Skip to content
Pillar III: Ethics, Transparency, Interpretability · § 07

Transparency in operations

When an AI fills a field, the user knows it was filled by AI and how confident the AI was. When a user talks to a chatbot, the user knows it is a chatbot. When the platform is degraded, the user knows without opening a ticket. These three signals (confidence, disclosure, status) are the operational face of transparency.

The most common cause of trust erosion in production AI is not error rate. It is ambiguity. Users find out, weeks in, a field they assumed they had typed was AI-completed. Or an answer they treated as company policy came from a chatbot’s best guess. Once it happens, every AI suggestion is suspect. Transparency at runtime exists to prevent the discovery.

ConfidenceVisualSystem actionUser action
≥ 95%Green background, check iconAuto-fill, write audit logOptional review
80–94%Light yellow, warning iconAuto-fill, highlightedReview recommended
< 80%Light red, alert iconHighlight only, do not auto-fillVerify required (HITL)
ManualWhite backgroundNoneEntered by the user

The visual encoding never relies on colour alone. Every confidence band carries an icon shape and text annotation so colour-vision deficient users see the same information (see §4 on inclusivity).

The confidence numbers are calibrated. When the model reports 95%, the empirical accuracy is between 94 and 96%. We verify calibration monthly against human-labelled ground truth. A confident-looking model whose confidence does not predict accuracy is worse than no confidence indicator at all.

“You are speaking with an AI Assistant. Answers are imperfect. Please verify financial information before making decisions.”

The disclaimer is shown at the top of the chat (not below the fold), once per session as a notice, and afterwards as a persistent icon next to the assistant’s name. It is one sentence, not a legal wall.

status.bizzi.vn (or its equivalent) shows the health of the core services. OCR, AI assistant, ERP sync, AI Gateway. Along with real-time latency, uptime, incident history with severity, and scheduled maintenance windows. The status page does not expose tenant-level metrics, customer identities, or internal tooling errors invisible to end users.

  • Model upgrades. Communicated via release notes. If behaviour changes meaningfully, an in-app notice runs for the affected feature.
  • Kill-switch triggers. Affected customers are notified immediately with an ETA.
  • Scheduled maintenance. Production maintenance windows longer than 30 minutes are announced at least 7 days in advance.

Customers reach their tenant’s audit trail directly through the Customer Portal. No ticket required. The audit log lists every AI action on tenant data, the model and prompt versions in effect, confidence scores per field, HITL approver and timestamp. It exports cleanly to CSV for internal audit workflows.