Data Privacy in AI Marketing: Consent, GDPR and Customer Data

By Camellia · Updated Oct 11, 2026

Data privacy in AI marketing is the set of obligations that apply the moment you paste, upload or sync customer data into an AI tool. The key questions are simple to state. Do you have a lawful reason to use this data for this purpose? Did you tell people? Is the vendor handling the data only on your instructions? And where does it end up geographically? This article is general information, not legal advice, so have privacy counsel review your actual setup.

It's a different job from collecting data well. The sibling article on first-party data for AI marketing covers what to collect and how to keep it clean. Here the data already exists, and the question is what you're allowed to do with it once an AI tool enters the picture. The examples lean on the EU's GDPR because it's the most detailed rulebook and many other regimes borrow from it, with notes on the UK, France, the US federal regulator and California.

Why AI changes the privacy question

Marketing teams have always shared customer data with software. What's new is how casually it happens. A campaign manager can drop a CRM export into a chat window to "find patterns." An email tool can switch on an AI feature that reads engagement history. A plug-in can summarize call recordings. Each of those is a new use of personal data, often by a new third party, and often decided by one person in an afternoon.

Three things make AI use different from a normal software rollout:

• Purpose drift. Data given for one reason (an order confirmation) ends up powering something else (a propensity model).
• New recipients. The AI vendor, and possibly its own sub-vendors, now handle the data.
• Opaque retention. Once data is used to train or fine-tune a model, deleting the original record may not remove its influence.

Regulators have said so directly. The European Data Protection Board's opinion on AI models, adopted on 18 December 2024, was issued at the request of the Irish supervisory authority to address data protection aspects of AI model processing. Its press summary adds that a model developed with unlawfully processed personal data could affect the lawfulness of its deployment, unless the model has been duly anonymised.

Key Facts: AI Marketing Data Privacy

• GDPR Article 6(1) lists six lawful bases, including consent and legitimate interests; legitimate interests apply only if they aren't overridden by the person's rights and freedoms (GDPR Article 6).
• Personal data must be collected for "specified, explicit and legitimate purposes" and be "adequate, relevant and limited to what is necessary" (GDPR Article 5).
• A processor may act only on "documented instructions from the controller" and needs prior written authorisation to engage a sub-processor (GDPR Article 28).
• People have the right to object "at any time" to processing for direct marketing, including related profiling, and processing for that purpose must then stop (GDPR Article 21).
• Transfers to third countries are allowed only if the Chapter V conditions are met (GDPR Article 44).
• Maximum GDPR fines reach EUR 20 million or 4% of global annual revenue, whichever is higher, for the most serious infringements (GDPR Article 83).
• The US FTC says it may be unfair or deceptive to adopt more permissive data practices, such as AI training, and tell consumers only through a surreptitious, retroactive change to terms (FTC, 13 February 2024).

Step one: know your role and your lawful basis

Under the GDPR you're usually the controller: the party that decides why and how personal data is used. The AI vendor is usually your processor, acting on your behalf. That split matters because the controller carries the main duties to customers.

Every use of personal data needs a lawful basis. Article 6(1) gives six options: consent, contract, legal obligation, vital interests, public interest, and legitimate interests. For marketing, three come up in practice.

The EDPB's AI opinion points to a three-step test for legitimate interests as the way to assess whether that basis works for AI-related processing. Its examples, a conversational agent and AI used for cybersecurity, can rely on legitimate interests only if the processing is strictly necessary and the balancing is respected. Marketing teams should read that as a warning against waving legitimate interests around as a catch-all.

Where does that leave a typical team? A practical split:

• Consent for anything that tracks behavior across sites or sends electronic marketing, where your local law requires it.
• Legitimate interests for lower-risk uses with a written balancing note, such as scoring existing customers for service relevance.
• A fresh look whenever the AI use differs from what the customer was originally told.

Step two: check purpose limitation before you upload anything

Article 5 requires personal data to be collected for specified, explicit and legitimate purposes, and not further processed in a way that's incompatible with them. This is the rule that most often catches AI projects.

Here's a concrete case. A customer gave their email to receive a webinar reminder. Months later, the team exports the full list and feeds it into an AI tool that builds lookalike audiences and writes personalized sales emails. The webinar notice never mentioned profiling. That's a purpose gap, and the fix is either a new notice and consent, or a documented legitimate-interests assessment, or dropping that dataset from the project.

France's regulator takes a pragmatic view of reuse. In its February 2025 AI recommendations, the CNIL says reusing databases, including online ones, is often possible provided the data was not collected unlawfully and the reuse fits the original collection purpose. The same page notes that minimisation doesn't prevent the use of large training datasets, though data should generally be selected and cleaned to avoid unnecessary processing of personal data.

A simple pre-flight test before any upload:

1. What was the person told when we collected this data?
2. Does the AI use fit inside that description?
3. If a customer saw exactly what we're about to do, would they be surprised?

If the answer to the third question is yes, stop and rethink.

Step three: minimize what the tool sees

Minimization is the cheapest privacy control you have. Article 5 requires data to be adequate, relevant and limited to what is necessary for the purposes. For AI marketing, that translates into a habit: send the tool the fields it needs, not the whole export.

Pseudonymization (swapping names for IDs) lowers risk but doesn't make data anonymous, because the key still exists somewhere. Treat pseudonymized data as personal data. Real anonymization is hard to prove; the EDPB says anonymity of a model is assessed case by case by data protection authorities, which is a reason not to assume your vendor's marketing claim settles it.

Special categories of data (health, religion, sexual orientation and similar) deserve extra caution. Don't let a model infer them to target people, and keep them out of ordinary prompts.

Step four: read the vendor terms like a contract, because it is one

Most privacy failures in AI marketing happen at the vendor boundary. A tool that's free for staff to sign up to can still receive customer data. Before any customer data goes in, your privacy lead should confirm four things.

A processor agreement exists. GDPR Article 28 says processing by a processor must be governed by a binding contract, in writing including electronic form, setting out the subject matter, duration, nature and purpose of processing, types of data and categories of people. The processor must act only on your documented instructions.

Sub-processors are controlled. The same article says the processor shall not engage another processor without prior specific or general written authorisation, and the first processor remains fully liable to you if a sub-processor fails. Ask for the current sub-processor list, and how you're notified of changes.

You know whether your data trains the model. This is the clause that matters most for AI. Look in the vendor's own current terms and data-processing addendum for statements on whether inputs, outputs or uploaded files are used to train or improve models, whether that's on by default, and whether you can switch it off. Terms for consumer accounts and for business plans often differ, and they change. If you can't find a clear answer, treat that as a "no" until the vendor confirms in writing.

Deletion and return are defined. Article 28 requires the processor, at your choice, to delete or return all personal data when the service ends. Ask how that applies to logs, prompts, embeddings and fine-tuned models, not just to tables in a database.

The FTC adds a second angle from the US side. Its 13 February 2024 post says it may be unfair or deceptive for a company to adopt more permissive data practices and tell consumers only through a surreptitious, retroactive amendment to its terms or privacy policy. That applies to you as much as to your vendor. If you decide to use existing customer data for AI work, update your notices openly and before you start, not after.

Step five: handle cross-border transfers deliberately

Many AI tools process data on servers outside your customers' home region. Under the GDPR, Article 44 says a transfer to a third country or international organisation, including onward transfers, may take place only if the Chapter V conditions are met. In practice that means one of the recognized mechanisms applies, such as an adequacy decision or appropriate safeguards. Article 46 lists standard data protection clauses adopted by the European Commission among those safeguards.

For a marketing team, the working questions are:

• Where does the vendor store and process data, including support access and sub-processors?
• Which transfer mechanism does the vendor rely on, and is it written into the agreement?
• Can you choose a regional data location, and is that choice real or just a setting for storage while processing happens elsewhere?

Keep a one-page record per tool: region, mechanism, sub-processors, date reviewed. If your company sells into several markets, local rules may add their own conditions, so don't assume one EU-style clause covers everything.

Step six: honor objection, profiling and transparency rights

AI marketing usually involves profiling: scoring, segmenting, predicting. Two GDPR rights bite here.

First, the right to object. Under Article 21, when personal data is processed for direct marketing, the person can object at any time, including to related profiling, and the data can no longer be processed for that purpose. So your suppression list has to reach every AI tool, not just the email platform. A customer who opts out and still appears in a model-built audience is a complaint waiting to happen.

Second, solely automated decisions. Article 22 gives people the right not to be subject to a decision based solely on automated processing, including profiling, where it has legal or similarly significant effects. Most routine marketing doesn't reach that bar. Decisions about credit offers, pricing that excludes people from essential services, or eligibility do, so involve counsel before automating them.

Transparency ties it together. Your privacy notice should say, in plain words, that you use customer data with AI-assisted tools for personalization and analytics, and name the purposes. California's attorney general summarizes the state's equivalent: businesses must give a notice at collection listing the categories of personal information collected and the purposes, and must honor a user-enabled global privacy control as a valid opt-out request. California residents can also ask businesses to correct or delete their data and to limit use of sensitive information. The state's privacy agency has also published regulations covering automated decisionmaking technology, risk assessments and cybersecurity audits; its regulations page lists them among completed packages, so check current requirements with counsel if you serve California customers.

Step seven: assess risk before you scale

When a use of AI is likely to result in a high risk to people, Article 35 requires a data protection impact assessment before you start. The article names new technologies as a factor, and it specifically calls for one where there's a systematic and extensive evaluation of personal aspects based on automated processing, including profiling, that supports decisions with legal or similarly significant effects. Many pure marketing uses won't meet that trigger, but large-scale profiling or sensitive data might meet the general high-risk test.

The UK's ICO publishes guidance on AI and data protection, with chapters on lawfulness, security, minimisation and accountability, including what to consider in an assessment. The ICO notes that its page is under review following the Data (Use and Access) Act and may change, so read it as current direction, not fixed text.

A lightweight assessment for each new AI use is worth the hour it takes. Record: purpose, data fields, lawful basis, vendor and region, retention period, who can see outputs, and what you'll do if a customer objects.

Privacy risks by marketing function

A practical rollout checklist

1. Inventory your AI tools. Include features that were switched on inside existing platforms, and the unofficial tools staff use on their own.
2. Map each tool's data. Which fields go in, from which system, for which purpose.
3. Assign a lawful basis per purpose. Write down consent or the legitimate-interests reasoning.
4. Minimize. Strip identifiers and unneeded fields before any upload or sync.
5. Review the vendor agreement. Confirm processor terms, sub-processors, training use, retention, deletion and region.
6. Document transfers. Record the mechanism for each cross-border flow.
7. Wire up suppression. Make sure opt-outs and objections reach every tool.
8. Update notices. Say plainly how AI-assisted processing is used, before you start.
9. Write the staff rule. One page: approved tools, what data can go where, who to ask. A longer treatment is in AI marketing policy.
10. Review quarterly. Vendor terms, sub-processors and laws change.

The EU's separate AI rulebook is explained in EU AI Act. If your AI use extends to autonomous agents with system access, see AI agent data privacy. For the lead-capture side of GDPR, the guide on GDPR lead capture in the EU is the companion read.

Common pitfalls

• Treating the AI tool as "just software." It's a processor handling personal data, and it needs a contract.
• Pasting exports into consumer chat tools. The most common leak, and the easiest to prevent with an approved-tool list and a clear rule.
• Assuming consent covers everything. Consent is tied to specific purposes. A broad checkbox buried in terms is weak evidence.
• Skipping the training question. If you don't know whether the vendor trains on your inputs, you don't know your exposure.
• Forgetting suppression. Opt-outs that stop at the email platform leave customers in model-built audiences.
• Quietly widening use. Changing notices after the fact is the pattern the FTC has warned about.
• Ignoring where data goes. Regional storage doesn't settle where processing and support access happen.
• No owner. Privacy duties that belong to "everyone" belong to no one. Name a person per tool.