Privacy, Compliance & Enterprise

Enterprise AI is a trust product

Selling AI to enterprises means answering security questionnaires about data flow, not demoing clever prompts. Engineering implements what legal promises in the MSA.

Data flow boundaries

Document clearly:

• What user data enters prompts
• Which subprocessors (OpenAI, Anthropic, embedding vendors)
• What is logged, retained, and for how long
• Whether data trains vendor models (API tier vs consumer tier differs)

Zero retention and no training clauses — verify in contract, not marketing page.

Regional residency

Requirements:

• EU data stays in EU (GDPR)
• Some nations require in-country inference

Options:

• Regional vendor endpoints
• Self-hosted models in-region
• Metadata-only cross-border (never move raw PII)

Architecture decision, not checkbox — wrong region = deal blocker.

Compliance frameworks

Framework	AI-specific concern
SOC 2	Access control, audit logs, change management for prompts
HIPAA	PHI in prompts — BAA with vendor or no cloud LLM
GDPR	Right to erasure includes vectors derived from user docs
PCI	Never put card numbers in prompts — tokenize first

OWASP LLM Top 10: sensitive disclosure, supply chain risks — map controls to each.

Customer data isolation

Multi-tenant products need:

• Logical isolation (namespace, encryption per tenant)
• No cross-tenant retrieval (metadata filters + tests)
• Per-tenant audit export for compliance reviews

Pen-test retrieval ACLs — #1 enterprise fear is "their data in my answer."

Audit and explainability

Enterprise buyers ask:

• Who asked what, when, which model answered
• Was retrieval used? Which documents?
• Human override trail for regulated decisions

Immutable audit log (WORM storage) for some industries.

Vendor diligence

Before adding a subprocessor:

• Security whitepaper, pen-test summary
• Data processing agreement
• Incident notification SLA
• Model change notification policy

Employee access

Production prompts and logs contain customer secrets. Restrict:

• PII scrubbed from internal debug tools
• Break-glass access logged
• No prod data on laptops for LLM fine-tuning experiments

Interview framing

"Design AI for a bank's internal knowledge base":

1. VPC/private link to vendor or on-prem inference
2. No training on customer data; zero retention API
3. PII scanner on ingest and egress
4. Full audit trail; EU tenant data in EU region
5. Human review for compliance-tagged queries

Senior signal: Mention right to deletion — user leaves company → purge their docs from vector index and caches, not just OLTP rows.

Track completion

You've covered the production AI systems loop: foundations → architecture → serving → knowledge → safety → eval → cost → shipping. Cross-link to sibling tracks as you go deeper on any pillar.