemented in accordance with the authorized plan. PCCPs can be part of 510(k), De Novo or premarket approval submissions. [1]

The page sets three conditions for what a PCCP can contain: the modifications would otherwise require a new marketing submission (21 CFR 807.81(a)(3) or 21 CFR 814.39(a)), they maintain the device within its intended use, and, for 510(k) devices, they allow the device to remain substantially equivalent to the predicate. A PCCP generally has three components: a Description of Modifications, a Modification Protocol with pre-defined acceptance criteria and update procedures, and an Impact Assessment that covers the individual and cumulative impact of each modification and how one modification affects another. [1]

The page also carries a list of authorized devices with PCCPs, running from the Apple irregular rhythm notification feature (decision of July 21, 2023) to decisions of September 30, 2026. The list is not limited to AI: it includes antimicrobial susceptibility test systems, cardiac pacing leads, neurostimulators and ablation catheters. Trade press and an independent analysis of the list’s Excel export count 240 entries. For AI-enabled software, FDA’s final guidance of August 2025 on PCCPs for AI-enabled device software functions is the relevant recommendation document. [1] [5] [33]

Case 1: hospital-localized sepsis alert system

The hypothetical device aids clinicians in identifying patients at risk of developing sepsis within 24 hours in acute care hospitals, using routinely collected electronic health record data, and it generates a risk score and alert. Its global model is trained on a multi-site dataset. The PCCP covers local adaptation to each hospital system through three modifications: adjusting the weighting of structured nursing documentation inputs (mental status assessments and skin color/temperature flags) within an allowable range, re-training the final output layer(s) with site-specific EHR data, and modifying the alert threshold within an allowable range. The base model architecture, feature set and risk score calculation stay the same, free-text fields are excluded, and local adaptations are implemented manually with review and approval by the manufacturer before deployment at each new site. The device with the global model remains available for sites that do not use a locally adapted model, and as a fallback. [2]

The method has six steps: a data sufficiency review against minimum sample size and data quality, training and locking of the adapted model, validation on an independent test dataset that was sequestered before training and measured against the authorized device, updated site-specific labeling and performance documentation including model version history, deployment through a controlled software update process under the quality management system, and site-level post-market surveillance of alert positive predictive value, sensitivity and clinical response. Data are sourced from the deploying hospital, de-identified before use, and their provenance is documented. FDA also defines the dependencies: input weighting affects fine-tuning and the threshold, fine-tuning affects the threshold, and the modifications are therefore implemented in the sequence 1, 2, 3, while the threshold alone can be changed individually. [2]

Case 2: “continuously learning” arrhythmia detection software

The second example goes further. The device assists cardiologists in detecting clinically significant arrhythmias, including atrial fibrillation, ventricular tachycardia and third-degree heart block, in inpatients on continuous telemetry. The hypothetical PCCP accommodates automated, periodic re-training of the deep learning model across all deployed sites as newly annotated ECG data accrue, together with adjustment of the classification threshold for one or more arrhythmia classes. The modifications are implemented automatically and globally by software at the end of each cycle, without human input, action, review or decision-making, after automated validation checks. [2]

The modification protocol contains ten steps: an automatic re-training trigger when a prespecified amount of newly annotated data has accrued, de-identified data preparation with a sequestered independent test set, re-training with the prespecified pipeline, re-evaluation of the operating point, automated validation per arrhythmia class and per predefined subgroup against acceptance criteria based on the authorized device’s performance, an automated deployment decision that blocks deployment and alerts the manufacturer if any criterion is not met, staggered deployment across sites, a user interface that displays the current device and model version and the date of the last update together with updated instructions for use, a complete audit trail of every cycle, and automated post-market monitoring with predefined drift thresholds that trigger a manufacturer review. Re-training must precede the threshold re-evaluation, because it changes the model’s output probabilities. FDA states openly that automatic implementation presents risks because no human is in the loop, and it relies on deployment control and post-deployment monitoring to address them. [2]

What the examples do and do not show

FDA describes the examples as illustrations consistent with statute, regulations and guidance. They do not include everything expected in a PCCP, although they may represent the level of detail in a public device summary. All modifications are premised on being appropriate for a PCCP under the conditions above. The examples therefore do not establish general acceptance of self-learning systems, and a PCCP is only as broad as the plan FDA authorizes. They illustrate the level of structure and the types of controls that may be relevant to a PCCP, including prespecified acceptance criteria by class and subgroup, management of interdependent modifications, controlled deployment, traceability and post-deployment monitoring. They should not be treated as a universal checklist or as a statement of FDA requirements for every PCCP. [2]

FDA: foundation models, LLMs and the public summary

FDA’s AI-Enabled Medical Device List page states that the list is not comprehensive and that devices were identified primarily by AI-related terms in the summary descriptions of their marketing authorization documents and/or by their classification, based on FDA’s Digital Health and Artificial Intelligence Glossary. The page also states that FDA will explore methods to identify and tag devices that incorporate foundation models, from LLMs to multimodal architectures, and that sponsors are encouraged to include appropriate information in their public summaries to facilitate this. A third-party cached copy of the page dated April 9, 2026, already contains the encouragement to include appropriate information in public summaries and the plan to tag foundation-model devices, so this does not appear to be a new announcement; the page creates no new content requirement for submissions. [3] [28]

On August 18, 2026, FDA’s Digital Health Center of Excellence issued a discussion paper on the regulation of generative-AI-enabled medical devices and opened docket FDA-2026-N-7874, with comments due October 19, 2026. The paper outlines a possible two-axis heuristic for risk assessment. One axis describes the activity performed by the function in four categories: informational and non-directive, informational and action-directing, action-taking under healthcare-professional supervision, and fully autonomous action-taking. The other axis reflects the severity of harm that could result from relying on an incorrect output. FDA presents this as a discussion framework and seeks stakeholder feedback; it is not an adopted classification scheme. The paper also describes a competency-assessment approach to premarket evaluation consisting of non-clinical benchmarking and clinical confirmation, possible approaches to postmarket monitoring, and considerations on foundation models and agentic systems. FDA seeks feedback on the feasibility of voluntary Foundation Model Device Master Files: under the concept discussed, model developers could submit information confidentially, and device sponsors could reference it in a premarket submission with the file holder’s authorization. According to section VII.A of the paper, submission of such a file would not constitute authorization of the underlying model for any device intended use, and sponsors would remain responsible for independently demonstrating the safety and effectiveness of their own device. The paper is not a draft or final guidance, and the comment period has closed. [4] [20] [21] [30] [32]

For sponsors, the practical question is what an “appropriate” public description is. Public documents differ by pathway. For a 510(k), the submitter writes either a 510(k) summary or a 510(k) statement; a summary must be detailed enough to explain the basis for substantial equivalence, and FDA makes it available within 30 days of the substantial-equivalence determination (21 CFR 807.95(d)). For De Novo requests, FDA voluntarily publishes a decision summary describing the scientific evidence supporting the grant decision. For premarket approval, FDA prepares the Summary of Safety and Effectiveness Data (SSED) for original PMA approvals and panel-track supplement decisions. For De Novo and PMA the public summary is an FDA document, and the sponsor contributes through the underlying submission and supporting record. In our view, a controlled internal description of each model-dependent function (model category, medical role, inputs and outputs, degree of independence and human oversight, versions, third-party dependencies, performance evidence and limitations, change control) is the sensible source for every public statement, so that nothing in a public summary exceeds the authorized intended use. [6] [7] [23] [24]

Australia: when an AI scribe becomes a medical device

On October 16, 2026, TGA published the outcome of its digital scribe review and updated guidance. TGA identified around 100 suppliers in 2025, surveyed them (40 responded) and reviewed 12 products, including those with the largest market share. None of the 12 suppliers had classified its product as a medical device. TGA found that eight included functions likely to meet the definition of a medical device under section 41BD of the Therapeutic Goods Act 1989, two needed more information and two appeared unlikely to be devices. The review did not test the products, assess safety or performance, evaluate detailed regulatory compliance, or assess privacy and data protection, and TGA has not named which products fall into which group. [8] [10] [11]

Under the updated guidance, a scribe whose sole purpose is to record, transcribe and convert a clinical conversation into structured text, including through templates, without analyzing or interpreting clinical information, is not a medical device. The guidance compares functions in a table: for example, presenting a complete patient record is not a device function, but identifying, selectively displaying, prioritizing or analyzing information to generate diagnostic or treatment suggestions is; assigning codes that reflect what was documented is not, but deducing diagnoses or codes from implicit information is; retrieving reference information is not, but applying it to an individual patient is; and chatbot or “insight” functions that generate new clinically relevant conclusions the practitioner did not identify are device functions. Worked examples include an optional insights feature accompanied by a disclaimer that it does not replace professional decisions, which is still a device; an administrative tool marketed with claims that it guides diagnosis and treatment planning, which is also a device. TGA states that statements or disclaimers are not sufficient in determining whether a product is a medical device. [9] [10] [31]

For suppliers of regulated functions, TGA describes four routes: conditional exemption as clinical decision support software, modification of the product and its claims, inclusion in the Australian Register of Therapeutic Goods before supply, or ceasing supply. According to a summary of the guidance, an exempt CDSS also requires notification to TGA and compliance with the Essential Principles, and exempt CDSS software remains subject to medical device regulation. Since November 1, 2026, the “kinds of medical devices” cell of item 2.15 has been replaced under instrument F2026L01167 (the conditions cell was not amended). As reported for the amended wording, the criteria are cumulative: the software’s sole purpose is a recommendation to a health professional; it does not process or analyze medical images or signals from another medical device; it is not intended to replace clinical judgment; it is not intended to make a clinical diagnosis or treatment decision, including where it is IVD software; and it displays the guidelines, calculations or logic used so that the health professional can interpret and verify the recommendation. We summarize the criteria here and recommend checking the exact wording in the instrument. TGA describes the amendment as clarifying when the exemption applies, and it says the review findings will guide targeted compliance activities where a scribe’s functionality meets the definition of a medical device. [8] [10] [11] [25] [26] [27] [31]

Japan: basic principles for AI in regulatory activities

On November 2, 2026, PMDA’s Center for Regulatory Science posted Basic Principles for the Use of AI in Regulatory Activities (Early Consideration), as a provisional English translation of the Japanese original, which prevails. The document applies to settings in which AI, including generative AI, is used in activities involving regulatory judgment and decision-making across the life cycle of pharmaceuticals, quasi-drugs, cosmetics, medical devices and regenerative medical products, from development including non-clinical and clinical studies to manufacturing and post-marketing. The development and operation of such AI may also fall within scope; basic research and general medical practice are not its primary focus. The audience is PMDA, applicants, marketing authorization holders, manufacturers, sponsors and contract research organizations, as well as developers and providers of AI systems used in such activities. [12]

It is not a rule for AI-based medical devices. PMDA states that the document is not intended to establish new regulatory requirements and that addressing only the matters it describes would not be sufficient to meet all obligations. Its three tiers are a human-centric principle (AI supports human judgment, and ultimate responsibility rests with humans), six core principles (risk-based approach, clear context of use and scope, data governance, fairness, accountability and transparency, multidisciplinary expertise, and adherence to laws and standards) and five implementation and management principles (design and development, evaluation, life cycle management, human oversight and AI literacy). On data, it asks for documentation of data characteristics, sources, processing steps, change histories and the basis for decisions, with traceability. On evaluation, it asks that assessment consider not only AI performance but also operation in the actual use environment. [12]

PMDA has a separate document for AI-based products: on October 14, 2026, it published an English translation of Review Points for Medical Devices Utilizing Generative AI (Draft), which targets products that qualify as medical devices and is intended to gather a wide range of opinions. The draft itself states that no generative-AI-enabled medical device has yet been reviewed for approval by PMDA, and RAPS reported the same on October 19, 2026; such a status can change. [14] [15] [29]

United Kingdom: from commission report to government program

On October 6, 2026, the UK government accepted all 44 recommendations of the National Commission into the Regulation of AI in Healthcare, whose report was published on September 10. The response states that the current medical device regulations are not fit for purpose for AI, calls for a proportionate, risk- and lifecycle-based approach, and sets out an action plan. Alongside it, MHRA opened applications for phase 3 of its AI Airlock sandbox, with the first wave of innovators to be selected in November 2026; trade coverage reports an application deadline of October 31 for that wave. [16] [17] [22]

The commitments are mostly stated as intentions with dates. MHRA intends to launch a consultation on qualification and classification of software and AI-enabled devices by spring 2027, followed by secondary legislation covering an updated definition of a medical device (including possible carve-outs for certain administrative, wellbeing and decision support software), a revised classification approach and an updated definition of intended purpose that considers design, functionality and context of use. It intends to issue draft guidance on PCCPs by December 2026, describing an approach in which manufacturers define boundaries for the scope of allowable change rather than only prespecified modifications, including for adaptive AI-enabled devices and while maintaining intended use. Further commitments include a toolkit for tailored post-market surveillance, a possible function-based approach, exploration of Master Files for general-purpose models, reporting of dependencies on such models in submissions, guidance on operational conditions for safe deployment and risk controls, unique device identifiers with version control, civil sanctions, a regularly updated list of authorized AI-enabled devices, staged authorization pathways and recognition and reliance arrangements with trusted regulators. Most of these carry the date “further details by spring 2027,” and a full implementation roadmap is due by then. [16]

On October 19, MHRA announced that its Deputy Director for AI and Software would present the agency’s approach at the AdvaMed conference on October 20, covering lifecycle-based regulation, the AI Airlock, regional partnerships with NHS London and Manchester University NHS Foundation Trust and closer collaboration with FDA. MHRA’s 2030 strategy, reported on October 14, adds a planned Centre for AI and Digital Health and wider reliance and recognition routes with trusted regulators, including FDA. These statements describe direction, not new obligations. [18] [19]

What this means for planning

Plan PCCPs as engineering and quality documents. FDA’s examples illustrate controls that may be relevant: defined modifications, ordered dependencies, acceptance criteria per class and subgroup, sequestered test data, controlled deployment with the ability to block deployment, version display, audit trails and defined monitoring. This is our practical reading of the examples, not a checklist. Whether a particular change fits a PCCP depends on the three conditions above and on what FDA authorizes. [1] [2]

Do not read the examples as acceptance of self-learning devices. The arrhythmia example is a hypothetical in which every retraining cycle runs through predefined gates, and FDA names the missing human in the loop as a risk. A sponsor considering automated updating should, in our view, treat it as an illustration of the kind of control concept to discuss with FDA, ideally in a Pre-Submission. [2] [4]

Qualify functions, not products. TGA’s function table, the UK proposals for a function-based approach and for clearer intended-purpose rules, and the FDA discussion paper’s per-function risk framework, which looks at the final deployed device and not at the foundation model alone, all ask the same question. For products with several functions, administrative and clinical functions should be assessed separately, and claims, marketing material, user interface prompts and actual behavior should be consistent. [9] [16] [32]

Treat third-party models as a change-control problem. FDA’s discussion paper, the UK response (Master Files, dependency reporting) and PMDA’s traceability principles converge on version notice, evaluation access and the ability to constrain a changed component. These are discussion or policy proposals today, and they are a sensible basis for supplier agreements now. [4] [12] [16]

Use PMDA’s principles for AI in your own regulatory work, as a benchmark. For AI-supported drafting or review of dossiers, safety reports and quality documents, the principles translate into an inventory of use cases, defined human review points, documented model versions and configuration, validation against the actual use case and supplier control. They do not require any SOP change by themselves. [12]

Image notice: The illustrative and header images were generated using AI and do not depict real people, places, or events. The portrait photographs show real people.

Relevant for your project?

Similar questions in your current project?

In a first call we clarify what is specifically relevant for your situation, without obligation.

Request a call →

Regulations & standards considered

  • Section 515C of the FD&C Act (predetermined change control plans), 21 CFR 807.81(a)(3) and 21 CFR 814.39(a)
  • FDA, Novel Medical Devices with Predetermined Change Control Plans: Hypothetical Case Examples (October 20, 2026)
  • FDA discussion paper on generative-AI-enabled medical devices, docket FDA-2026-N-7874 (comment period closed October 19, 2026)
  • Therapeutic Goods Act 1989, section 41BD, and Therapeutic Goods (Medical Devices) Regulations 2002, Schedule 4, Part 2, item 2.15 (CDSS exemption), amended by F2026L01167 with effect from November 1, 2026
  • PMDA, Basic Principles for the Use of AI in Regulatory Activities (Early Consideration), November 2, 2026
  • UK government, Government Response to the National Commission's Recommendations on the Regulation of AI in Healthcare (October 6, 2026)

FAQ

Frequently asked questions

No. FDA describes them as illustrations consistent with statute, regulations and guidance, and states that they do not include or address all information expected in a PCCP. They illustrate the level of structure and the types of controls that may be relevant to a PCCP. They should not be treated as a universal checklist or as a statement of FDA requirements for every PCCP.

Sources

AI and software

Which of your software functions will a regulator treat as a device?

Describe the product, its functions and the markets you plan to enter.

Your inquiry goes to our team, which brings in the subject-matter experts as needed.

Prefer direct? +39 02 8904 1000
contact@theentourage.it

  • Reply usually within one working day
  • 3 locations: DE · CH · IT · projects in the US too
  • 100% life sciences