Practical guide
What evidence shows regulatory change management is operating?
A regulatory change process is not shown to operate by a subscription to legal updates or a policy that assigns responsibility. Operating evidence must connect an authoritative change to a scoped obligation, impact decision, approved implementation, communication, testing, exceptions, and follow-up. Use the PROOF test: Provenance, Relevance, Ownership, Operation, and Follow-up. This guide offers.
A regulatory change process is not shown to operate by a subscription to legal updates or a policy that assigns responsibility. Operating evidence must connect an authoritative change to a scoped obligation, impact decision, approved implementation, communication, testing, exceptions, and follow-up. Use the PROOF test: Provenance, Relevance, Ownership, Operation, and Follow-up. This guide offers a control-testing method, not legal advice. The applicable authority, effective date, jurisdiction, regulated entity, and legal interpretation require current professional review.
Control objective
The organization identifies relevant legal, regulatory, and supervisory changes in time to assess their effect on AI-enabled work, assigns accountable action, updates affected controls and operations, verifies implementation, and preserves exceptions and residual uncertainty for decision makers.
Define the scope before testing:
- legal entities, jurisdictions, products, services, and user groups;
- AI systems, vendors, data flows, policies, models, prompts, and workflows in scope;
- authoritative sources and monitored channels;
- change types, including enacted rules, effective dates, guidance, enforcement signals, consultations, and internal interpretations;
- accountable legal or compliance owner and operational owners;
- materiality, triage, escalation, and review rules.
P: Provenance of the change
Select a sample from the official source, not only from an alert summary. Capture:
The FCA Handbook, for example, maintains consolidated rules, legal instruments, timeline comparison, and structured current content. Those features can support monitoring, but a feed or comparison is only an input. The firm still must determine relevance and implement the applicable requirement. FCA: Handbook of rules and guidance FCA Handbook FAQ
Fail provenance when the source is an undated article, the text version cannot be reconstructed, the effective date is confused with the publication date, or a secondary summary silently replaces the controlling authority.
- issuing authority and stable source link;
- instrument, article, rule, guidance, or notice identifier;
- publication, adoption, application, and effective dates where distinct;
- authoritative text version and retrieval time;
- jurisdiction and entity scope;
- amendment, correction, transition, or dependency information;
- internal interpretation owner and review date.
Sources: [1] [2]
R: Relevance and impact decision
The impact record should show the reasoning path, not just a red, amber, or green label.
| Field | Required evidence | | --- | --- | | Scope test | entity, geography, activity, system, data, and user conditions | | Obligation or expectation | source anchor and interpretation | | Current state | affected process, control, policy, contract, or product behavior | | Gap | exact difference between current and required or expected state | | Consequence | operational, customer, data, reporting, or enforcement exposure | | Decision | no action, monitor, implement, escalate, or seek advice | | Authority | reviewer and approver | | Uncertainty | unresolved interpretation or dependency |
Test both inclusions and exclusions. For a “not applicable” decision, reproduce the scope reasoning from the official text and the organization's current facts. For a monitoring decision, define the event that will reopen assessment.
The consolidated EU AI Act illustrates why date and scope discipline matter. Different obligations and supporting measures have distinct application conditions and timelines; post-market monitoring duties for relevant high-risk systems require systematic collection and analysis over the system lifetime. This source is a concrete example, not a claim that every organization or AI workflow is covered. EUR-Lex: Regulation (EU) 2024/1689, consolidated text
Sources: [3]
O: Ownership and approved plan
Trace the impact decision into an action register:
| Action | Owner | Dependency | Due date | Approval | Evidence | Status | Exception | | --- | --- | --- | --- | --- | --- | --- | --- | | | | | | | | | |
Verify that owners have authority, capacity, and access. The plan should cover all affected layers, which may include policy, product behavior, data handling, vendor terms, records, user communication, training, monitoring, incident response, and reporting.
Check that dependencies do not create an impossible sequence. A policy update marked complete while product controls, contracts, and training remain unchanged is not full implementation.
O: Operation and implementation evidence
Inspect the artifacts that show the change entered normal operation:
Reproduce a risk-based sample. Start with the official change and follow it through impact, action, implementation, test, and current operation. Then reverse the direction: select a changed control or system behavior and trace it back to authority and approval.
The U.S. Department of Justice's Evaluation of Corporate Compliance Programs asks how organizations monitor and implement policies that reflect changing legal and regulatory risks, how policies are designed and updated, and whether compliance programs work in practice. It is prosecutorial guidance for its context, not a universal control standard, but it supports testing operation rather than accepting a paper process. U.S. DOJ: Evaluation of Corporate Compliance Programs
- approved policy or procedure diff;
- configuration, code, model, prompt, retrieval, or workflow version;
- contract or vendor-control update;
- training and communication audience, content, delivery, and exceptions;
- access or entitlement change;
- test plan, expected result, observed result, and reviewer;
- deployment, release, or process-change record;
- updated monitoring, incident, and reporting rules;
- evidence that superseded material is withdrawn or clearly marked;
- dated approval and separation of duties where required.
Sources: [4]
F: Follow-up, exceptions, and continuing review
The control should not close at deployment. Verify:
The 2025 GAO Green Book requires federal agencies to monitor internal controls, evaluate issues, and remediate deficiencies on a timely basis. Other organizations may adopt it as a framework, but its federal authority should not be misrepresented. Its principles are useful for testing whether identified gaps lead to documented correction and follow-up. U.S. GAO: 2025 Green Book
- open interpretations and dependencies have review dates;
- late or partial actions remain visible to the decision owner;
- exceptions include rationale, exposure, compensating action, approver, and expiry;
- failed tests create corrective actions and challenge retests;
- monitoring checks the changed behavior in operation;
- incidents, complaints, audit findings, and supervisory feedback can reopen the change record;
- the source remains monitored for amendments or implementation guidance;
- closure is approved only after evidence is complete.
Sources: [5]
Copyable operating-evidence record
Use one record per sampled change:
| Evidence stage | Artifact | Test performed | Result | Reviewer | Gap or exception | | --- | --- | --- | --- | --- | --- | | Official source | | version, date, scope | | | | | Impact | | reproduce applicability and gap | | | | | Action | | confirm owner, due date, dependencies | | | | | Implementation | | inspect changed artifact | | | | | Test | | reproduce expected behavior | | | | | Operation | | sample live or current process | | | | | Follow-up | | inspect monitoring and closure | | | |
Rate the sample as:
Do not calculate an organization-wide percentage unless the sample design and population support that interpretation.
- supported: every material stage is traceable and the test result matches;
- partially supported: a bounded gap exists without invalidating all evidence;
- not supported: a material stage is absent, contradictory, or not reproducible;
- insufficient evidence: the control may have operated, but the record cannot support a conclusion.
Worked test: new AI transparency requirement
A fictional financial-services firm monitors a new transparency requirement affecting one customer-facing AI assistant. The example is synthetic and does not interpret a real law.
The change register links to the official instrument and records publication and application dates. Legal identifies that the assistant is in scope because it generates customer-facing explanations. The action plan assigns product, compliance, legal, support, and vendor owners. It requires updated interface wording, an explanation route, revised support scripts, a vendor evidence request, and post-release sampling.
The tester selects the change after its recorded closure date. PROOF finds:
The tester rates the change partially supported, reopens it, assigns the PDF defect, extends monitoring to exports, and keeps the vendor action visible. A challenge retest verifies the web and PDF journeys, the support route, and the updated monitoring sample. Only then does the owner approve closure.
The narrow conclusion is that the sampled change passed its revised test on that date. It does not establish legal conformity across the organization or future operation.
- provenance is supported by a frozen official source and dated interpretation;
- relevance is supported by a system inventory and customer-journey map;
- ownership is partial because the vendor evidence request remains open;
- operation fails one scenario because the disclosure disappears in an exported PDF;
- follow-up is weak because the post-release sample covers only the web interface.
Failure conditions
Hold or fail the operating-evidence conclusion when the official source cannot be reconstructed, scope reasoning is missing, a “not applicable” decision is unsupported, material actions lack owners, policy changed without operational implementation, tests were written after results to match the outcome, exceptions disappeared at closure, or the current system differs from the reviewed version.
Two Prune can place a participant in a realistic regulatory-change scenario and evaluate how they locate authority, scope an obligation, use AI, verify evidence, manage exceptions, and communicate a recommendation. It does not provide legal advice, certify conformity, operate a compliance programme, or make employment decisions.
Sources
FCA: Handbook of rules and guidance
- FCA Handbook FAQ
- EUR-Lex: Regulation (EU) 2024/1689, consolidated text
- U.S. Department of Justice: Evaluation of Corporate Compliance Programs
- U.S. GAO: 2025 Green Book
Sources: [1] [2] [3] [4] [5]
Sources
- 1.FCA: Handbook of rules and guidance · www.fca.org.uk
- 2.FCA Handbook FAQ · handbook.fca.org.uk
- 3.EUR-Lex: Regulation (EU) 2024/1689, consolidated text · eur-lex.europa.eu
- 4.U.S. DOJ: Evaluation of Corporate Compliance Programs · www.justice.gov
- 5.U.S. GAO: 2025 Green Book · U.S. Government Accountability Office