Trust Center

1. Executive Summary

Prepp is an AI-powered sales onboarding platform that helps sales teams reach competency faster through guided practice, roleplay, and manager-visible progress insights.

2. Product Overview

What Prepp Does

• Content onboarding: Admins upload approved product playbooks and training materials
• AI coaching: Sales reps complete guided practice and roleplays with an AI coach
• Manager insights: Managers review aggregated progress and coaching outcomes

Key Workflows

Admin uploads content → Rep practices with AI coach → Manager reviews progress

3. Architecture & Data Flow

High-Level Architecture

Infrastructure Components

Data Residency

• Primary region: GCP Israel (me-west1)
• Data stored in Israel: Application data, user profiles, coaching logs, recordings
• Data leaving Israel: Beyond core hosting in Israel, certain features process content in other regions or subprocessor regions as configured:
  • Ongoing coaching / external call import: Call audio may be sent to ElevenLabs for speech-to-text transcription. Transcripts and related context (seller names, companies, counterparty fields, and similar metadata supplied by integrations or admins) may be sent to LLM providers (OpenAI, Anthropic, Google Gemini, etc.) and to the agent-service (application logic for scoring, coaching, and related workflows). These flows are not limited to “generic prompts”; they can include real call content where customers enable those features.
  • LLM features: Text prompts and model responses are typically processed in provider regions (often the United States) according to each provider’s terms and DPA.
  • Optional tracing: When LangSmith (or similar) tracing is enabled in agent-service configuration, trace metadata and LLM inputs/outputs (as configured) may be sent to that subprocessor for observability.

For a technical walkthrough of the ongoing coaching import pipeline, see docs/ongoing-coaching/call-import-pipeline-ELI5.md.

4. Authentication & Access Control

Identity & SSO

Role-Based Access Control (RBAC)

Internal Service-to-Service Authentication

Internal communication is authenticated at two layers:

• Backend webhooks (worker -> backend): A single shared secret (BACKEND_INTERNAL_WEBHOOK_SECRET) is sent in the X-Prepp-Internal-Webhook-Secret header and verified by the InternalWebhookSecretGuard on all webhook controllers (call-import, coaching-snapshot, worker progress). On GCP Cloud Run, requests additionally carry Google-issued OIDC identity tokens.
• Agent-service: Callers send either a Google OIDC ID token (verified against AGENT_SERVICE_AUDIENCE on Cloud Run) or a shared AGENT_SERVICE_API_KEY as a Bearer token for local/dev environments.

Audit Logging

• Sign-in events
• Administrative actions
• Content changes
• Organization-scope mismatch warnings on call-review webhook payloads
• Export available on request

5. Data Handling

Data We Collect

Sensitive Data Types We Do Not Require as Standalone Onboarding Fields

• ❌ Government identifiers (social security, national ID numbers, etc.) as a required product field
• ❌ Sensitive customer records outside normal business use of the product
• ❌ Call recordings (unless explicitly enabled)
• ❌ Integration with CRM/HR systems (unless explicitly connected)

Calls and transcripts: Sales calls and coaching workflows may naturally include personally identifiable information (names, companies, phone numbers, and other content spoken or written in transcripts). Ongoing coaching and call import features process that content as part of the service. Processing is governed by the DPA and subprocessor disclosures in this document and the customer agreement—not by an assumption that calls are “generic” or PII-free.

Tenant Isolation

Multi-tenant SaaS with logical tenant separation. All records are scoped by organization_id with org-scoped queries enforced at the service layer. Ongoing-coaching review results derive their organization scope from the stored call record (database-authoritative), not from the inbound webhook payload; mismatches are logged and the payload scope is ignored.

6. Encryption

In Transit

At Rest

Key Management

• Encryption keys managed by GCP
• Customer-managed encryption keys (CMEK) available on request

7. Security Controls

Network Security

• Cloud Run services deployed in GCP Israel region
• Private VPC connector for database access
• HTTPS-only ingress with TLS 1.2+
• Hatchet workflow engine runs on an internal-only GCE VM with firewall rules restricting access to the VPC connector IP range
• Hatchet dashboard access via HTTPS load balancer with Google Identity-Aware Proxy (IAP); only authorized Google accounts can reach the UI
• Cloud Armor available for WAF/DDoS protection on request

SSRF Protection

• Outbound audio-fetch requests are validated before execution: URL scheme (HTTPS required), port restriction, DNS resolution checked against private/reserved IP ranges (RFC 1918, link-local, loopback, metadata endpoint), and redirect targets re-validated per hop (max 3 redirects)
• GCP metadata endpoint (metadata.google.internal) is explicitly blocked

Secure Development Lifecycle (SDLC)

• Code review required for all changes (PR-based workflow)
• CI pipeline with automated tests and linting
• Dependency scanning via Dependabot
• Secret scanning in CI
• Staging environment for pre-production validation

Debug & Documentation Exposure

• Agent-service /docs, /redoc, and debug/test routes are disabled in staging and production environments by default
• Controlled via ENVIRONMENT env var; opt-in override requires explicit ENABLE_AGENT_DEBUG_ROUTES flag

LLM Trace Privacy

• Sensitive ongoing-coaching LLM invocations (call evaluation, coaching snapshots, seller context) suppress content tracing to LangSmith in staging and production
• Opt-in re-enablement via ENABLE_LANGSMITH_CONTENT_TRACING for authorized debugging

Logging & Monitoring

• Application logs in GCP Cloud Logging
• Audit logs for sign-ins, admin actions, content changes
• Organization-scope mismatch detection and logging on webhook payloads
• Alerting via GCP Cloud Monitoring
• SIEM integration available on request (log export supported)

Backup & Recovery Testing

• Automated daily backups for Cloud SQL
• Quarterly restore testing schedule
• RPO 24 hours; RTO 48 hours

8. AI Governance

LLM Providers

Data Usage Policy

LLM Data Retention (Provider Side)

Note: OpenAI explicitly states: "Data sent to the OpenAI API is not used to train or improve OpenAI models."

9. Data Retention & Deletion

Default Retention

Configurable Options

• Retention windows adjustable per customer contract
• Ongoing-coaching artifact retention configurable per deployment (default 90 days)
• Opt-out of storing prompts/responses available
• Voice recordings can be disabled entirely

Deletion Process

10. Security Operations

Vulnerability Management

Detection Methods

• Weekly dependency scans (CI)
• Dependabot alerts and automated PRs
• Cloud provider security alerts
• Manual review of major advisories

Penetration Testing

Incident Response

Severity Levels

Security Contact

• Email: security@prepp.tech
• Response: Within 24 hours

11. Business Continuity

Backup Strategy

Recovery Objectives

Disaster Recovery

• Zonal Cloud SQL in GCP Israel with automated backups + PITR
• HA/failover available when regional Cloud SQL is enabled
• Quarterly restore testing

Service Availability

12. Compliance Status

Current Status

Compliance Roadmap

Israel Privacy Law (חוק הגנת הפרטיות)

13. Subprocessors

14. Contact Information

This document consolidates Prepp's security and compliance information for enterprise due diligence. For additional questions, contact security@prepp.tech.