1.0
Test the support contract
Evals make sure the product still classifies correctly, blocks unsafe sends, and cites approved sources after model or prompt changes.
- Classifier checks
- Policy checks
- Knowledge extraction
Home
Resource
Evals
Replayable tests for classification, risk, policy decisions, knowledge extraction, and draft quality before support automation changes ship.
Replay
Examples
Known support scenarios test behavior.
Quality
Measured
Draft and policy outcomes are checked.
Release
Safer
Changes can be evaluated before launch.
2.0
Use real support shapes
Examples cover refund requests, hidden HTML instructions, prompt injection, and normal inbound support mail.
- Prompt injection
- Duplicate inbound event
- Refund request
Operations
What this means in production.
These pages describe the product contract behind the UI, not a decorative brochure. Each surface should connect back to the same governed support loop.
Control boundary
AI outputs are schema-validated.
Operator workflow
Support moves through secure ingest, AI preparation, human approval, outbound execution, and audit evidence.
Configuration impact
Provider setup, project policies, knowledge trust, retention, and billing limits decide how this behaves for a real workspace.
Proof points
Designed for governed support, not hidden autonomy.
AI outputs are schema-validated.
Examples are replayable.
Policy regressions can be caught before release.
References
Official stack and legal references.
These external links point to provider documentation, privacy notices, data-processing terms, or platform terms that shape production operation.
Next