AI Governance & Operations
Ensure AI reliability, safety, and cost control.
UC-AIML-003: AI Guardrails & Safety
Purpose: Prevent hallucinations and ensure brand-safe AI interactions.
| Property | Value |
|---|---|
| Actor | System / AI Manager |
| Trigger | AI Generation Event |
| Priority | P0 |
Capabilities:
-
Content Filtering: Block profanity, PII (Personal Identifiable Information), or competitor mentions in generated messages.
-
RAG / Search Endpoints:
-
Context Injection: Retrieve relevant docs (e.g., "Refund Policy") before answering queries.
-
Citation: AI links to the source policy "According to [Refund Policy]...".
-
-
Budgeted Automation Limits:
-
Cost Control: "Stop AI generation if monthly API bill > $500".
-
Volume caps: "Max 500 AI calls/day per tenant".
-
Main Success Scenario:
- User asks AI: "Write a rude message to the client."
- Guardrail detects negative sentiment/profanity logic.
- System overrides output: "I cannot generate offensive content. Here is a professional alternative."
UC-AIML-004: Automated ML Retraining
Purpose: Keep models fresh with recent data.
| Property | Value |
|---|---|
| Actor | MLOps System |
| Trigger | Schedule (Weekly) / Drift Detection |
| Priority | P1 |
Capabilities:
-
Pipeline Automation:
-
Ingest new booking/cancellation data.
-
Re-train "No-Show Prediction" model.
-
Compare accuracy vs previous model (Shadow Mode).
- Auto-promote new model if Accuracy > Threshold.
- Rollback if performance degrades.
-
Acceptance Criteria:
- [ ] Pipeline runs automatically on schedule (e.g., Weekly).
- [ ] Drift detection alerts admins if model accuracy drops < 80%.
- [ ] Shadow mode allows testing new models without affecting live users.
Related Use Cases
-
Training Pipelines: Baseline logic for model training.
-
AI Clienteling: Consumes the "Booking Probability" model.