Skip to content
Pillar III: Ethics, Transparency, Interpretability · § 08

Transparency in serving. Explanation API

Customers do not need to file a ticket to audit an AI decision in their own tenant. Bizzi exposes an Explanation API and a Customer Portal which return the data, the reasoning, and the version metadata behind any decision the AI has made. Bounded by tenant isolation and rate limits.

The traditional vendor pattern is opaque. A customer asks “why did your AI flag this invoice?”. A support engineer queries logs. A written response arrives three days later. The latency is incompatible with month-end close. We invert the model. The customer queries directly. Support involvement is the exception, not the path.

GET /api/v1/ai-decisions/{transaction_id} returns for any decision:

  • Model version in effect at the time of the decision.
  • Prompt version as registered in the prompt registry.
  • Input after PII redaction.
  • Output as written to the tenant.
  • Confidence per field.
  • Reasoning trace. The agent’s chain of reasoning captured through the observability layer.
  • HITL approver if applicable, with timestamp.
  • Audit trail timestamps for each lifecycle event.

Access uses a customer API key scoped to the tenant. Calls are rate-limited and every access is itself logged for audit. The endpoint is part of the contracted product, not a special exception.

The Portal exposes the same data through a UI for non-engineering users:

  • Audit log viewer. Search, filter, export.
  • Activity timeline. Transactions plotted over time with AI confidence overlaid.
  • Drill-down. Click a transaction to see the full reasoning trace and citations.

Internal audit teams use the Portal to spot-check batches. Investigators use it to dig into a specific transaction. Trainers use it to show new accountants how the AI reasons. External auditors use it as evidence in their own workpapers.

For concerns which need investigation beyond what the Portal exposes, the customer files an “AI Behaviour Inquiry” ticket. We respond with a detailed report within five business days. If the behaviour is a bug, it is fixed under the severity-appropriate SLA. If the behaviour is intended, the response explains it and links to the Model Card or documentation which confirms the design.

Some AI decisions do not have a classical explanation. High-dimensional transformer outputs are not derived from a feature-importance table. When this is the case, we are explicit. We provide proxy evidence. Confidence scores, similar past decisions, evaluation results. We tell the customer when classical explanation is not available. If the customer does not accept the proxy, the fallback is manual processing, not a fabricated rationale.

We do not disclose:

  • Bizzi’s proprietary system prompts. Their content would enable prompt injection against our customers.
  • Vendor model weights. We license them. We do not sublicense.
  • Cross-tenant data. A customer’s explanation contains only the querying customer’s data.

Customers audit behaviour in detail. They do not have a right to audit Bizzi’s implementation details.