Principal-led architecture for critical systems

Resource

AI System Intake & Release Decision Record

A reusable record for registering an AI system, defining its authority and dependency boundary, capturing evaluation evidence, and recording a bounded release, pause, rollback, or retirement decision.

Decision recordOne current record from intake through retirement
Resource type
Review worksheet
Version
1.48.0
Decision
What exactly is this AI system, who owns its consequences, what evidence supports the current state, and under what conditions may it proceed, pause, roll back, or retire?

Purpose and use

AI use cases, agents, models, vendors, and embedded features can enter review without one current record of ownership, intended use, dependencies, human authority, risk, evidence, release conditions, and retirement.

Who should use it

Business owners, technical owners, risk, security, privacy, legal, accessibility, procurement, operations, and delivery stakeholders who need one shared record of the reviewed system state.

Decision supported

What exactly is this AI system, who owns its consequences, what evidence supports the current state, and under what conditions may it proceed, pause, roll back, or retire?

Lifecycle note: Treat the record as versioned evidence. Update it when the use, users, model, vendor, data, tools, authority, environment, risk, evaluation evidence, or operating conditions materially change.

Identity and ownership

Record the current answer, owner, supporting evidence, unknowns, and next review trigger. Use the organization's own risk language rather than treating this worksheet as a universal scoring model.

  • System, use-case, agent, model, vendor, and business-process name
  • Business owner, technical owner, risk or policy owner, and release authority
  • Lifecycle state, review date, next review trigger, and retained record location
  • Intended users, affected people, environments, and organizational boundary

Intended use and prohibited use

Record the current answer, owner, supporting evidence, unknowns, and next review trigger. Use the organization's own risk language rather than treating this worksheet as a universal scoring model.

  • Business problem, expected benefit, and decision the system supports
  • Approved context, user population, geography, and operating conditions
  • Explicit prohibited uses, unsupported decisions, and out-of-scope populations
  • Known assumptions, unknowns, and conditions that invalidate the current approval

System, data, and dependency boundary

Record the current answer, owner, supporting evidence, unknowns, and next review trigger. Use the organization's own risk language rather than treating this worksheet as a universal scoring model.

  • Models, prompts, retrieval, agents, tools, APIs, vendors, and downstream actions
  • Data sources, classifications, trust states, access controls, retention, and deletion
  • External dependencies, failure modes, fallback paths, and concentration risk
  • Versioned configuration and change sources needed to reproduce the reviewed state

Consequence and human authority

Record the current answer, owner, supporting evidence, unknowns, and next review trigger. Use the organization's own risk language rather than treating this worksheet as a universal scoring model.

  • Consequential outputs or actions and who may be affected
  • Proposal, review, approval, execution, appeal, override, and escalation roles
  • Actions the system may never take automatically
  • Evidence and interface cues reviewers need to make an informed decision

Risk and concern inventory

Record the current answer, owner, supporting evidence, unknowns, and next review trigger. Use the organization's own risk language rather than treating this worksheet as a universal scoring model.

  • Organization-defined risk tier and the rationale for that tier
  • Safety, security, privacy, fairness, accessibility, reliability, legal, operational, and vendor concerns
  • Impact, likelihood, uncertainty, affected parties, mitigations, owners, and residual risk
  • Issues that require deeper review, independent expertise, or a stop decision

Evidence and evaluation

Record the current answer, owner, supporting evidence, unknowns, and next review trigger. Use the organization's own risk language rather than treating this worksheet as a universal scoring model.

  • Golden, edge, ambiguous, adversarial, refusal, access-control, and failure cases
  • Evaluation environment, fixtures, measures, thresholds, reviewers, and observed results
  • Known limitations, disagreement, exceptions, missing evidence, and unresolved questions
  • Links to versioned system cards, test results, data or model documentation, and decision records

Release decision and conditions

Record the current answer, owner, supporting evidence, unknowns, and next review trigger. Use the organization's own risk language rather than treating this worksheet as a universal scoring model.

  • Decision: proposed, discovery, pilot, conditional approval, approved for bounded use, paused, or retired
  • Approved scope, release conditions, exceptions, expiry, and required approvers
  • Rollback, blocked-action, incident, notification, and recovery criteria
  • Monitoring, change review, retraining or vendor-update triggers, and evidence-retention expectations

Retirement and handoff

Record the current answer, owner, supporting evidence, unknowns, and next review trigger. Use the organization's own risk language rather than treating this worksheet as a universal scoring model.

  • Decommissioning trigger, owner, user communication, and continuity path
  • Data, record, model, prompt, integration, and credential disposition
  • Client-owned artifacts, export formats, retained knowledge, and successor-system dependencies
  • Final review outcome, remaining obligations, and archive location
Usage boundary: A planning and decision record, not certification, legal advice, an impact assessment, or a complete implementation of NIST AI RMF, ISO/IEC 42001, or any regulatory regime. NIST's AI RMF 1.0 is voluntary and was under revision as of this resource's review date; the NIST playbook is not a checklist to follow in its entirety.

Need a measured baseline?

Turn the record into a bounded review and release decision.

Start with public-safe context. Detailed evidence moves only after scope, responsibility, and an approved channel are clear.

Private local search

Find a service, capability, evidence record, resource, or insight

Press / to open search when focus is not in a form field.

Search runs locally against the public site index.