When AI Begins to “Infer”: The Boundaries of Personal Data Are Being Rewritten, How Should Enterprise Data Architecture Respond?

When AI Begins to “Infer”: The Boundaries of Personal Data Are Being Rewritten

Introduction

In July 2026, Financier Worldwide published a special feature, “Data privacy and protection,” inviting six data protection and compliance professionals from Greece, France, Canada, Belgium, the United Kingdom, and Italy to discuss a seemingly basic question that in fact determines the direction of enterprise architecture: After AI systems generate, infer, and aggregate information at scale, what counts as personal data?

The core conclusion emerging from the discussion is that the boundaries of personal data are expanding from “directly identifiable individual” to “inferable, linkable, and traceable back to an individual.” Algorithmic outputs, behavioral profiles, biometric characteristics, and even neuro-derived data are being brought—or are about to be brought—within the scope of regulation. For companies, this means data privacy is no longer a paper compliance document that can be filed away, but a set of operational constraints that must be written into cloud platforms, data pipelines, and the model lifecycle.

Event Background: A Cross-Border Discussion About “Definitions”

This discussion deserves the attention of enterprise technology managers because it reveals a structural fact: the criteria for determining what constitutes personal data are diverging across jurisdictions, while corporate data flows rarely remain within a single jurisdiction.

Greece: The Hellenic Data Protection Authority (HDPA) adopts an expansive, dynamic interpretive approach: algorithmic outputs and behavioral profiling mechanisms no longer enjoy anonymous status as long as they can enable indirect identification, targeting, or retrospective association with a natural person. Under the EU Artificial Intelligence Act framework, the regulatory focus is clearly concentrated on the strict prohibition of emotion recognition systems in workplace and educational settings.

France: In its 2025 recommendations, the French data protection authority (CNIL) adopted the position of the European Data Protection Board (EDPB) Opinion No. 28/2024: as long as training data can be extracted—including “regurgitation” in generative systems—the model falls within the scope of the GDPR.

Canada: The statutory definition has not formally changed and remains “information about an identifiable individual,” but regulators are clearly expanding its application in practice: from focusing only on data that can directly identify an individual to also covering inferred and derived information, including behavioral profiles, AI-generated outputs, and biometric and neural data.

Belgium does not redefine personal data through separate legislation; instead, it applies the GDPR and the EU Artificial Intelligence Act to emerging data types. The Belgian Data Protection Authority (BDPA) holds that AI-generated inferences—behavioral profiles, scores, predictions—constitute personal data as long as they relate to an identifiable individual, even when they come from indirect signals such as browsing records or tone of voice.The United Kingdom, by contrast, is slowly moving toward more relativized, contextualized standards: identifiability is assessed from the perspective of the specific entity processing the data, and the same information may be personal data for one processor but not for another. The Data Use and Access Act and Information Commissioner’s Office (ICO) guidance both reflect this direction, although legislators have not expanded the definition to the point of potentially jeopardizing the UK’s adequacy determination.

Italy’s data protection authority, the Garante, likewise favors a broad interpretation: as long as information can directly or indirectly contribute to identifying, profiling, or singling out an individual, it falls within the regulatory scope, and the regulatory focus is shifting from the technical form of information to the actual risk of re-identification or inference.

Technical Analysis: Three Pathways from “Identifiable” to “Inferable”

For non-technical managers, understanding this boundary change can be approached through three technical pathways.

First, model extractability. Recital 26 of the GDPR provides a functional definition: any element that can be used to single out an individual through “means reasonably likely to be used.” EDPB Opinion 28/2024 and CNIL’s 2025 recommendations extend this standard into generative models themselves—if training data can be reverse-extracted, the model itself is subject to the GDPR. Technically, this means that “we only store weights, not the original data” is no longer an automatically valid defense.

Second, the legal effect of automated inference. In the 2023 Schufa case, the Court of Justice of the European Union (CJEU) ruled that when a third party “heavily relies” on an automated probability score to determine a contractual outcome, that score constitutes a decision within the meaning of Article 22 of the GDPR. The Belgian BDPA further clarified that an AI-generated score has no probative value in itself, and that any legally significant decision must be preceded by meaningful human review. The Greek HDPA, in turn, pointed out that pseudonymization alone is insufficient to eliminate re-identification risk; therefore, data produced by algorithmic predictions and neuro-derived insights constitutes personal data and requires a mandatory Data Protection Impact Assessment (DPIA) under the GDPR.

Third, the special status of biometric and neural data. Biometric data constitutes a special category of data under Article 4(14) and Article 9(1) of the GDPR; since 2019, the French CNIL has extended this protection to algorithmic outputs derived from original biometrics but from which the original data cannot be reconstructed. In January 2025, the Belgian Constitutional Court confirmed that facial recognition requires strict proportionality, explicit consent, and immediate deletion after use. Neuro-derived data is the next frontier: although it has not yet been included in Article 9 of the GDPR, the Council of Europe’s T-PD Committee issued draft guidelines in 2025 applying the principles of the upgraded Convention 108 to neural data.To summarize in one sentence: whether data is “anonymous” or “personal” no longer depends on what it looks like, but on whether it can be linked back and used to make decisions.

Business Impact Analysis

Cost Impact: A Two-Way Rise in CAPEX and OPEX

The technification of compliance requirements will simultaneously raise capital expenditure and operating expenditure. Capital expenditure is reflected in data partitioning, regionalized training and inference environments, audit log infrastructure, and the development of data lineage and deletion capabilities; operating expenditure is reflected in ongoing DPIAs, model extractability assessments, manual review roles, and cross-jurisdictional data flow mapping.

Enforcement data explains why these costs are hard to avoid. France’s CNIL conducted 323 inspections in 2025 and issued 83 sanctions totaling nearly €487 million, 67 of which were completed through the fast-track procedure introduced in 2022—meaning the sanction cycle is shortening. Belgium’s BDPA opened 157 new cases in 2024, up 83% year on year, with total fines of €708,371, including a €174,640 fine against a data broker and a €250,000 fine against IAB Europe.

Deployment and Operations Impact: Compliance Moves Upstream to the Architecture Layer

Once “inferences are personal data” becomes a regulatory consensus, compliance cannot be remedied after the fact at the application layer. Enterprises need to answer several questions at the architecture level: in which region training data lands, whether inference results flow across borders, whether data subject rights requests can actually be executed technically, and whether model outputs can be traced back to specific data sources.

UK practice also points to another complexity: identifiability may vary by processor, meaning the same data may have different legal attributes in different business units and different cloud environments; data catalogs and permission models need to have the corresponding granularity.

Security and Compliance Impact: Breach Notification Becomes an Enforcement Entry Point

Data from multiple jurisdictions point to the same trend—breach notification is no longer merely a compliance obligation, but an enforcement entry point. France recorded 6,167 breach notifications in 2025, up about 10% year on year, while insufficient Article 32 security measures remained one of the main enforcement grounds. Belgium had 1,455 breach notifications in 2024, up 13% year on year, with ransomware attacks under active investigation; the BDPA also routinely sends follow-up inquiries to controllers, requiring them to explain the methodology used to assess the level of breach risk.

Litigation risk is rising in parallel. In Canada, class actions now often follow major breaches, and companies need to manage regulatory and litigation risks simultaneously. In Belgium, 10 appeals were filed with the Brussels Market Court in 2024, of which 6 rulings were partially or fully overturned; more notably, the Market Court has confirmed that a single incident can trigger a comprehensive GDPR audit of the entire organization.

Market Competition Analysis

This change is redistributing bargaining power across the cloud computing and data industry chain.On the cloud provider side, data residency and regionalized processing capabilities have shifted from a “bonus” to an “entry requirement.” When model extractability, traceability of cross-border data flows, and enforceability of data subject rights become part of contract terms, cloud platforms that can offer stronger data partitioning, more granular auditing, and localized processing options will gain an advantage in bids from customers in regulated industries; conversely, the simplified narrative of a unified multinational architecture will encounter resistance.

The data brokerage and adtech sectors are under the most obvious pressure. Belgium’s penalties against data brokers and against IAB Europe show that business models relying on indirect signals to build profiles are in a zone of heightened regulatory scrutiny.

On the enterprise SaaS side, data processing terms for AI features that are enabled by default will face stricter scrutiny. When behavioral inference is deemed personal data, SaaS vendors need to provide explainable inference logic and data paths that can be turned off in product design; otherwise they will transfer compliance risk to customers.

On the AI infrastructure side, the extractability of training data will directly affect model release strategies. The usability of the claim “anonymized” in marketing and compliance documents is declining, and model cards, data provenance statements, and extractability tests may become standard attachments for model delivery.

Industry Trends Observations

First, from paper compliance to operational readiness. A recurring judgment in the feature is that what enterprises need is a pragmatic approach based on operational readiness, not reliance on documents. Actions such as DPIAs, human review, and extractability testing are essentially turning compliance into engineering practice.

Second, the relativization and contextualization of identifiability standards. The relative standard being advanced in the UK, as well as the subjective standard reintroduced by the CJEU in the EDPS v SRB case, both suggest a possibility: anonymity can be conditional and contextual, rather than an inherent property of the data itself.

Third, sovereignization and regionalization will continue to permeate the infrastructure layer. When data attributes depend on the processing entity and processing environment, the design freedom of cloud architecture will be redrawn by compliance boundaries.

Fourth, enforcement is moving from sporadic actions to structured programs. CNIL’s shift from scattered cases to systematic inspection programs, and Belgium’s shift from passively handling complaints to proactive systematic reviews, indicate that regulatory capacity itself is becoming industrialized.

CloudTechDaily Insight

The most important significance of this discussion is not whether a particular jurisdiction has redefined personal data, but that the underlying assumptions of regulatory logic have shifted: whether a piece of data is protected is no longer determined by its form, but by whether it can be linked, inferred, and used to make decisions. For AI systems, this is a standard that is difficult to “circumvent” through technical means, because it assesses system capabilities, not the way data is stored.There are three direct implications for enterprise IT strategy. First, data governance must shift down from the application layer to the platform layer. Data partitioning, lineage tracking, regionalized inference, and deletion capabilities should be designed as first-class citizens of cloud architecture, rather than patched in hastily before a compliance audit. Second, the launch process for AI features needs to embed compliance decision points: whether the sources of training data can be extracted, whether inference results will be used for decisions with legal significance, and whether meaningful human review is retained. The answers to these questions should be provided during the architecture review stage, not after an incident. Third, cost models need to be reassessed. The scale of penalties and the fast-track mechanism in France, and the precedent in Belgium where a single incident triggered a comprehensive audit, both mean that the expected loss from compliance failure is rising. Making compliance capabilities reusable platform capabilities is one of the few paths that can simultaneously reduce risk and long-term operating costs.

For the cloud computing industry, the most noteworthy long-term implication is this: privacy compliance is changing from a topic for legal departments into a component of infrastructure product strength. Whoever can make regionalized processing, verifiable data governance, and AI governance capabilities into default-available platform capabilities will be more likely, in the coming enterprise AI procurement cycle, to turn compliance from a cost center into a competitive barrier.