For Regulators & Policymakers

1. What You Are Worried About

If you supervise organizations that use AI to file, approve, or move things in the real world, you are likely concerned with:

• Unauthorized or opaque decisions
  AI or automation taking actions that no one clearly authorized.
• Lack of reconstructable evidence
  After an incident, there is no reliable record of who approved what, under which rules, and when.
• Fragmented governance
  Policies exist on paper, but there is no proof that they were applied at the moment of action.
• Cross-border and cross-vendor risk
  Sensitive data, models, and workflows span multiple vendors and jurisdictions.

A sealed pre-execution governance runtime gives you a concrete pattern to point to — one that lets firms automate safely while preserving authority, accountability, and evidence.

3. Supervisory Expectations Checklist

You can adapt this as an internal or public checklist.

Regulators SHOULD EXPECT that regulated entities:

1. Identify high-risk AI workflows where AI or automation can submit filings, approve decisions, or move client or customer assets.
2. Front those workflows with a pre-execution governance runtime that is separate from the AI model and from vendor business logic.
3. Use client-owned policy, identity, and matter data as inputs to the runtime, rather than bespoke configurations hard-coded into applications.
4. Operate the runtime on a fail-closed basis, so that ambiguity in rules or identity results in refusal, not best-effort execution.
5. Record every approve, refuse, and override as a sealed, tamper-evident artifact under the firm’s control.
6. Store artifacts in a tenant-owned audit environment, under the firm’s retention, access, and jurisdiction rules.
7. Run new AI workflows initially in “observe-only” mode, to identify and address governance gaps before full enforcement.
8. Periodically review artifacts and metrics through internal audit, risk, or ethics functions, and make those reviews available to supervisors on request.
9. Ensure runtime decision data is not used for cross-tenant model training without explicit consent and appropriate safeguards.
10. Maintain clear accountability for policy content: business leaders, not vendors, own the rules that the runtime enforces.

5. Model Language You Can Reuse

Use these as templates for discussion papers, supervisory statements, or rules.
They are intentionally neutral and architecture-focused.

A. Supervisory Statement (Example)

High-Risk AI Governance – Pre-Execution Runtime Expectation

Where AI systems or automated workflows are capable of initiating, approving, or executing high-risk actions (including but not limited to regulatory filings, client instructions, payments, and record changes), supervised entities are expected to:

1. Front such workflows with a pre-execution governance runtime that:

• evaluates who is acting, on what, under which authority and risk policies, prior to execution;
• issues explicit approve, refuse, or supervised-override decisions; and
• records each decision as a sealed, tamper-evident artifact under the entity’s control; and

    2. Demonstrate, upon request, that this runtime operates on a fail-closed basis and is not bypassed for governed workflows.

B. Rule / Guideline Text (Example)

[X]. Pre-Execution Governance Runtime for AI-Enabled Actions

(1) A regulated entity that deploys AI or automated systems capable of initiating or executing material actions affecting clients, counterparties, or markets SHALL implement a pre-execution governance runtime for those actions.

(2) The runtime SHALL:

• (a) rely on entity-owned policy, identity, and contextual information, including applicable laws, mandates, and risk limits;
• (b) provide explicit approval, refusal, or supervised-override outcomes for each governed action;
• (c) default to refusal where policy, identity, or context is incomplete or inconsistent; and
• (d) generate sealed, tamper-evident decision artifacts that record, at a minimum, the time, actor, action type, applicable policy version, and outcome.

(3) Regulated entities SHALL ensure that high-risk actions cannot be executed through channels that bypass the governance runtime, except under documented contingency procedures approved by the entity’s governing body.

(4) Decision artifacts produced by the runtime SHALL be stored in an entity-controlled environment, under retention and access rules that support regulatory investigations, audits, and enforcement.

C. Examination / On-Site Procedure (Example)

Objective: Assess whether the firm’s use of AI in high-risk workflows is governed by an effective pre-execution runtime and whether decision evidence is available and reliable.

Steps:

1. Identify at least three AI-assisted workflows that can submit filings, approve transactions, or change critical records.
2. For each workflow, determine whether a pre-execution runtime gate is present and describe how it is invoked.
3. Obtain and review a sample of sealed decision artifacts for a recent period, including approvals, refusals, and supervised overrides.
4. Verify that artifacts contain sufficient metadata (actor, action, matter/context, policy version, outcome, timestamps) to support supervisory review.
5. Assess whether alternate technical paths exist that would allow actions to execute without passing through the runtime.

Evaluate the firm’s process for reviewing artifacts and metrics (e.g., by risk, compliance, audit, or ethics committees) and how issues identified are remediated.