3.6 KiB
Saudi compliance & AI governance register (design-time)
Not legal advice. This is an engineering readiness register for building Dealix as a Tier-1 operating system in KSA/GCC. Legal review remains required for production claims and customer contracts.
Canonical trust model: trust-fabric.md. Product legal texts: salesflow-saas/docs/legal/.
1. PDPL / personal data (design checklist)
When processing data that may identify individuals in the Kingdom:
- Inventory data categories, purposes, lawful basis, retention, subprocessors, and cross-border transfers (if any).
- Minimize collection; default deny for exports and bulk analytics on personal fields.
- Consent and notices aligned with product copy (
salesflow-saas/docs/legal/consent-policy-ar.md, privacy / data protection docs). - AI-specific: training, enrichment, search, scoring, messaging, and logs can all be processing — classify sensitivity (S0–S3) per approval-policy.md and route S2/S3 away from unreviewed third-party models/tools.
- Subject rights / export: define operational runbooks before offering enterprise SLAs.
References (external): Saudi PDPL / SDAIA knowledge center and official guidance — verify current text with counsel.
2. NCA cybersecurity posture (readiness, not certification)
Design so the platform can align with ECC and related cloud/data controls (DCC, CCC) as the customer tier demands:
- Asset inventory, patch cadence, access control, logging, incident response hooks.
- Segregation of prod/staging; break-glass for admin; audit streaming for long retention (pair with github-and-release.md audit notes).
References (external): NCA published controls and updates (e.g. ECC 2-2024 track) — map controls to features in an ADR when pursuing attestation.
3. AI governance (NIST + OWASP)
Use as a risk and testing frame for agentic features:
| Frame | Use in Dealix |
|---|---|
| NIST AI RMF | Govern, map, measure, manage — tie to release gates and evidence packs |
| NIST Generative AI profile | Supplement for LLM-specific risks |
| OWASP Top 10 for LLM Apps | Prompt injection, insecure output handling, excessive agency, sensitive disclosure — explicit test cases in CI where feasible |
Pair with trust-fabric.md: red-team workflows, structured output validation, tool allowlists, and rollback plans for Class B / R2+.
References (external): NIST publications portal; OWASP LLM Top 10 and GenAI security project pages.
4. Arabic-first execution (product, not theme)
Beyond RTL UI:
- Arabic classification and summaries for internal notes where policy allows.
- Partner memos and notification templates with terminology normalization (sector-specific).
- Retrieval quality for Arabic queries (embedding model + chunking + evaluation).
- Trust cues in UX (support, compliance, local expectations).
See design-and-arabic.md.
5. Review cadence
- Quarterly: re-read this register against shipped features and incident postmortems.
- Per major release: update PDPL/NCA mapping appendix when product surface area changes.
See also: technology-radar-tier1.md, ../execution-matrix-90d-tier1.md, pdpl-nca-ai-control-matrices.md.