DAX. Enterprise Technical Whitepaper · 2025
Download PDF Try Pilot →
Technical Whitepaper

Enterprise AI Governance
Without Compromise

The case for zero-knowledge BYOC architecture as the only governance model that doesn’t trade data sovereignty for visibility.

DAX AI, Inc.June 2025v1.0
Download PDF (239 KB)

The Problem

Every major SaaS AI governance platform today shares a common architecture: your prompts flow through their servers, your completions are processed on their infrastructure, and your employees’ AI interactions are stored in their databases. In exchange, you receive dashboards, policy enforcement, and audit trails.

This is the governance paradox: the tools built to protect your data require you to surrender it.

The SaaS Governance Paradox: You cannot govern AI data sovereignty by sending your data to a third-party governance vendor. The solution to a data risk cannot be another data risk.

Consider what governance actually requires: tracking which models were called, by whom, with what policies applied, and what the outcomes were. None of that requires seeing prompt content. Token counts, cost metrics, latency, and status codes are sufficient for comprehensive governance — and they carry zero confidential information.

The regulatory pressure

The EU AI Act (enforcement begins February 2026 for high-risk systems), NIST AI RMF, and ISO/IEC 42001 all impose requirements on organizations deploying AI in production. Regulated industries face additional obligations: HIPAA for patient data, GDPR for EU personal data, and sector-specific guidance from regulators still writing the rules.

These requirements create a dilemma: enterprises need governance tooling, but the tooling itself must not create new compliance exposure. A HIPAA-covered entity cannot route clinical prompts through a SaaS vendor’s servers, even with a BAA.

Zero-Knowledge Architecture

DAX Enterprise is built on a single architectural principle: the governance layer should never see the governed data.

This is achievable because inference governance can be decomposed into two planes with different data requirements:

Control plane (governance metadata only)

The control plane manages policies, identities, approvals, audit logs, and budgets. It receives usage events containing: model ID, user ID, token counts, cost estimate, latency, HTTP status code, and timestamp. It never receives prompt content, completion text, or any user-generated content.

Data plane (inference only)

The data plane is an OpenAI-compatible inference proxy that runs in the customer’s GCP project exclusively. It validates access tokens, enforces rate limits, applies guardrail policies, and proxies requests to the configured provider.

JWT offline validation: The gateway validates tokens using a public key fetched once at startup. The control plane is never contacted during inference — no per-request callback, no session lookup, no hot-path dependency.

The zero-knowledge guarantee

The guarantee is architectural, not contractual. The control plane has no endpoint that accepts prompt content. The gateway has no code path that transmits prompts to the control plane. The usage event structure contains only numeric and identifier fields. This is not a privacy policy — it is a code constraint.

BYOC Design

Bring-Your-Own-Cloud means both planes run in the customer’s GCP project. DAX operates no infrastructure on the customer’s behalf and has no access to their data, network, or credentials.

ComponentLocationDAX Access
Control plane APICustomer Cloud RunNone
Portal (Next.js)Customer Cloud RunNone
Inference gatewayCustomer Cloud Run / GKENone
PostgreSQLCustomer Cloud SQLNone
RedisCustomer MemorystoreNone
Provider API keysCustomer Secret ManagerNone
Prompt / completion dataNever storedN/A

Workload Identity for Vertex AI

When routing inference to Google Vertex AI, the gateway uses Workload Identity — the GCP service account bound to Cloud Run is granted roles/aiplatform.user directly. No API key is needed and there is no secret rotation burden for Vertex models.

Compliance Coverage

DAX maps natively to the frameworks enterprise buyers, legal teams, and auditors actually require. No custom integrations, no compliance spreadsheets — every control is a built-in platform feature.

StandardRequirementHow DAX Covers It
NIST AI RMFGOVERN — AI policiesNamed guardrail policies, RBAC, audit log evidence trail
NIST AI RMFMAP — Risk identificationModel catalog with risk_tier, autonomy_tier, use-case context per access request
NIST AI RMFMEASURE — AnalyticsUsage events across org/model/user/team/project + incident severity metrics
NIST AI RMFMANAGE — ResponseGuardrail engine, budget hard blocks, incident lifecycle (open → resolved)
EU AI ActArt. 6 — Risk classificationrisk_tier + eu_ai_act_category on every model, compliance coverage dashboard
EU AI ActArt. 14 — Human oversightautonomy_tier enforcement — autonomous models require guardrail policy
EU AI ActArt. 13 — Transparencyexplainability_doc_url + bias_assessment_url fields, coverage metrics
EU AI ActArt. 73 — Incident reportingIncident log with severity lifecycle, auto-creation from guardrail violations
ISO/IEC 42001AI system inventoryModel catalog with lifecycle status, provider, version, governance metadata
ISO/IEC 42001Shared responsibilityBuilt-in printable shared responsibility matrix, documented BYOC boundary
SOC 2 Type IIAccess controlRBAC, model approval workflow, token scoping — least-privilege by default
SOC 2 Type IIAudit loggingAppend-only audit log, 25+ event types, CSV export for external auditors
GDPR / HIPAAData sovereigntyBYOC: prompts never leave customer cloud — architectural, not contractual

EU AI Act Alignment

The EU AI Act imposes tiered obligations based on risk level, with enforcement for high-risk systems beginning February 2026.

Risk classification (Articles 6, 51)

DAX implements four-tier risk classification on every model: high, medium, low, and minimal. Each model also carries an eu_ai_act_category field. The compliance dashboard tracks classification coverage.

Human oversight (Article 14)

The autonomy tier field classifies each model as Assistive (HITL), Conditional (HOTL), or Autonomous (HOOTL). Autonomous models require an assigned guardrail policy before activation. The approval workflow satisfies Article 14’s oversight requirements.

Transparency (Article 13)

The explainability_doc_url field links each model to its decision-scope documentation. The compliance dashboard measures coverage and reports a pass/fail check.

Incident reporting (Article 73)

The incident log provides the workflow required for serious incident reporting. Guardrail blocks above a configured severity threshold automatically create incidents, flowing through open→investigating→resolved with timestamped resolution notes.

NIST AI RMF Alignment

The NIST AI RMF organizes AI governance into four functions: Govern, Map, Measure, and Manage. DAX implements controls across all four.

GOVERN

RBAC with five built-in roles plus custom role definitions satisfies accountability requirements. Named guardrail policies serve as org-wide AI usage policies. The audit log provides the accountability evidence trail.

MAP

The model catalog with risk tier, autonomy tier, EU AI Act category, and use-case context captured via justification and intended_use on access requests satisfies MAP requirements.

MEASURE

Usage events with org/model/user/team/project dimensions give the measurement data MEASURE requires. The incident log with severity classification provides incident metrics.

MANAGE

The guardrail engine implements risk treatment controls. Budget enforcement with hard blocks prevents runaway spend. The incident lifecycle implements response and recovery processes.

Shared Responsibility Model

AreaDAXCustomer
Governance engine codeProvides + maintainsDeploys + operates
Inference gatewayProvides + maintainsDeploys in own GCP
Prompt / completion dataNever receivesSolely responsible
Provider API keysSecret Manager integrationCreates + manages secrets
Model risk classificationProvides fields + dashboardClassifies each model
Guardrail policy contentProvides engine + templatesConfigures policies
SSO / identityProvides SAML frameworkManages IdP + enforces MFA
GCP IAM + VPCNo accessConfigures + manages
Formal certificationReadiness toolingEngages auditors

The critical distinction is in the prompt/completion row: DAX has no access because there is nothing to access. The gateway never sends this data to the control plane. This is not a contractual promise — it is an architectural constraint enforced in code.

Getting Started

Option 1: Self-hosted (Community, free)

git clone https://github.com/dax-ai/dax-enterprise
cd dax-enterprise
docker compose up --build
# Portal → http://localhost:3001
# API docs → http://localhost:8090/docs

Option 2: Live pilot (hosted by DAX)

The full platform is running at app.enterprise.dax.studio. Sign up with your work email — no credit card, no sales call.

Option 3: Deploy to your GCP project

cd data-plane/infra/gcp
terraform init
terraform apply -var="project_id=YOUR_GCP_PROJECT"
Design partner program: The first 10 enterprise customers receive Pro tier at no cost for 6 months in exchange for direct product feedback. Contact hello@dax.studio.