Every company that uses AI for marketing faces the same question: can I put this data into the tool? The short answer is "it depends on what data, with what legal basis, and under what contract." The long answer is what this article organizes.
The risk is not abstract. Customer data in a tool without a processing contract, without a defined legal basis, and without retention control is exactly the type of situation that generates notification to ANPD, requests from data subjects, and, ultimately, legal action.
First: what is personal data here
Many people think that only CPF and email count. No. Personal data is any information that identifies or can identify someone: name, phone number, email, IP, device identifier, browsing history, approximate location, service recording, photo, voice audio.
And there is the sensitive category: health, biometrics, racial origin, religious belief, sexual orientation, children's data. For these, the standards are much stricter — and this is where many AI projects in health and education fail.
The questions before uploading any database
1. What is the legal basis?
For marketing, the usual bases are consent, legitimate interest, or contract execution. Each requires different documentation. Legitimate interest is not a free pass: it requires a registered assessment that balances your interest and the rights of the data subject.
2. Does the data need to be identifiable?
To train a model, almost never. Anonymizing or pseudonymizing beforehand drastically reduces exposure. If the tool does not need to know who the person is, do not disclose who they are.
3. Does the provider train with your data?
Check the contract and settings. Many tools use input data to improve models by default. Disable this and document the decision.
4. Where do the data stay and for how long?
Storage country, retention period, deletion process. International transfer requires a specific clause. "I don't know" is not an acceptable answer in an audit.
5. How does the data subject exercise their rights?
There needs to be a real path for access, correction, and deletion — including data that is in third-party tools. If the data is in five systems, deletion needs to reach all five.
Practices that reduce risk without killing the project
- Minimization: send only the necessary columns. Name and CPF are rarely needed for segmentation.
- Pseudonymization: replace identifiers with internal codes; keep the map separate.
- Controlled environment: use corporate instances, with access logs and no training on your data.
- Processing contract: operator document, with purpose, security, and subprocessors.
- Decision records: one page per project explaining legal basis, data used, and measures taken. It’s worth more than a generic policy on the site.
The most common mistake
Putting a lead spreadsheet with name, phone number, and conversation history into a free tool to "see what AI does." The productivity gain is marginal; the exposure is total. If the test is necessary, use synthetic data or an anonymized sample.
Also read:
- consulting on AI agents
- data privacy in marketing
- first-party data without cookies
- responsible adoption roadmap for AI agents
Compliance as a business advantage
Companies that can answer "yes, we have a contract, legal basis, and deletion process" close larger B2B contracts more easily. Data security has become a purchasing criterion — and the marketing team that organizes this beforehand delivers a sales argument, not just a legal obligation.
The golden rule is simple: treat customer data as you would treat cash. It does not leave the company without control, record, and reason.

