Practical guide

Change-Control Policy for AI-Enabled Work

A useful change-control policy defines which AI-enabled work changes require control, who may approve them, what evidence must exist before release, and how exceptions are reviewed. It should make a change traceable from request to post-release disposition without pretending that documentation alone proves the control operated effectively.

By Two Prune

A useful change-control policy defines which AI-enabled work changes require control, who may approve them, what evidence must exist before release, and how exceptions are reviewed. It should make a change traceable from request to post-release disposition without pretending that documentation alone proves the control operated effectively.

Policy objective and scope

Objective: ensure that material changes to AI-enabled workflows are identified, assessed, authorized, tested, released, monitored, and retained in a reproducible record.

Apply the policy to changes in models or providers, system prompts, retrieval sources, tools, permissions, data flows, decision rules, evaluation methods, user instructions, human-review routes, monitoring, and decommissioning. The policy owner maintains the scope; workflow owners classify individual changes; privacy, security, legal, domain, and accessibility reviewers participate when their stated triggers apply.

NIST SP 800-53 CM-3 calls for organizations to define controlled changes, review and approve or disapprove them with security and privacy impact analysis, document decisions, implement approved changes, retain records, monitor change activity, and coordinate affected parties. Adapt those control ideas to the organization's actual context. NIST SP 800-53 Rev. 5, CM-3

Sources: [1]

Classify changes before work begins

The initiator opens a change record with an identifier, purpose, affected workflow and users, current and proposed versions, source changes, expected benefit, plausible harm, reversibility, requested date, and accountable owner. The owner assigns one of three routes:

A change is material when it can alter outputs, evidence available to users, data exposure, access, oversight, assessment interpretation, or the consequence of error. Splitting one material change into small tickets does not make it standard.

  • Standard: repeatable, low-consequence change with a pre-approved procedure and evidence set.
  • Normal: material change requiring impact review, named approvals, test evidence, and a release decision.
  • Emergency: time-critical response to an incident or material exposure, with restricted authority, minimum pre-release checks, and a time-bound retrospective.

Required operating procedure

For every normal change, complete these steps in order:

NIST's AI RMF Core treats governance as continuous across the lifecycle and calls for policies, monitoring, clear roles, documented impacts, incident processes, and change management. This supports a lifecycle procedure rather than a one-time approval meeting. NIST AI RMF Core

  • Impact map: identify affected people, decisions, data, systems, sources, controls, monitoring, and downstream artifacts.
  • Risk and claims review: state changed risks, assumptions, prohibited uses, and any public or internal claims that must be revised.
  • Test plan: define representative cases, expected results, failure cases, comparison baseline, reviewer, and acceptance criteria before testing.
  • Approval: obtain the accountable workflow owner's decision plus triggered specialist reviews. Approval binds to the exact configuration or artifact hash.
  • Release: record implementer, environment, version, time, migration or rollback steps, and user communication.
  • Post-release check: verify the deployed version, monitoring signals, access, evidence capture, and unexpected effects during the stated window.
  • Closure: link results, exceptions, incidents, residual risks, follow-up owner, and next review date.

Sources: [2]

Approval and segregation rules

The initiator may prepare the record but may not approve a normal change they alone implement when the change can materially affect users or evidence. The workflow owner approves business acceptability. Specialist reviewers approve only within their mandates. The release operator confirms that approvals match the version being deployed.

Any edit after approval that changes behaviour, sources, data, permissions, tests, or claims invalidates that approval and returns the record to the appropriate review step. Administrative corrections that do not alter meaning may use the standard route if the policy owner has defined it.

Evidence retained for testing

Retain the request, impact map, triggered reviews, predeclared test plan, test inputs and outputs, comparison result, approval identity and time, exact version identifier, release evidence, post-release check, exception record, and closure decision. Record unavailable evidence as unavailable; do not convert it into a successful result.

The NIST AI RMF Govern Playbook recommends integrating AI risk management into existing policies, standardizing current documentation, defining metrics and tests, and tracking risks and responses. Treat its suggestions as design prompts and tailor them to the organization's risk tolerance and obligations. NIST AI RMF Govern Playbook

Test policy operation periodically by reconciling the complete change population with records, sampling across risk classes and outcomes, checking version-bound approvals, reproducing selected test results, and examining exceptions and emergency changes. State the population, period, sample method, missing records, and limitations.

Sources: [3]

Exceptions and emergency changes

An exception must name the policy requirement, reason, affected scope, compensating action, approver, start time, expiry, and evidence required for closure. Expired exceptions become unresolved findings; they do not silently renew.

Emergency authority is limited to named roles and circumstances. The record must explain why the normal route could not be used, what minimum checks occurred, what was changed, and how rollback works. A retrospective review is due within the policy's stated window. It may confirm the release, require repair, or reverse it, but it must not erase the emergency route from the record.

Worked change: policy-assistant prompt and source update

A service team proposes a new system prompt and a replacement policy corpus for an internal guidance assistant. The change owner opens AI-CC-042, classifies it as normal, and identifies effects on source selection, response wording, user escalation, and retained citations. The test plan includes twelve representative questions, three conflicting-source cases, two unsupported-exception traps, and a rollback comparison against version 3.4.

The domain and privacy reviewers approve hash 8af2... after the cases pass. Before release, an engineer makes a “small” prompt edit that changes the escalation instruction. The release operator sees that the deployed hash no longer matches the approval and stops the release. The owner reruns the three escalation cases, obtains new version-bound approval, deploys version 3.5, and checks citations and escalation events for seven days. One missing event stream remains open as an instrumentation finding rather than being counted as successful operation.

Review, disposition, and boundary

The policy owner reviews this document at least on the stated date and sooner after a material incident, regulatory change, new provider or model class, repeated exception, failed operating test, or change in organizational risk tolerance. Each review records keep, revise, or retire, plus owner and due date.

This policy is an educational starting point. Its existence does not establish legal compliance or control effectiveness. Those conclusions require current jurisdiction-specific review and evidence that the policy operated over a defined population and period. Two Prune may provide task-bounded assessment evidence; it does not approve system changes, audit the organization's controls, or make employment decisions.

Sources

  1. 1.NIST SP 800-53 Rev. 5, CM-3 · National Institute of Standards and Technology
  2. 2.NIST AI RMF Core · National Institute of Standards and Technology
  3. 3.NIST AI RMF Govern Playbook · National Institute of Standards and Technology