AI and Agent Governance Framework

"Gate" in this framework always means a lifecycle gate - a decision point a client's agent passes through.

1. Purpose and applicability

This framework answers one operating question:

Who may approve, change, operate, suspend and retire an agent, through which lifecycle gates, with which registry record and retained evidence?

It applies to a production AI system that includes an agent, from intake through retirement. It may also be used for a non-agentic AI system where the same decision rights are useful. The client maps its job titles and committees onto the functional accountabilities defined here.

NIST describes AI risk management as a continuous lifecycle activity organized through GOVERN, MAP, MEASURE and MANAGE. GOVERN is cross-cutting and informs the other three functions; NIST also states that documentation can improve transparency, human review and accountability. [G01] [G02] The NIST Playbook contains voluntary suggestions rather than a checklist that every organization must follow. [G03] This document therefore uses NIST as a source basis but presents an Alpha AI recommended operating method, not a NIST conformance claim.

NIST published the Generative AI Profile as a companion to AI RMF 1.0. [G04] NIST states that the Profile can help organizations identify risks specific to generative AI and select actions aligned with their goals and priorities. [G01] ISO describes ISO/IEC 42001 as an AI management-system standard using Plan-Do-Check-Act to govern organizational AI risks and opportunities. [G05] The OECD AI Principles address trustworthy AI across the AI lifecycle and were updated in May 2024. [G06]

Source and assurance limits

• The licensed ISO/IEC 42001 text was not consulted. [G05]
• This document makes no clause-level ISO statement, mapping, alignment or conformance claim.
• It is not legal advice and does not determine which law or regulation applies.
• Detailed threats, controls, regulatory mappings, compliance reports and audit packs are addressed in separate companion frameworks.
• Client-specific job titles, risk appetite, legal obligations and approval thresholds remain deployment-brief decisions.

2. Governance principles

The following are Alpha AI recommended design principles, informed by NIST's treatment of governance as continuous, cross-cutting and lifecycle-wide. [G01] [G02]

1. One accountable owner. Every production agent has one accountable business owner, even when several teams build or operate it.
2. Authority follows impact. Approval authority and evidence depth increase with the agent's potential impact and ability to act.
3. No unrecorded production agent. A production deployment requires an active registry record and current owner.
4. Evidence before approval. A gate passes only when its exit condition is met and the named evidence is retained.
5. Change is part of the lifecycle. A material change returns the agent to the applicable earlier gate.
6. Exceptions expire. Every exception has an owner, compensating measures, an expiry date and a defined closure test.
7. Stop authority is explicit. Operations and risk functions may suspend an agent within their defined authority without waiting for a routine governance meeting.
8. Unknowns remain visible. An unresolved legal, technical or evidence question is recorded as an open dependency, not converted into an assumption.

3. Functional accountabilities and decision rights

These are accountabilities, not job titles. One person may hold several roles for a lower-tier agent if conflicts are documented. For Tier 3 and Tier 4, the approver and independent assurance role must be different people.

Decision-rights matrix

A means final accountability, R prepares or executes, C must be consulted, I must be informed and V must record a positive independent validation before the decision passes. The client records named people against these functions in the agent registry.

4. Risk-tier interface

NIST states that MAP establishes context for risk decisions, that MEASURE supplies documented testing and monitoring, and that MANAGE prioritizes and treats mapped and measured risks. [G02] This framework defines a four-tier interface so GOV1 can operate before GOV3 supplies the detailed control checklist and GOV4 supplies jurisdiction-specific regulatory mappings.

These tiers are an Alpha AI governance method. They are not legal classifications and must not be labelled with a jurisdiction's statutory risk categories.

Tier assignment rules

The product owner prepares an initial tier proposal. The business owner is accountable for the proposal, and independent assurance challenges Tier 2-4 classifications. The assigned tier is the highest tier triggered by the agent's data, users, decisions, actions, scale or failure impact.

Escalate at least one tier, or record why escalation is unnecessary, when the agent:

• gains a new tool, action, data class, user population or external communication path;
• moves from recommendation to execution;
• removes or weakens human review;
• changes deployment context, model, memory behavior or operating scale in a way that could change impact;
• experiences a serious incident or repeated control failure; or
• reaches a use case for which regulatory applicability has not been determined.

Open GOV4 dependency: GOV4 must define the regulatory applicability assessment and the jurisdiction-specific triggers that can set a tier floor, require additional evidence or prohibit deployment. Until that record exists, unresolved applicability is Tier 4 and cannot be downgraded by the product or technical owner.

5. Lifecycle gates

NIST treats AI risk management as iterative across the lifecycle rather than a one-time approval. [G01] [G02] The eight gates below implement that concept as an Alpha AI recommended operating process. Every gate has a binary exit condition, a retained artifact and a role accountable for retention.

Material-change rule

A change is material when it could alter purpose, users, affected parties, data, decision rights, action scope, risk tier, acceptance criteria or the evidence supporting an approval. The product owner records the classification and routes a material change to the earliest gate whose evidence is no longer valid. Independent assurance may override a non-material classification for Tier 2-4.

6. Agent and model registry

The registry is the authoritative governance record, not necessarily the runtime's technical inventory. NIST describes documentation as supporting transparency, human review and accountability, and its GOVERN function includes processes and organizational schemes across the lifecycle. [G02]

Minimum record

Each production agent record contains:

• stable agent ID, name, status and environment;
• intended purpose, prohibited purposes and business workflow;
• accountable business owner, product owner, technical owner, operations owner, assurance owner, data owner and regulatory applicability owner;
• risk tier, rationale, assignment date and next review date;
• affected users or parties and human-decision points;
• data classes, sources, memory/persistence description and approved retention reference;
• model, prompt, tool, runtime and dependency version references;
• action boundary, approval boundary, stop conditions and kill-mechanism owner;
• acceptance criteria and links to current validation evidence;
• lifecycle gate status, approvals and evidence links;
• open findings, exceptions, expiry dates and compensating measures;
• incidents, material changes and last periodic review;
• regulatory applicability status and GOV4 mapping link;
• suspension, restart and retirement state.

The model record may be separate, but the agent record must resolve to the exact approved model and configuration version. A shared model does not replace an agent record because the agent's purpose, tools, data and decision rights can differ.

Attestation and record quality

At each scheduled review, the business owner attests continued need and ownership; the product owner attests record accuracy; operations attests current operating evidence; and assurance records unresolved findings. Missing ownership, an expired approval or an overdue Tier 3-4 review blocks new material changes and triggers a suspension decision.

7. Exceptions and risk acceptance

An exception does not silently change a gate. It temporarily records why a requirement cannot be met and how exposure is constrained.

Every exception record includes:

• agent ID, tier, affected requirement and reason;
• impact and time-bounded risk statement;
• compensating measures and their evidence;
• accountable owner and approving authority;
• start date, expiry date and closure test;
• monitoring trigger and immediate suspension condition;
• related incident, change and regulatory applicability records.

The approving authority is the same authority required for production at that tier, except Tier 4 and expired exceptions always require the governance authority. An exception cannot waive an applicable law; GOV4 determines the applicability interface. On expiry, the agent is suspended unless the requirement is met or a new exception is approved using current evidence.

8. Management reporting and evidence handoffs

The OECD Principles frame accountability across the AI lifecycle, while NIST describes MANAGE as regular prioritization, treatment, monitoring and improvement based on mapped and measured risk. [G02] [G06] GOV1 therefore defines the management questions and handoffs, while GOV5 defines the report and GOV6 defines the audit evidence pack.

The governance authority receives, at least quarterly:

• active agents by tier, owner, status and overdue review;
• gate approvals, suspensions, restarts and retirements;
• material changes and tier changes;
• open findings and exceptions by age and expiry;
• incidents, repeated failures and overdue corrective actions;
• unresolved regulatory applicability determinations;
• missing or failed evidence collection;
• decisions required and named accountable owners.

Handoffs to later deliverables:

9. FDE adoption sequence

1. Map client titles and committees to the functional roles; record conflicts.
2. Agree tier thresholds, risk appetite and delegated approval limits.
3. Configure the registry schema and migrate known production agents.
4. Select one representative agent in each active tier and reconstruct its gate evidence.
5. Test suspension, restart, exception expiry and material-change routing.
6. Assign the GOV4 regulatory applicability owner and open unresolved determinations.
7. Start the tier-based review cadence and management decision log.
8. Transfer registry administration, evidence ownership and gate operation to the client.

10. Reusable checklists

Governance setup

• Functional roles have named people and delegated authority.
• Conflicts and required separation of duties are recorded.
• Tier definitions, thresholds and review cadence are approved.
• Registry custodian and authoritative system are named.
• Regulatory applicability owner and GOV4 dependency are named.
• Emergency suspension and restart authority are documented.
• Exception expiry produces an enforceable action.

Gate decision record

• Agent ID, version, environment and tier
• Gate and decision date
• Exit condition result
• Evidence artifact links and versions
• Open findings and exception links
• Decision: pass or fail
• Approver functions and named people
• Conditions, owners and due dates
• Next gate or route-back gate
• Next review date

Periodic governance review

• Ownership and business need remain current.
• Tier and regulatory applicability remain current.
• Gate evidence still describes the deployed version.
• Material changes were routed correctly.
• Incidents, findings and repeated failures were resolved or escalated.
• Exceptions are current and compensating measures still work.
• Acceptance criteria and monitoring remain sufficient.
• Suspension, restart and retirement paths remain operable.

References

All references resolve to sources/SOURCES.md.

• [G01] NIST, Artificial Intelligence Risk Management Framework 1.0.
• [G02] NIST AI Resource Center, AI RMF Core.
• [G03] NIST AI Resource Center, NIST AI RMF Playbook.
• [G04] NIST, AI RMF: Generative Artificial Intelligence Profile.
• [G05] ISO, ISO/IEC 42001:2023 official record.
• [G06] OECD.AI, OECD AI Principles.