AI Model Preview and Staged Release Policy
Establish an internal policy that distinguishes preview and experimental AI system access from approved production deployment, and requires documented governance sign-off at each release stage before a system advances to broader use.
Objective
Prevent AI systems from entering production without adequate governance review by defining release stages with distinct access controls, evaluation requirements, and approval authorities for each transition.
Maturity Levels
Initial
There is no distinction between experimental AI access and approved production use. Individual teams may deploy AI systems acquired during a preview program without governance review.
Developing
The organization informally distinguishes between pilot and production AI deployments, but the criteria for transitioning between stages are not defined and approval is informal.
Defined
A documented staged release policy defines at least two release stages (e.g., experimental/preview and production-approved), with explicit criteria and required approvals for each stage transition. Experimental deployments are tracked in the AI inventory with a status flag indicating non-production status.
Managed
The staged release policy is enforced through the intake workflow (MGV-002). Stage transitions are logged and auditable. Preview deployments that have not transitioned after a defined period are flagged for review or retirement.
Optimizing
Release stage criteria are reviewed annually against current model capabilities and regulatory requirements. The policy covers new modality types (agentic, multimodal, ambient) as they enter the environment. Preview program participation is governed to prevent capability access that has not been risk-assessed.
Evidence Requirements
What an auditor or assessor would expect to see for this control.
- —Staged release policy document defining release stages, permitted use at each stage, data handling restrictions, and transition criteria.
- —AI inventory records showing experimental vs. production-approved status for all tracked AI deployments.
- —Stage transition approval records for deployments that advanced from experimental to production-approved in the past 24 months.
- —Evidence of periodic review of experimental deployments to confirm they are either transitioning or being retired.
Implementation Notes
Why preview and production are governance-distinct states
AI vendors, including major frontier model providers, routinely offer access to new model capabilities under preview or beta programs before general availability. These programs have practical value: they allow organizations to evaluate and prepare for capabilities before they are fully released. But they also create a governance gap.
Preview models are often assessed under less rigorous safety evaluation than GA releases. Capabilities may be incomplete, inconsistent, or subject to change without notice. The vendor may not yet have completed its own red-team cycle for preview features. System prompt behavior, tool-use characteristics, and safety filter coverage may differ substantially from what is documented in the GA model card.
When organizations use preview access for exploratory development but allow those deployments to continue in production after GA release — or when preview deployments are never formally transitioned and reviewed — they inherit risk from the preview model's lower assurance level. The governance failure is not the use of preview access; it is the absence of a defined transition gate.
Defining release stages
A minimal staged release policy requires at least two stages:
Experimental / Preview: Access to AI capabilities under a vendor preview program, or internal AI systems under active development. Permitted use: internal testing, evaluation, non-production workflows. Data handling: no regulated or production data unless explicitly approved. Oversight: team-level, with AI inventory flag. Not subject to full governance review.
Production-Approved: AI systems that have completed the intake and approval workflow (MGV-002), passed required evaluations, received governance sign-off, and are authorized for use with production data and business processes. Oversight: per approved deployment parameters.
Larger organizations may add intermediate stages (e.g., limited production, phased rollout) with corresponding controls.
Stage transition requirements
For a deployment to transition from Experimental to Production-Approved:
- Complete AI system intake (MGV-002) if not already done.
- Complete model evaluation appropriate to the use case (PRC-003 for external models; MGV-006 for RAI benchmark alignment).
- Confirm the GA model being deployed matches the preview model used in development, or document any differences and re-evaluate as needed.
- Update the AI inventory entry from experimental to production-approved status.
- Obtain sign-off from the AI governance committee or designated approval authority.
What this policy covers beyond typical change management
This control is distinct from the model deployment gate (CHM-002), which governs updates to already-approved production systems. Staged release policy governs the initial deployment path: when a system transitions from evaluation to production for the first time. Both controls are needed: CHM-002 handles what happens after production; MGV-001 handles what happens before.
Example Implementation
AI Release Stage Policy — Summary Table
| Stage | Description | Permitted use | Data restrictions | Approval authority | Max duration |
|---|---|---|---|---|---|
| Experimental | Preview/beta access; internal dev/eval only | Evaluation, non-production testing, PoC | No regulated data; no production data without explicit exception | Team lead | 90 days; extension requires AI governance committee approval |
| Limited production | Production use with restricted user population or use case scope | Defined scope only; production data within approved scope | Per intake approval | AI governance committee | 180 days; full production approval required to extend |
| Production-approved | Fully reviewed and approved for production use | All uses within approved scope | Per intake approval | AI governance committee | Ongoing; subject to periodic re-review and change management |
Stage Transition Log (excerpt)
| System | Previous stage | New stage | Transition date | Approval | Notes |
|---|---|---|---|---|---|
| Contract summarization assistant | Experimental | Limited production | 2026-03-15 | AI Governance Committee | Approved for legal team use only; full production pending DPIA completion |
| Procurement risk classifier | Limited production | Production-approved | 2026-04-28 | AI Governance Committee | DPIA complete; usage extended to all procurement workflows |
| Competitor analysis agent | Experimental | Retired | 2026-05-01 | Team lead | Use case deprioritized; experimental access terminated |