Data protection and privacy under Decree 13/2023

Invoices and financial documents are full of personal data. Buyer names, phone numbers, individual tax IDs, sometimes national ID numbers. The moment that data crosses into a commercial LLM API, three risks open up. The vendor retains it to train their model. It leaks through logs or shared response paths to another tenant. Someone unauthorized accesses it. Decree 13/2023 requires you to control all three. This section explains how.

Context

The seven principles in Decree 13/2023 are not abstract aspirations. They are enforceable rules with breach-notification obligations attached. Your job as a B2B AI processor is to map each principle to a concrete product control. An auditor or a customer DPO must trace any of your claims back to a running mechanism.

How we implement

The seven principles of Decree 13/2023 and the Bizzi control implementing each.

#	Principle (Decree 13)	How Bizzi implements it
1	Lawful, fair, transparent	Customer Disclosure Policy. AI Disclaimer in the UI
2	Purpose limitation	Every AI feature has documented data lineage. No re-purposing outside the registered purpose
3	Data minimization	PII Redaction at the gateway before any LLM call. Only the minimum required data leaves your perimeter
4	Accuracy	Full audit trail on every modification. DSAR response in under 30 days
5	Storage limitation	Retention policy per the Data Classification Standard (Pillar IV §3). Automatic deletion at expiry
6	Integrity and confidentiality	AES-256 at rest, TLS 1.3+ in transit. Access controls per Pillar V §9
7	Accountability	Named DPO. Live ROPA. Internal audit twice a year (Pillar I §10)

PII Redaction is the central technical control

Bizzi Practice: PII Redaction at the gateway

Every piece of personal data appearing on an invoice or document is auto-redacted at the gateway before the data is persisted or sent to any LLM. Buyer name, individual phone number, national ID, individual tax ID. The LLM never sees the original PII.

The pipeline.

1. OCR extracts text from the PDF or image.
2. The PII detector scans the OCR output and identifies personal fields.
3. Detected PII is replaced with placeholders. [PERSON_1], [PHONE_1], [NATIONAL_ID_1].
4. The redacted text is sent to the LLM for downstream processing.
5. The LLM response returns to the Bizzi backend. The placeholder-to-real-value mapping lives only in the internal OLTP database, behind strict access controls.
6. The UI re-hydrates real values for authorized users. Reports and audit logs carry only placeholders.

Consent management for training data

Bizzi does not use customer personal data to train models without explicit consent. The concrete rules.

• Base model training data. Anonymized or synthetic only.
• Customer fine-tunes. Confined to that customer’s tenant, never cross-tenant, governed by the DPA.
• Evaluation data. Anonymized, or covered by explicit consent.

DSAR. Data Subject Access Requests

Data subjects have the right to access, correct, and delete their data. The DSAR workflow.

1. The customer (as Controller) forwards the request to Bizzi through the contracted channel.
2. Bizzi verifies the subject’s identity.
3. Bizzi retrieves the relevant records via audit trail and data lineage.
4. Response is delivered within 30 business days with the full report.
5. Deletion is propagated across OLTP, OLAP, vector DB, and backups to the technically feasible boundary defined in the retention policy.

Standards mapping

Standards Mapping

Decree 13/2023 Articles 4 to 19 (principles and subject rights), Article 22 (DPIA). ISO/IEC 42001 A.7.2 (data quality). EU AI Act Article 10 (data and data governance).