Right to explanation

In B2B finance, an AI decision which rejects a payment or flags an anomaly has real consequences for the business behind it. The person affected, a vendor, an employee, or the customer themselves, has the right to understand why. BAGF commits to the right and defines exactly how we deliver on it.

Context

The right to explanation is established in Decree 13/2023 (Art. 19), in GDPR (Art. 22), and now in the EU AI Act (Art. 86). What varies between jurisdictions is the procedure. Who asks. How fast we respond. What level of detail we provide. We standardize on a single internal process which satisfies all three.

Principle: The explanation rule

Three classes of requester demand an explanation. The customer (data controller) for any decision affecting their tenant. The data subject for any automated decision which materially affects them. The regulator under audit. Each class gets a defined response within a defined window.

Who asks

• The customer (data controller). Any AI decision affecting their tenant.
• The data subject. Any individual whose data was processed in a materially impactful automated decision (Decree 13 Art. 19, GDPR Art. 22).
• A regulator. Any decision in scope of an active audit or inspection.

Self-service. The first layer

Most explanation requests are answered without a ticket through:

• Confidence indicators in the UI.
• Citation beneath each decision.
• Audit log viewer in the Customer Portal.
• Explanation API for programmatic access (see §8).

A customer’s audit team usually answers their own question in minutes. The Portal is the right tool at the scale.

Through support. The second layer

For requests which need a written investigation:

1. The customer submits a request through the contracted channel with “Explanation Request” priority.
2. Bizzi responds within 30 working days, matching Decree 13 Art. 19.
3. The response covers four points. An overview of the AI logic for this use case. The top factors which influenced the decision. The HITL approver if one was involved. A counterfactual. If the input had been different in this specific way, the output would have been different in this specific way.

Three levels of detail

The depth of explanation matches the audience:

• Level 1. User-facing. Plain language for end users. Focused on why, not how.
• Level 2. Technical. More detail on model, prompt, and confidence. For the customer’s IT or audit team.
• Level 3. Forensic. Full reasoning trace, model version, raw input and output. For compliance investigations or legal proceedings.

Level 1 and Level 2 are available through the Portal without ticket. Level 3 needs a verified-identity request because of the detail it exposes.

What the right does not cover

• Not every AI suggestion needs full explanation. Low-stakes assistance (high-confidence auto-fill, category suggestions) is adequately explained by the UI’s confidence indicator.
• Not every explanation is technically achievable. As discussed in §9, high-dimensional transformer outputs do not have classical explanations. We provide proxy explanations (citation, confidence, comparable cases) instead, and we are explicit when we do.
• Bizzi’s implementation is not disclosed. Customers have a right to behaviour. They do not have a right to system prompts or model internals.

Standards mapping

Standards Mapping

Decree 13/2023. Đ. 19 (Automated decision-making), Đ. 14 (Data subject rights). GDPR. Art. 22 (Automated individual decision-making), Art. 15 (Right of access). EU AI Act. Art. 13 (Transparency), Art. 86 (Right to explanation of individual decision-making). ISO/IEC 42001. A.6.2 (Roles), A.8.2 (Performance).