Practical guide

Regulatory change management policy for AI-enabled work

This policy template establishes how an organization identifies, assesses, implements, and reviews regulatory changes that may affect AI-enabled work. It is an operating design, not legal advice or evidence that any organization meets a legal requirement. Use WATCH: Watch official sources, Assess applicability, Translate obligations, Change operations, and Hold evidence. Adapt the policy to the.

By Two Prune

This policy template establishes how an organization identifies, assesses, implements, and reviews regulatory changes that may affect AI-enabled work. It is an operating design, not legal advice or evidence that any organization meets a legal requirement. Use WATCH: Watch official sources, Assess applicability, Translate obligations, Change operations, and Hold evidence. Adapt the policy to the organization's entities, jurisdictions, activities, systems, risk appetite, and professional advice.

Policy objective

The organization will identify relevant legal, regulatory, and supervisory changes early enough to determine applicability, assign action, update affected operations, test implementation, manage exceptions, and retain a traceable decision record.

This policy covers enacted rules, effective dates, formal guidance, corrections, transitional measures, consultations that require preparation, and material enforcement or supervisory communications. It covers AI systems, third-party services, data, policies, contracts, user journeys, monitoring, records, and workforce practices where a scoped change may apply.

It does not authorize an AI system to interpret law, approve applicability, accept risk, or close a change.

Roles and accountability

| Role | Minimum responsibility | | --- | --- | | governing owner | approves policy, risk tolerance, material exceptions, and reporting | | legal or regulatory owner | confirms authoritative sources, interpretation, scope, and need for advice | | change coordinator | maintains register, triage, dependencies, dates, and evidence | | operational owner | implements changes in the affected workflow or system | | risk, privacy, security, records, or workforce reviewer | reviews impact within assigned expertise | | independent tester | tests critical implementation against approved requirements | | internal audit or assurance | independently evaluates design or operation when commissioned |

Names, delegates, authority, and escalation routes must be recorded. A RACI label without access, capacity, or decision rights is not sufficient.

NIST's Generative AI Profile includes a suggested action under Govern 1.1 to align development and use with applicable laws and regulations. It is voluntary U.S. guidance, not a legal determination, but it supports assigning governance and documenting applicability. NIST: Generative Artificial Intelligence Profile

Sources: [1]

W: Watch authoritative sources

The legal or regulatory owner must approve a source inventory for each entity and activity. The inventory records issuing authority, monitored channel, subject, jurisdiction, frequency, backup owner, and last successful check.

Alerts from vendors, newsletters, and AI summaries may help discovery but cannot replace the official instrument. Each candidate change must retain:

Automated monitoring must record failures and route unresolved items to a person. A silent feed failure is a control failure.

  • official source and stable identifier;
  • authoritative or controlling text version;
  • publication, adoption, application, and effective dates where different;
  • jurisdiction, entity, product, activity, and user scope;
  • amendment, correction, transition, and dependency information;
  • retrieval time and reviewer.

A: Assess applicability and materiality

Open one change record per distinct decision. Within the triage period set by risk, the legal or regulatory owner assigns one disposition: not applicable, monitor, assess, implement, seek advice, or escalate.

The record must show the reasoning path:

| Field | Required content | | --- | --- | | authority | exact article, rule, guidance, or notice | | scope test | entity, geography, activity, system, data, and affected people | | current state | policy, process, contract, control, or system behavior today | | gap | difference between current and required or expected state | | materiality | consequence, likelihood, reach, reversibility, and timing | | uncertainty | disputed interpretation or missing fact | | decision | disposition, approver, date, and reopen trigger |

“Not applicable” requires reproducible reasoning. “Monitor” requires a trigger and date. The consolidated EU AI Act illustrates why scope and timing must be explicit: provisions and operator obligations can have different application conditions and dates. The official text must be reviewed for the actual case. EUR-Lex: Regulation (EU) 2024/1689, consolidated text

Sources: [2]

T: Translate the change into testable requirements

Legal interpretation must be translated into an approved requirements record without losing the source anchor or uncertainty.

| Requirement | Source anchor | Affected artifact | Expected behavior | Owner | Test | Due date | | --- | --- | --- | --- | --- | --- | --- | | | | policy / UI / contract / data / monitoring | | | | |

Requirements should cover all affected layers, including:

Where interpretation remains uncertain, the record must state the assumption, exposure, advice sought, interim action, owner, and review date. Do not hide uncertainty by converting it into a definitive technical requirement.

  • policy and procedure;
  • product behavior and user communication;
  • data collection, access, retention, and deletion;
  • model, prompt, retrieval, and vendor configuration;
  • contracts and assurance requests;
  • records and reporting;
  • training, support, and escalation;
  • monitoring, incident response, and change control.

C: Change operations under control

The operational owner creates an implementation plan with dependencies, resources, approvals, release sequence, communication, testing, rollback, and closure criteria.

Required controls are:

NIST's AI RMF Core calls for clear roles, documented legal and regulatory requirements, testing in relevant conditions, and ongoing monitoring. Apply it as risk-management guidance, not a substitute for jurisdiction-specific obligations. NIST AI RMF Core

  • Critical requirements receive an independent review proportionate to risk.
  • Expected results are approved before final testing.
  • Testing uses the same relevant versions and conditions as intended operation.
  • Failed and incomplete tests remain visible.
  • Superseded policy, guidance, or configuration is withdrawn or clearly marked.
  • People affected by the change receive the information and practice necessary for their role.
  • Monitoring and incident routes are updated before closure.
  • Material divergence from the approved interpretation returns to legal or regulatory review.

Sources: [3]

H: Hold evidence, exceptions, and review

The change coordinator retains a versioned evidence packet containing source, assessment, approvals, requirements, implementation artifacts, test plan, results, communications, exceptions, monitoring changes, and closure decision.

An exception must state:

Expired exceptions automatically reopen. Closure requires confirmation that critical tests passed, failed items have approved treatment, current operation matches the reviewed version, monitoring is active, and remaining uncertainty is visible to the decision owner.

At least annually, and after a material incident or control failure, the policy owner reviews source coverage, late triage, overdue actions, exception aging, failed tests, monitoring findings, and changes closed without sufficient evidence. The review produces actions, owners, and dates.

  • unmet requirement and source;
  • affected operation and exposure;
  • reason and alternatives considered;
  • compensating action;
  • accountable approver;
  • expiry and review date;
  • trigger for escalation or withdrawal.

Escalation and hold conditions

Escalate or hold implementation when:

Only an authorized owner may accept residual risk. An AI assistant may summarize or compare source text, but its output must be checked against the official source and reviewed by the appropriate professional.

  • the official source or controlling version cannot be reconstructed;
  • applicability is disputed and exposure is material;
  • an effective date may be missed;
  • a critical requirement lacks an owner or test;
  • the running system differs from the approved evidence;
  • a vendor dependency blocks implementation;
  • a test fails or relevant evidence is missing;
  • an exception lacks authorized approval or has expired;
  • people affected by the change face an unaddressed material risk.

Worked policy record: disclosure change for an AI assistant

A fictional financial-services organization identifies a new disclosure requirement that may affect one customer-facing AI assistant. This example does not interpret a real law.

Watch. The monitor captures the official instrument, dates, jurisdiction, amendment history, and retrieval time. A vendor newsletter is attached only as a discovery aid.

Assess. Legal concludes that the web and exported-PDF journeys are in scope. The record links the conclusion to the system inventory and identifies uncertainty about a third-party conversation archive. The governing owner approves an interim hold on expanding the assistant while advice is obtained.

Translate. Requirements cover interface wording, persistent disclosure in exports, a human explanation route, support scripts, archive retention, vendor evidence, and post-release sampling. Each has a source anchor, owner, expected behavior, and test.

Change. Product updates the web interface and PDF renderer. Support rehearses the explanation route. Records and privacy reviewers approve the archive treatment. The vendor response remains open under a dated exception with restricted exposure.

Hold evidence. Independent testing finds that disclosure is visible on the web but absent from a PDF created through an older route. The change stays open. Product repairs the renderer and the tester repeats web, PDF, accessibility, and support scenarios against the released version. Monitoring samples both journeys for four weeks.

The closure decision states only that the specified requirements passed the recorded tests on that version and date, with the vendor exception still monitored. It does not state organization-wide legal conformity.

Policy record template

| Stage | Owner | Decision or artifact | Evidence link | Result | Exception | Review date | | --- | --- | --- | --- | --- | --- | --- | | watch | | official change | | | | | | assess | | scope and materiality | | | | | | translate | | approved requirements | | | | | | change | | implementation and tests | | | | | | hold | | closure and monitoring | | | | |

Practitioner QA

Two Prune can assess how a participant handles a realistic regulatory-change scenario: finding authority, scoping impact, using AI carefully, verifying evidence, managing exceptions, and communicating a recommendation. It does not provide legal advice, operate a compliance programme, certify conformity, or make employment decisions.

  • Can a reviewer reconstruct the official source and date?
  • Is applicability reasoning tied to current facts?
  • Are legal interpretation and operational requirement visibly connected?
  • Do owners have authority, capacity, and access?
  • Were expected test results frozen before testing?
  • Do failures and incomplete work remain visible?
  • Are exceptions approved, bounded, dated, and monitored?
  • Does the current system match the reviewed version?
  • Can incidents, guidance, or amendments reopen the record?
  • Does reporting avoid claims beyond the sampled evidence?

Sources

EUR-Lex: Regulation (EU) 2024/1689, consolidated text

  • NIST: Artificial Intelligence Risk Management Framework, Generative AI Profile
  • NIST AI Risk Management Framework Core

Sources: [2] [1] [3]

Sources

  1. 1.NIST: Generative Artificial Intelligence Profile · National Institute of Standards and Technology
  2. 2.EUR-Lex: Regulation (EU) 2024/1689, consolidated text · eur-lex.europa.eu
  3. 3.NIST AI RMF Core · National Institute of Standards and Technology