The promise is seductive: AI models that predict what a buyer will do before the buyer knows it themselves. Intent signals drawn from web behavior, content consumption, third-party data exchanges, and cross-platform interactions are being assembled into probabilistic profiles that marketers use to trigger campaigns, adjust scoring, and allocate budget. The MarTech industry has spent the last two years evangelizing this capability as the natural successor to demographic targeting.
But a question has gone largely unasked in the rush toward intent-driven everything: where did the data come from, and does the organization have a defensible legal basis for using it this way?
The answer, for most enterprise marketing teams, is uncomfortable.
1. Historical context
Customer intent modeling is not new. Google's introduction of "micro-moments" in 2015 formalized the idea that search queries, browsing patterns, and content engagement could reveal purchase readiness. What has changed is the scale and sophistication of the modeling, and the erosion of the consent frameworks that once governed the underlying data.
For most of the 2010s, intent data lived in two buckets. First-party data, collected through owned properties with some form of opt-in, and third-party data, aggregated by vendors like Bombora, TechTarget, and G2 from publisher networks. The implicit bargain was straightforward: users consumed free content, publishers monetized behavioral signals, and marketers purchased aggregated intent feeds. Privacy policies existed, but enforcement was lax and consumer awareness was minimal.
GDPR's enforcement beginning in May 2018 introduced the concept of "purpose limitation" into mainstream marketing operations. Data collected for one purpose could not be repurposed without additional consent. The California Consumer Privacy Act (CCPA), effective January 2020, added the right to opt out of the sale of personal information. Brazil's LGPD, Canada's updated PIPEDA guidance, and a growing list of U.S. state-level privacy laws have since created a patchwork of obligations that vary by jurisdiction, data type, and processing activity.
During this same period, the MarTech stack expanded dramatically. Scott Brinker's 2024 MarTech Landscape counted over 14,000 solutions. Many of these tools pass behavioral data between systems through integrations, APIs, and shared cookies with minimal documentation of the consent basis for each transfer. The result is an intent data supply chain where the provenance of any given signal is often untraceable.
Third-party cookie deprecation, which Google has repeatedly delayed but the market has largely anticipated, has pushed organizations toward first-party cookie strategies and proprietary data collection. Yet even first-party data, when processed through AI models that infer new attributes or predict future behavior, raises questions about whether the original consent covers the derived insights.
"Data is a toxic asset. Holding data is a liability. It should be collected only if there is a specific use for it, and deleted as soon as that purpose is complete."
2. Technical analysis
Modern AI intent models operate on a fundamentally different technical basis than the rule-based scoring systems they replace. Traditional lead scoring assigned static point values to discrete actions: downloading a whitepaper earned 10 points, visiting a pricing page earned 25. The data inputs were explicit, auditable, and directly tied to a user's observable actions on owned properties.
AI-driven intent models work differently. They ingest high-dimensional datasets, often combining first-party behavioral data with third-party intent feeds, firmographic enrichment, technographic signals, and in some cases social media activity. Machine learning algorithms then identify patterns that correlate with conversion outcomes. The output is a probability score or segment assignment, but the reasoning behind it is opaque. A prospect might be flagged as "high intent" based on a combination of 47 features, none of which individually would trigger a traditional scoring rule.
This opacity creates three distinct privacy problems.
Consent chain fragmentation
When an AI model combines data from six different sources to generate an intent score, the consent basis for the output depends on the consent basis for each input. If one source collected data under a "legitimate interest" basis (permitted under GDPR Article 6(1)(f)) and another collected it under explicit consent (Article 6(1)(a)), and a third is a purchased dataset with unclear provenance, the composite score inherits the weakest link in the chain. Most marketing automation platforms, whether Oracle Eloqua or Adobe Marketo, do not natively track consent basis at the individual data-attribute level. The consent record, if it exists, typically covers the contact record as a whole, not the 200 behavioral signals feeding the model.
Derived data as new personal data
The Article 29 Working Party (now the European Data Protection Board) clarified in 2017 that data derived from other personal data is itself personal data. An AI model that infers a prospect's budget authority, purchase timeline, or competitive evaluation status from browsing behavior has created new personal data. Under GDPR, the data subject has rights over this derived data, including the right to access, rectify, and erase it. Few enterprise marketing teams have mechanisms to surface AI-derived intent attributes in response to a data subject access request (DSAR).
Cross-border data flows in model training
Intent models trained on global datasets may process personal data from EU residents on infrastructure located in the United States or elsewhere. The Schrems II decision (July 2020) invalidated the EU-U.S. Privacy Shield, and while the EU-U.S. Data Privacy Framework adopted in July 2023 provides a replacement mechanism, its long-term stability is uncertain. Organizations training intent models on cross-border data need to ensure that their data processing agreements, standard contractual clauses, and transfer impact assessments cover not only the raw data but the model outputs.
As we discussed in our analysis of identity resolution under agentic AI, the gap between what AI systems can technically do with identity data and what they are legally permitted to do is widening.
3. Strategic implications
The strategic exposure here is asymmetric. The upside of AI intent modeling is incremental: better targeting, higher conversion rates, more efficient campaign spend. The downside is categorical: regulatory fines (up to 4% of global annual turnover under GDPR), class-action litigation, reputational damage, and the operational cost of unwinding a data architecture built on shaky consent foundations.
Three structural forces make this exposure difficult to ignore.
Enforcement is accelerating
The Irish Data Protection Commission fined Meta 1.2 billion euros in May 2023 for transferring EU user data to the U.S. without adequate safeguards. The Italian Garante temporarily banned ChatGPT in March 2023 over data collection concerns. The French CNIL has issued multiple enforcement actions against adtech companies for cookie consent violations. These are not theoretical risks. Regulators have demonstrated willingness to pursue large penalties against companies whose data processing outpaces their compliance infrastructure.
B2B is not exempt
A persistent misconception in enterprise marketing is that B2B data processing faces lighter regulatory scrutiny than B2C. This is false. GDPR applies to the processing of personal data of natural persons regardless of whether the context is consumer or business. A B2B prospect's work email address, IP address, and browsing behavior on a vendor's website are personal data. The UK Information Commissioner's Office explicitly confirmed this in its 2023 guidance on direct marketing. Intent data derived from a B2B prospect's behavior is subject to the same consent, purpose limitation, and transparency requirements as consumer data.
Vendor risk is multiplying
Every third-party intent data provider in the marketing stack is a data processor (or in some cases a joint controller) under GDPR. The organization using that data is responsible for ensuring the processor complies with applicable privacy law. When a marketing team purchases intent signals from a provider that aggregated them from a publisher network spanning 15 countries, the buying organization inherits the compliance risk for the entire chain. A privacy assessment of the vendor ecosystem is no longer optional due diligence. It is a legal requirement.
The CDP consolidation wave has compounded this problem by concentrating more data processing in fewer platforms, each of which acts as both controller and processor depending on the use case.
Source: GDPR Enforcement Tracker by CMS Law (enforcementtracker.com), as of Q4 2024
"We're moving from a world of 'What data can I collect?' to 'What data should I collect?' And that shift changes everything about how marketing technology gets built."
4. Practical application
Enterprise marketing operations teams can take concrete steps to close the gap between their intent modeling ambitions and their privacy compliance reality.
Conduct a consent provenance audit
Map every data source feeding your intent models back to its consent basis. For each source, document: (a) the legal basis for collection, (b) the purpose for which consent was obtained, (c) whether that purpose covers AI-driven profiling and automated decision-making, and (d) the geographic scope of the data subjects. This audit will almost certainly reveal gaps. Addressing them before a regulator does is preferable to addressing them after.
Organizations running data management programs across multiple marketing platforms should build this provenance mapping into their existing data governance workflows rather than treating it as a one-time compliance exercise.
Implement purpose-bound data layers
Rather than feeding all available data into a single intent model, segment your data by consent basis and purpose. Data collected with explicit consent for personalized marketing can be processed more aggressively than data collected under legitimate interest. Data purchased from third parties should be isolated and its use limited to the purposes specified in the vendor agreement. This architecture is more complex than a single unified data lake, but it creates defensible boundaries that survive regulatory scrutiny.
Build DSAR response capability for derived data
If your AI models generate intent scores, segment assignments, or predictive attributes, you need a mechanism to include those outputs in data subject access request responses. Under GDPR Article 15, data subjects have the right to know about "the existence of automated decision-making, including profiling" and to receive "meaningful information about the logic involved." Documenting your model's feature importance and decision logic, even at a summary level, prepares your team to respond to these requests without triggering a compliance breach.
Review subscription and preference management
Your subscription center should reflect the actual processing activities your marketing stack performs. If you are using intent signals to trigger personalized campaigns, the subscription preferences should give contacts visibility into and control over that processing. A subscription center that offers only "opt in to newsletters" and "opt out of all communications" is inadequate for an organization running AI-driven multi-touch campaigns based on behavioral intent signals.
Pressure-test your vendor agreements
Review the data processing agreements (DPAs) with every third-party intent data provider. Confirm that the DPA specifies the categories of personal data processed, the purposes of processing, the duration of processing, and the obligations of each party in the event of a data breach. If the provider cannot produce a DPA that satisfies these requirements, or if the DPA contains broad, vague language about "marketing purposes," that vendor represents unquantified risk.
5. Future scenarios
Over the next 18 to 24 months, three developments will reshape the relationship between intent modeling and data privacy.
Regulatory convergence on AI-specific rules
The EU AI Act, which entered into force in August 2024, classifies AI systems used for "evaluation of the creditworthiness of natural persons" and "risk assessment and pricing" as high-risk. While B2B intent modeling does not fall squarely into these categories today, the Act's risk-based framework creates a template that national regulators are likely to extend. The EDPB's forthcoming guidance on AI and data protection, expected in 2025, will clarify how GDPR principles like data minimization and purpose limitation apply to machine learning models. Organizations that wait for this guidance before addressing their compliance gaps will find themselves retrofitting systems under time pressure.
We explored the broader platform integration implications of advancing AI capabilities in our analysis of GPT-5.6 Terra. The compliance dimension of that integration challenge is only growing.
First-party intent will become the dominant signal
As third-party data becomes harder to source compliantly, organizations with strong first-party data collection, built on transparent consent and genuine value exchange, will have a structural advantage. Companies that have invested in form capture strategy, progressive profiling, and owned content experiences will be able to feed their intent models with data whose provenance is unambiguous. Those relying on purchased intent feeds will face declining data quality, rising costs, and increasing legal exposure.
This shift rewards organizations that treat buying behaviour analysis as a first-party data strategy rather than a vendor procurement exercise.
Privacy-preserving computation will mature
Techniques like federated learning, differential privacy, and secure multi-party computation offer a path toward training intent models on distributed datasets without centralizing personal data. Google's Privacy Sandbox initiative and Apple's on-device intelligence approach point in this direction. Within 24 months, expect at least one major marketing cloud vendor to offer privacy-preserving intent modeling as a product feature. Early adopters of these techniques will gain both a compliance advantage and a trust advantage with increasingly privacy-aware B2B buyers.
The AI personalization measurement problem will also intersect with these privacy constraints: if you cannot measure what a model is doing because the data is privacy-restricted, the business case for the model itself comes under pressure.
6. Takeaways
-
AI-driven intent modeling creates new categories of personal data (derived data, inferred attributes) that inherit the consent obligations of their source inputs. Most enterprise marketing teams lack the infrastructure to track consent at this level of granularity.
-
B2B data processing is subject to the same privacy regulations as B2C. The assumption that business contact data faces lighter scrutiny is incorrect and dangerous.
-
Every third-party intent data provider in the stack is a compliance dependency. Vendor data processing agreements should be audited for specificity, not assumed to be adequate.
-
Subscription and preference management must evolve beyond simple opt-in/opt-out binaries to reflect the actual profiling and automated decision-making that intent models perform.
-
The EU AI Act and forthcoming EDPB guidance on AI will introduce new obligations for organizations using machine learning in marketing. Waiting for final guidance before acting is a losing strategy.
-
First-party intent data, collected with transparent consent on owned properties, will become the most valuable and defensible signal for AI models. Organizations should invest now in the data collection infrastructure to support this shift.
-
Privacy-preserving computation techniques (federated learning, differential privacy) will enter the marketing technology stack within 24 months. Early evaluation of these approaches creates both compliance and competitive advantage.
