GDPR and AI Marketing Tools: What to Check Before You Generate a Single Ad

Before your team uses any AI tool to write or design ads, treat GDPR and AI marketing tools the way you would treat any supplier that touches personal data: check what goes into the prompt, what the vendor does with it, where it travels, how long it stays, and whether your customers were told. Most ad copy needs no personal data at all, so the safest setup is often the simplest. This checklist covers the GDPR questions, the EU AI Act points that touch marketing as of October 2026, and a table of questions to send your vendor.

For claim checking and brand voice, see our guide to AI-powered advertising; for choosing between tools, the buyer's guide to AI advertising platforms. This article stays on data protection.

1. Make an inventory of what would go in

The GDPR applies the moment personal data reaches the tool, whether or not the output is ever published. Personal data is any information relating to an identified or identifiable person (Article 4(1)). Walk through the inputs your team plans to use:

Input	Personal data?	Lower-risk alternative
Customer or lead lists (names, emails, purchase history)	Yes	A written audience description, such as "repeat buyers of running shoes, price-sensitive, mostly shopping on mobile"
CRM notes and sales call notes	Yes, often with opinions about the person	A summary of common objections with names, companies and contact details removed
Customer reviews and support emails	Yes, when they carry names, usernames or details that point to a person	Paraphrased themes ("buyers mention fast setup"), no named quotes
Photos or videos of customers, staff or models	Yes, if the person can be identified	Licensed images or images with written permission, plus a clear answer on what the tool does with faces
Campaign reports with user IDs, click IDs or IP addresses	Yes	Aggregated results by segment, keyword or creative
Product specs, prices, offer terms, brand guidelines	Normally no	Use freely, subject to your own confidentiality rules

On photos: under Recital 51 a photo counts as biometric data only when processed through specific technical means that allow unique identification. If a tool runs face matching on uploads, find out before anyone uploads a team photo.

Minimise first, then pseudonymise

Data minimisation (Article 5(1)(c)) is your strongest control, because copy for a segment rarely needs any individual's data. Where a task needs real records, such as replies to reviews, swap names and contact details for placeholders and keep the key yourself.

The EDPB's Guidelines 01/2025 on pseudonymisation confirm that data which can still be attributed to a person with additional information remains personal data. Pseudonymisation lowers risk and supports data protection by design (Article 25), but the GDPR still applies. The proposed Digital Omnibus on data would change how identifiability is judged for recipients of pseudonymised data, but as of October 2026 it is still in first reading and not law.

2. Pin down the vendor's role

An AI tool that processes personal data on your behalf is your processor (Article 4(8)). Article 28 requires a contract limiting processing to your documented instructions and covering confidentiality, security, help with data subject requests, deletion or return at the end, and audits. Our explainer on data processing agreements for marketers goes through it clause by clause. Three points are specific to AI tools:

• The model provider is a subprocessor. Most ad tools call a model run by another company, which belongs on the vendor's subprocessor list with its processing location. Under Article 28(2) a processor needs your prior authorisation for subprocessors and, under a general authorisation, must tell you about additions or replacements so you can object. Vendors switch models often, so subscribe to change notices.
• Liability runs down the chain. The vendor must impose the same obligations on its subprocessors and stays fully liable to you if they fail (Article 28(4)).
• Own-purpose use changes the role. A vendor that trains its own models on your inputs is deciding purposes and means. Article 28(10) makes a processor that does this a controller for that processing, and you would need a legal basis and transparency for handing the data over. For most teams the answer is to switch it off.

3. Training on your inputs: get the answer in writing

"Do you train on our data?" needs a precise answer. Ask separately about prompts, uploaded files and outputs; about the vendor and each model provider behind it; and whether the answer differs between free and paid plans or individual and team accounts. Then check where the commitment lives: a clause in the DPA is stronger than a settings toggle any user could switch back, so make sure only an admin can change it.

France's CNIL, in its first guidance on deploying generative AI (July 2024), tells organisations to work out how far the provider may reuse the data supplied, set rules on what staff may enter, and train users before rollout. That makes a good outline for an internal AI use policy.

There is also the model's past. In Opinion 28/2024 on AI models (December 2024), the EDPB said a model developed with unlawfully processed personal data can affect the lawfulness of its deployment unless it has been duly anonymised. Regulators should consider whether the deploying organisation made an appropriate assessment, for instance of the data's source and any infringement finding by a regulator or court, scaled to the risk. For copy written from product briefs the risk is modest; note what you checked.

4. Know where the data goes

If prompts or files are processed outside the EEA, Chapter V of the GDPR applies. Ask the vendor to name every processing region, including support access, and check the transfer tool for each:

• Adequacy. The Commission's adequacy list includes the UK, Switzerland, Japan, the Republic of Korea, Canada (commercial organisations) and, since January 2026, Brazil, among others. For the US, adequacy covers only organisations certified under the EU-US Data Privacy Framework, so confirm the vendor and its model provider are certified.
• The framework in court. On 3 September 2025 the EU General Court dismissed an action to annul it (Case T-553/23, Latombe v Commission); an appeal was lodged on 31 October 2025 (Case C-703/25 P). It is valid today, but keep a fallback in your contracts.
• Standard contractual clauses. Without adequacy, the usual tool is the Commission's 2021 clauses (Decision 2021/914) under Article 46(2)(c), with an assessment of the destination country's law.

Your privacy notice must mention transfers and their safeguards (Article 13(1)(f)).

5. Retention and deletion

Storage limitation (Article 5(1)(e)) applies to prompt logs as much as to your CRM. Ask how long the vendor keeps prompts, uploads, outputs and any abuse-monitoring copies; whether you can delete a single project rather than the whole account; whether deletion reaches the model provider and backups; and what happens when you cancel, since Article 28(3) requires deletion or return at the end of the service unless law requires storage. Record the periods in your Article 30 record.

6. Decide whether you need a DPIA

Article 35 requires a data protection impact assessment before processing likely to result in a high risk to people, particularly with new technologies. The WP248 guidelines endorsed by the EDPB list nine criteria, among them evaluation or scoring, combining datasets, large-scale processing, vulnerable data subjects and innovative technology, and say processing meeting two will in most cases need a DPIA. National authorities also publish lists of processing that always needs one.

Marketing use	DPIA likely?	Why
Writing ad copy from a product brief with no personal data	No	No personal data processed
Turning reviews into ad copy after removing names	Usually no; record the reasoning	Limited, pseudonymised input
Personalised messages per contact from CRM history and inferred interests	Likely yes	Evaluation, combined datasets, new technology, often at scale
AI that scores leads or decides who receives an offer	Yes, and check Article 22	Scoring plus decisions that can significantly affect people

Writing ad copy is not a decision about a person, so Article 22 rarely applies to content generation. It can apply where a system alone decides, with legal or similarly significant effects, who is never offered a price or a service. Then you need one of its exceptions (contract, law or explicit consent), human intervention on request, and meaningful information about the logic in your privacy notice (Article 13(2)(f)).

7. Tell your customers

If customer data goes into an AI tool, your privacy notice should cover the purpose and legal basis, recipients or categories of recipients (Article 13(1)(e)), transfers, and retention (Article 13(2)(a)). Document your legal basis, legitimate interest or consent, before the first upload. People can ask what you hold and who received it; see our post on access requests marketing teams get wrong. Uploading customer lists to ad platforms is a separate activity with its own rules, covered in GDPR and retargeting.

8. EU AI Act points that touch marketing

The AI Act separates providers, who place an AI system on the market, from deployers, who use one in a professional context (Article 3(4)). A business generating ads with a tool is usually a deployer. Two duties matter now:

• AI literacy (Article 4). Applies since 2 February 2025. The Digital Omnibus on AI (Regulation (EU) 2026/1744, in force since 27 July 2026) replaced the duty to ensure AI literacy with a softer one: deployers must take measures to support the development of AI literacy of their staff and others operating or using AI systems on their behalf, and are not required to guarantee any specific level. In practice: brief staff on what never goes into a prompt and how to check outputs, and keep a record.
• Deep fake disclosure (Article 50(4)). Since 2 August 2026, a deployer that uses AI to generate or manipulate image, audio or video content resembling existing persons, objects, places, entities or events, which would falsely appear authentic, must disclose that it is AI-generated or manipulated. A real spokesperson's cloned voice, a real customer's photo altered to show your product, or an edited image of a real venue can all count when the result would pass as authentic. The disclosure must be clear at the latest at first exposure (Article 50(5)). A fully invented face is less clear-cut, but presenting it as a real customer would mislead anyway, so disclose.

A real person's likeness is also their personal data, so the GDPR checks above apply too. Providers must separately mark synthetic output in a machine-readable way (Article 50(2)); for systems on the market before 2 August 2026, the Omnibus moved that deadline to 2 December 2026. Section 7 of our AI-powered advertising guide covers Article 50 and ad text in depth.

9. Questions to send your vendor

Question	A good answer	Why it matters
Is there a DPA, and does it cover the AI features?	A DPA you can sign before use that names the AI processing	Article 28(3)
Which model providers receive our prompts, and where?	Named subprocessors with processing regions	Article 28(2), Chapter V
Are prompts, uploads or outputs used to train any model, and who can change that?	"No" in the DPA for every plan you use; setting admin-only	Article 28(10)
How long are prompts, logs and outputs kept?	Stated periods, including abuse-monitoring copies, and deletion per project	Article 5(1)(e), Article 28(3)
Which transfer mechanism covers each non-EEA recipient?	Data Privacy Framework certification or SCCs, named per recipient	Article 13(1)(f), Article 46
How will we hear about subprocessor changes?	Advance notice with time to object	Article 28(2)
Do you analyse faces or voices in uploaded media?	A clear yes or no, with the purpose	Biometric data (Article 4(14))
How do you mark AI-generated images, audio and video?	A description of the machine-readable marking	AI Act Article 50(2)
How do you help with access requests, deletion and DPIAs?	A named process and contact	Article 28(3)

10. Common mistakes

• Pasting a CRM export "to make the ads more relevant" when a segment description would do.
• Staff using personal or free AI accounts for company work, outside any DPA or admin control.
• Checking the subprocessor list once and never again after the vendor changes model provider.
• Uploading photos of customers or staff without permission or without asking what the tool does with faces.
• Running a cloned voice or edited image of a real person or place without a disclosure.

11. Checklist before the first generation

1. List every AI tool in use, including free accounts staff opened themselves.
2. Record what data goes into each, using the table in section 1.
3. Replace personal data with segment descriptions or placeholders.
4. Sign the DPA, save the subprocessor list and subscribe to change notices.
5. Get a written answer on training; lock the opt-out at admin level.
6. Confirm the transfer mechanism for each non-EEA recipient.
7. Note retention periods and set your own for briefs and drafts.
8. Screen against the DPIA criteria; run one where two or more apply.
9. Update your privacy notice and Article 30 record.
10. Brief staff with a one-page AI use policy and add a disclosure step for realistic synthetic media.

Where Adverizeo fits

Adverizeo writes ad copy for 24 channels in 20+ languages, and every text ad gets a compliance score from 0 to 100 with suggested fixes. That score is a risk check, not a legal guarantee, and does not replace the vendor checks above. You can test an existing ad with the free GDPR ad checker, read our GDPR page, or start on the free plan (10 credits a month, no card). For consent and tracking beyond AI tools, see our guide to GDPR-compliant marketing for European SMEs.

This article is general information, not legal advice.