When a customer says their CPF or CURP on a call, a well-designed AI agent transcribes the number, replaces it with a token before sending it to the model, uses the real value only inside the tools that verify identity, keeps it out of the agent’s replies and stores the record without the full value. Each of those steps can be audited.
It is the second of the 12 questions a Risk team asks before approving an AI agent: what happens to a piece of personal data when the agent hears it. Many vendors answer “we protect it”. A bank needs a more precise answer: where the data travels, who sees it and what gets stored.
This article follows a CPF or a CURP through a conversation, summarizes what Brazil’s LGPD and Mexico’s personal data law require, and proposes five tests to run with a vendor. It is the foundation of any PII policy for AI agents.
Why a CPF or a CURP needs more care than other data
The CPF (Brazil) and the CURP (Mexico) uniquely identify a person and are used to open accounts, apply for credit and verify identity in almost every channel. With one of them and a couple of other details, a third party can try to impersonate the customer. That is also why they are the first data points someone looks for when trying to trick a conversational agent.
Voice agents add the recording. A voice is personal data, and if the institution uses it to identify the customer biometrically, the LGPD classifies it as sensitive data, with stricter rules.
The path of personal data in five stops
This is what the path looks like in an architecture built for banking, from the moment the customer says the number until the conversation is archived.
1. Capture: from audio to text
The customer says the number out loud and the speech recognition engine turns it into text, turn by turn. From then on the data exists in two forms, audio and transcript, and both need the same level of protection.
2. Before the model: tokenization at the edge
Before the transcript reaches the language model, the number is replaced with a token. The model knows the customer gave their CPF but never receives the value. Since in most architectures the model is provided by a third party, this step keeps the personal data from leaving the institution on every turn.
3. Use: verification happens in the tool
To confirm identity, the agent calls a tool that queries the bank’s systems. That query uses the real value inside the authorized perimeter and returns only the result to the agent: identity confirmed or not confirmed. The model decides the next step based on that answer, without handling the number.
4. Output: the agent does not repeat what it should not
A code-level filter checks every reply before it reaches the customer. If the agent tries to read a full ID number out loud, or someone tries to get it to reveal another person’s data, the filter blocks the output and the agent replies with the defined message or hands off to a human. This control has to live in code and not in the prompt, because an instruction to the model can be bypassed with a well-crafted question.
5. Record: the log keeps the interaction, without the data
The conversation is recorded for audit with the data tokenized or masked. Anyone who needs the full number, for example to handle a complaint, accesses it with an authorized role, and that access is logged too. What else needs to be recorded is covered in what should be recorded in every conversation. When the retention period ends or the data subject requests deletion, the process has to reach the recording, the transcript and the copies in connected systems.
What the LGPD and Mexican law require
Both laws treat the CPF and the CURP as personal data and ask the institution, as the party responsible for processing, for essentially the same things: inform the data subject, protect the data and be able to show how it was handled. The agent vendor acts as a processor, and its architecture is part of that evidence.
| Topic | Brazil · LGPD | Mexico · LFPDPPP |
|---|---|---|
| Authority | ANPD | Secretaría Anticorrupción y Buen Gobierno, since March 21, 2025 |
| Transparency | Inform the data subject of the purpose and form of processing | Privacy notice before processing |
| Data subject rights | Access, correction, deletion and review of automated decisions | ARCO rights: access, rectification, cancellation and objection |
| Security | Technical and administrative measures to protect the data | Administrative, technical and physical security measures |
In Brazil, if an incident may cause relevant risk or harm to data subjects, ANPD Resolution CD/ANPD 15/2024 gives the controller three business days to report it to the ANPD and to the people affected. To meet that deadline, the bank needs to know quickly where each piece of data was and who accessed it, and that information comes from the record.
In Mexico, the new Federal Law on the Protection of Personal Data Held by Private Parties took effect on March 21, 2025 and moved the INAI’s functions to the Secretaría Anticorrupción y Buen Gobierno. What each regulator expects from an AI agent is covered in LGPD, Banco Central do Brasil and CNBV.
Five tests to run with a vendor
- Say a test CPF or CURP on a call and then search for it in the logs: it should appear tokenized or masked.
- Ask to see what the model provider received on that same turn. The number should not be there.
- Try to get the agent to read another customer’s ID number out loud and check how the block was logged.
- Ask for the access log: who saw the full value, when and with which role.
- Ask how a deletion request is carried out and which systems it reaches.
If any of these tests cannot be run in the demo, it will come back as an objection in the Risk review.
How ADI Platform handles it
In ADI Platform, blocking personal data exfiltration is one of three mandatory guardrails, along with prompt injection protection and crisis signal detection. None of them can be turned off in any agent, and all three run in code, before and after the model.
Personal data is tokenized at the edge: it does not go out to the model provider, and the audit log records the interaction without the data. A configurable filter for sensitive data in outputs adds a second barrier on what the agent replies. Each institution’s data is isolated by organization, with server-side enforcement, and access is role-based, with every administrative action logged. When the bank’s policy requires it, the platform is deployed in a dedicated VPC or in the institution’s private network.
On voice, the agent’s pronunciation dictionary handles acronyms such as CPF, CURP or RFC so they sound the way customers say them in each market.
See ADI Platform security and guardrails
Frequently asked questions
Can an AI agent hear a CPF or a CURP without breaching the LGPD or Mexican law?
Yes, if the institution has a legal basis for the processing, informs the data subject and applies security measures. In an AI agent, that means tokenizing the data before the model, blocking it from the agent’s replies and logging who accesses the full value.
What does tokenizing personal data mean?
Replacing the real value with an identifier that carries no meaning. The model and the logs work with the token, and the real value is only used inside the institution’s authorized systems.
Does the language model provider receive the CPF or CURP?
It should not. With tokenization at the edge, the number is replaced before it reaches the model, so the LLM provider never receives it in any turn.
How long does a company have to report a personal data incident in Brazil?
Three business days from the moment the controller learns that the incident affected personal data, under ANPD Resolution CD/ANPD 15/2024, with up to 20 business days to complete the information.
Who is the personal data protection authority in Mexico?
Since March 21, 2025, the Secretaría Anticorrupción y Buen Gobierno, which took over the INAI’s functions under Mexico’s new federal law on personal data held by private parties.
Keep reading
- AI agent audit: what should be recorded in every conversation
- LGPD, Banco Central do Brasil and CNBV: what each regulator requires from an AI agent
Informational content. It does not constitute legal advice. Check the current regulation and your legal team for each case.
