Europe & the U.S.: FDA Reviews GenAI Medical Devices, FTC Targets Personalized Pricing, UK Tightens Child Chatbot & Agent Security

On August 18, 2026, the U.S. Food and Drug Administration (FDA) published Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback, seeking stakeholder input on risk assessment, premarket evaluation, postmarket monitoring, and the use of foundation models and agentic AI systems in generative AI-enabled medical devices. The document is a discussion paper and request for feedback, not formal guidance, regulation, or a new regulatory requirement. Comments are due by October 19, 2026.

The FDA discussion focuses on how a new regulatory-evaluation framework might be developed for generative AI-enabled medical devices rather than directly amending existing medical-device approval rules. The paper identifies two main directions. First, it explores a risk-dimension-based assessment framework that would determine regulatory focus according to the type of device task and the potential severity of harm. Second, it explores competency-based evaluation for premarket assessment, including benchmarking for non-clinical devices and clinical confirmation, to assess whether a generative AI-enabled medical device performs as intended. The paper also discusses risk-proportionate postmarket monitoring, including periodic device-performance assessment, review of real-world samples, and monitoring for performance degradation. In addition, the FDA specifically raises regulatory questions concerning foundation models and agentic AI systems in medical devices, including model changes, output uncertainty, and risk management in continuously operating environments.

For enterprises developing or deploying generative AI medical products, the main impact falls across three areas: model validation, deployment operations, and continuous monitoring. Traditional medical software can often rely on fixed-version testing to demonstrate safety and effectiveness, whereas generative AI systems may produce different results as models are updated, inputs change, and outputs remain open-ended. Enterprises therefore need to establish validation systems capable of explaining “why this model is appropriate for this medical use case.” In practice, records of training-data sources, fine-tuning processes, version changes, performance metrics, error types, and intended-use boundaries may become important for future submissions or internal quality management. Medical-device companies integrating third-party foundation models may need more than evidence of final-product performance; they may also need to define what capability evidence, version-control mechanisms, and change-notification processes the foundation-model supplier provides.

The postmarket phase has an even more direct effect on operational systems. The FDA discussion paper suggests that generative AI-enabled medical devices may need more continuous real-world monitoring to address risks that premarket testing cannot fully cover. Enterprises therefore need to consider how to collect device logs, user feedback, clinical outcomes, and data on changes in model performance. In deployment, this means defining triggers for performance anomalies, such as deterioration in output quality, changes in the intended population, shifts in input-data distribution, or outcome drift following model updates, together with human review, version rollback, and risk-escalation procedures. Healthcare organizations procuring such systems may likewise need to reassess whether suppliers can provide ongoing monitoring interfaces rather than focusing only on initial certification materials.

The discussion paper extends the FDA’s broader development of a lifecycle regulatory approach to AI-enabled medical devices. The FDA’s Digital Health Center of Excellence (DHCoE) has previously conducted regulatory-science work on AI-enabled medical devices, including performance evaluation, safety of continuously learning models, and postmarket monitoring methods. Historically, FDA regulation of AI-enabled medical devices has focused more heavily on specific algorithms and intended device uses. The new discussion paper extends the regulatory question to open-ended outputs, dependence on foundation models, and agentic operating modes introduced by generative AI. Unlike the EU Artificial Intelligence Act (AI Act), which brings medical AI systems into a high-risk AI regulatory framework, the FDA is currently using a discussion paper to gather technical and industry feedback and gradually develop regulatory-science approaches. For enterprises, the issue to watch is therefore not a single approval-rule change, but how R&D, quality management, supplier management, and clinical operations will form a traceable closed loop as medical AI moves from “one-time validation” toward “continuous validation and operational monitoring.”

On August 19, 2026, the U.S. Federal Trade Commission (FTC) published the Federal Trade Commission’s Proposed Enforcement Policy Statement Regarding Personalized Pricing and sought public comment, setting out an enforcement approach to businesses that use consumers’ personal data and algorithmic estimates of willingness to pay to set personalized prices. The document is currently a proposed enforcement policy statement and request for public comment, not a final regulation or a new categorical prohibition. Comments are due by September 18, 2026.

The FTC does not propose a blanket ban on personalized pricing. Instead, it states that it will use its existing authority under Section 5 of the Federal Trade Commission Act (FTC Act), which prohibits unfair or deceptive acts or practices, to focus on cases where the use of personal data in price setting is not adequately disclosed. The proposed statement identifies three main considerations. First, where a business uses browsing history, location, purchase history, or other personal data to infer a consumer’s willingness to pay and adjust prices, it should tell consumers whether the price has been personalized. Second, disclosure should cover not only whether the price is personalized but also the types of data used for personalization and the basis for the personalization. Third, where consumers reasonably expect that the price of a product will not vary based on their personal data and the business fails to make adequate disclosure, the FTC states that the practice may constitute an unfair or deceptive act or practice.

For enterprises using AI for marketing, recommendations, dynamic pricing, or customer-value analysis, the change primarily affects three stages: data use, model decision-making, and consumer interaction. Enterprises have traditionally focused on whether algorithms improve conversion or optimize inventory, but where a model uses user profiles, historical behavior, device information, or location data to generate different prices, the enterprise needs to be able to explain what data was used in price formation and how the model output affected the final price. The record that needs to be retained is not only the final price, but the full chain from data input and feature selection to model judgment and price generation. For example, where an e-commerce platform uses a machine-learning model to predict purchase propensity and dynamically adjust discounts, it needs to distinguish between ordinary promotional recommendations, dynamic pricing based on market supply and demand, and personalized pricing based on individual characteristics, and avoid presenting the third mechanism as a uniform market price where consumers would not reasonably understand the distinction.

At the operational and compliance level, enterprises also need to revisit privacy notices, user-interface design, and vendor-management processes. If the pricing model is supplied by a third-party marketing platform, ad-tech provider, or business-intelligence tool, the enterprise needs to determine whether the supplier uses external data to infer consumer characteristics and whether those inferences feed into price decisions. When consumers complain, the enterprise also needs to be able to answer “why did this user see this price?” rather than merely providing the result produced by an automated system. Because the FTC document remains a proposed policy statement, enterprises do not need to immediately redesign all pricing mechanisms around it. However, businesses using AI for price optimization can begin maintaining pricing-decision audit records that include the data categories used, model version, human rules, and consumer disclosures, reducing the cost of explanation in future enforcement or disputes.

This policy development continues the FTC’s recent investigative track around “surveillance pricing.” In 2024, the FTC used its investigative authority to issue information requests to several algorithmic-pricing service providers, examining how businesses use location, demographic data, browsing history, shopping behavior, and similar information to categorize consumers and set targeted prices. Preliminary research published in 2025 further suggested that some pricing systems may use consumer behavior and characteristics to vary prices or promotions. The proposed policy statement issued on August 19 therefore does not introduce the idea of personal-data-driven pricing for the first time; it moves the issue from the investigation stage toward clearer enforcement criteria. Unlike the European Union, which uses the AI Act to establish risk classifications and transparency duties, the U.S. approach here relies primarily on existing consumer-protection law to address data-driven business practices. Its key question is not whether the algorithm belongs to a particular category of AI system, but whether the enterprise has adequately disclosed how consumer data affects a transaction outcome. For companies operating across borders, the same pricing-optimization system may therefore be subject to different regulatory logics: the European market may place greater emphasis on high-risk AI and lawful data processing, while the U.S. market may focus more heavily on whether consumers are misled and whether the price-formation mechanism is transparent.

On August 19, 2026, the UK Department for Science, Innovation and Technology (DSIT) updated the outcome of Growing up in the online world: a national conversation, setting out further protections for minors using AI chatbots, including regular breaks in chatbot use and measures addressing mental-health risks. The proposal remains at the stage of government policy design and subsequent regulatory development rather than a set of specific legal duties already in force.

The UK’s latest move builds on its child online-safety consultation by introducing additional governance measures for risks specific to AI chatbots. The government states that chatbot services used by people under 18 will need to introduce mandatory breaks to reduce risks associated with prolonged interaction and emotional dependency. It also plans to work with regulators to address chatbot services that provide harmful, misleading, or unverified mental-health advice, while retaining the option of restricting certain services where other regulatory and safety mechanisms are insufficient. The policy is aimed primarily at interactive chatbot services used by children and excludes chatbots typically used in business or customer-service settings.

For enterprises deploying consumer-facing generative AI chatbot products, the change primarily affects product design, user safety, and operational monitoring. Chatbot safety assessments have traditionally focused on filtering harmful outputs, but the UK proposal goes further by examining the interaction mechanism itself, including prolonged conversations, anthropomorphic behavior, and risks of emotional dependency. Enterprises offering companion, educational, or advisory chatbots to minors need to reassess whether product features encourage excessive use, such as continuous-conversation incentives, proactive re-engagement, or simulated intimate relationships. Protecting minors is therefore no longer only a content-moderation issue; it also involves conversation design, usage-time controls, and risk-intervention mechanisms.

At the model and operations level, enterprises also need to consider how mental-health-related requests are handled. Where users express anxiety, self-harm tendencies, or psychological distress, relying solely on a general-purpose model to generate reassuring responses may not satisfy future regulatory expectations. Enterprises can begin establishing sensitive-topic detection, human-escalation channels, referrals to professional resources, and model-output testing, while keeping records of changes to relevant safety policies. Products operating across multiple markets also need to distinguish UK child-protection requirements from frameworks elsewhere, such as the EU AI Act’s high-risk and transparency obligations or the U.S. reliance on consumer-protection and sectoral rules for chatbot risks. Because the UK measures remain under policy development, enterprises should treat them as a design benchmark for products used by minors rather than as an already effective compliance checklist.

The proposal continues the UK’s broader shift from conventional online-content regulation toward the governance of interactive AI risks. The Online Safety Act 2023 primarily established platform responsibilities around illegal content and content harmful to children, but the government’s 2026 consultation on children online highlighted that some risks from AI chatbots do not arise solely from specific content. They also arise from product design, including simulated human relationships, encouragement of prolonged engagement, and children’s reliance on chatbots for emotional support. The addition of mandatory breaks and mental-health safeguards therefore extends the regulatory focus from “what did the chatbot output?” toward “how does the chatbot interact with a child?” This differs from recent U.S. scrutiny of personalized algorithms through consumer-protection law and the EU’s risk-classification approach to AI systems. The UK is currently focusing more directly on the relationship between a specific user group—children—and interaction-design risks, rather than beginning with a uniform classification of model types. For enterprises deploying AI products used by children across regions, this means age assurance, interaction limits, content safety, data protection, and human escalation all need to be handled at the product level rather than relying on a single model-safety policy across every market.

On August 20, 2026, the UK National Cyber Security Centre (NCSC) published practical cybersecurity advice for operators of agentic AI systems, recommending controls including sandboxing, distinct identities, activity monitoring, and emergency shutdown capabilities.

The NCSC publication is not legally binding regulation, but interim practical advice for enterprises and system operators on reducing cybersecurity risk when deploying AI agents with significant levels of autonomous execution. The advice focuses on permissions, operating environments, and behavioral controls throughout agent execution. Key measures include: running AI agents inside a robust sandboxed environment to limit the data, networks, and system resources they can access; giving agents an identity that is clearly distinct and attributable rather than simply inheriting a human user’s credentials, combined with least-privilege access; establishing logging, auditing, and real-time monitoring; and retaining an emergency shutdown capability so an agent can be stopped immediately when abnormal or unintended activity occurs.

The NCSC also advises organizations to assess before deployment how much autonomy an agent actually needs, identify scenarios in which unintended behavior could occur, and use human oversight, access controls, and technical isolation together to constrain the potential impact.

For enterprises already deploying AI agents, the most important implication is not model capability itself, but the operating environment once the model receives execution authority. Traditional generative AI applications usually follow a “user prompt—model output” interaction pattern. Agentic AI can invoke tools, access systems, and execute tasks, requiring enterprises to reassess the full control chain between model access and business-system actions.

At the deployment layer, enterprises need to redesign agent permission boundaries. For example, if an agent that automatically handles customer support tickets also has database write access, email-sending permissions, and external API access, its potential impact is already close to that of an automated business account. The NCSC’s emphasis on attributable identities and least privilege means enterprises should not simply reuse employee credentials when connecting agents. Instead, they need service-account-like mechanisms covering access approval, credential lifecycle management, and restrictions on the scope of access.

At the operations and security-management layer, enterprises need additional mechanisms to monitor agent behavior. Traditional software exceptions can often be traced through system logs or error reports, whereas an agent may execute multiple actions in sequence. Enterprises therefore need to record which tools were invoked, what data was accessed, and what external actions were performed. For example, when a sales agent automatically modifies customer records or a procurement agent invokes a supplier system, the enterprise should be able to reconstruct the precise action chain and determine whether abnormal behavior originated from model judgment, permission configuration, or external input.

At the model-governance and vendor-management layer, enterprises selecting third-party agent platforms or building internal agents need to add new dimensions to security assessment. Beyond model quality, cost, and functionality, they should confirm whether the supplier supports runtime isolation, permission control, log export, human intervention, and fail-stop mechanisms. For agents connected to core enterprise systems, the absence of these controls can make it difficult to locate accountability quickly when something goes wrong.

The broader context is that NCSC’s focus on agent security extends its existing lifecycle approach of secure design, secure development, secure deployment, and secure operation. In 2023, NCSC and partner agencies from the UK, U.S., and other countries published the Guidelines for Secure AI System Development, emphasizing that AI-system security must run across design, development, deployment, and operation rather than relying only on post-deployment risk response. The August advice on agentic AI moves the question further from “is the model secure?” toward “how many actions can the model take on behalf of the organization, how many resources can it access, and how can it be constrained?”

This development also illustrates differences in regional approaches to governing AI agents. The EU AI Act primarily establishes a regulatory framework around risk levels, transparency, and obligations for high-risk systems. The UK NCSC is currently taking a cybersecurity-practice approach, focusing more on how agentic AI, as software with execution capabilities, changes the attack surface. For enterprises, the two are complementary rather than contradictory: the former affects whether an AI system meets regulatory requirements, while the latter affects whether the deployed system is controllable, traceable, and recoverable in operation. As enterprises move from AI as an assistive tool toward agents that can operate business systems, permission management, runtime isolation, and incident response will become foundational controls that need to be considered across jurisdictions.