AI Security & Data Handling Brief
A plain-English summary of how Northstar Relay AI’s support-triage product uses models, handles customer data, and manages AI-specific risks. This sample deliberately preserves open gaps instead of presenting unverified claims as facts.
01 · Product boundary
Support Relay
Support Relay classifies inbound B2B support requests and drafts a suggested response for an authorized customer-support agent. It does not send messages, change account settings, issue refunds, or make eligibility decisions without a human action.
In scope: ticket text, approved help-center articles, model prompts and responses, application logs, and human approval. Out of scope: autonomous customer communication and production account changes.
02 · AI data flow
03 · Model & provider statement
| Provider | Purpose | Customer data | Training use | Retention |
|---|---|---|---|---|
| Example Model API | Ticket classification and draft generation | Ticket text and selected help-center content | Contract says API data is not used to train general models | 30-day provider abuse monitoring; zero-retention eligibility under review |
04 · Control statements & evidence
Human approval before customer communication
Draft responses remain in the agent workspace until an authenticated support user edits or sends them.
Retrieved-content boundary
Retrieval is limited to articles approved for the customer’s tenant; retrieved text is treated as untrusted content.
Provider zero-retention configuration
Northstar plans to request provider approval for zero data retention. Current buyer statements use the documented 30-day provider window until approval is verified.
Quarterly prompt-injection regression coverage
Ad hoc tests exist, but the team has not verified a repeatable quarterly test suite. No recurring-test claim should be made.
05 · Buyer answer examples
Does customer data train your models?
Northstar does not train its own general-purpose models on customer data. Its documented enterprise API terms state that API inputs and outputs are not used to train the provider’s general models. Provider and feature scope should be confirmed for the proposed deployment.
Can the AI take actions autonomously?
No. The in-scope product creates a draft for an authenticated human support agent. Sending, account changes, refunds, and eligibility decisions require separate human actions.
How do you prevent prompt injection?
Northstar restricts retrieval by tenant, treats retrieved text as untrusted, limits tool access, and requires human approval before a response is sent. A recurring regression-test cadence is not yet verified and is tracked as an open gap.