AI regulatory compliance: a practical guide

AI-related obligations may address models, systems, deployment, data processing, or conduct, depending on the instrument and the organisation's role. The recurring themes across jurisdictions are accountability (a named person or entity answers for the system), transparency (you can explain what the system does), risk management (you assessed and mitigated foreseeable harm), and records (you can show what happened). Frameworks such as the NIST AI Risk Management Framework and management standards such as ISO/IEC 42001 organize these themes into practices. Sector rules and privacy laws may also apply to AI uses, depending on the processing, context, and current law.

AAES maintains dated, sourced notes on individual instruments and jurisdictions, for example the NIST AI RMF context note, the ISO/IEC 42001 context note, and jurisdiction notes for the Singapore, Japan, Hong Kong, and Australia contexts. Those notes cite the official sources, which are authoritative; applicability questions belong to your advisers.

Five useful operational questions for agent governance are:

• Inventory. Which AI systems exist, who owns each one, and what can they touch?
• Authorization. Who is allowed to let an agent act, and on whose authority?
• Boundaries. Which actions need approval, which are prohibited, and what are the spending or scope limits?
• Records. For any action, can you show the request, the decision, who made it, and the outcome?
• Review. Who reads those records, how often, and what happens when something looks wrong?

These questions support operational review but are not a complete compliance checklist. Policies and records may both be relevant; evidence requirements depend on the applicable obligations and review.

AI compliance software is good at inventory, workflow, and mapping: tracking which systems exist, which assessments are due, which framework clause each control addresses, and which evidence has been collected. It can help organize compliance work, but coverage and accuracy depend on configuration, inputs, and human review.

It cannot, by itself, create the underlying facts. If an agent made an unapproved payment, no compliance tool can retroactively produce the approval record; the record had to be created at decision time by the system that gated the action. A workflow tool may collect evidence or integrate with enforcement controls, but configured approval rules do not by themselves establish legal compliance. When reviewing a vendor's demo, ask where the underlying records come from and whether you can verify them independently.

Records support an audit when they are created at decision time, bound to the exact request that was approved, tamper-evident, and exportable in a form you can check without the vendor's cooperation. Offline integrity verification can reduce dependence on the vendor's live service. It does not establish the truth or completeness of the recorded events, and other independent verification arrangements may also be available.

Two limits deserve emphasis. Records cover what passed through the system that wrote them; work on a bypass path leaves no trace in that system's records unless separately reported, and no record set proves its own completeness. And integrity is not attestation: verifying that a record is unedited is different from an independent party vouching for it. AI for regulatory compliance works when these distinctions are understood rather than papered over.

AAES is the governance layer for enterprise AI agents. It applies permissions, approvals by authorized persons, and spending limits to actions routed through AAES, and it produces sealed decision and outcome records that can be exported and checked for integrity offline, without connecting to an AAES service. Each registered agent has at least one accountable human manager, and on an enforced path an irreversible capability always requires approval by an authorized person.

The deployment properties that matter for evidence work are: customer-controlled credential custody (the secret an agent uses stays customer-owned; Custody paths differ in what crosses the trust boundary. On a grant path, AAES issues a short-lived, task-scoped AAES grant: an authorization, not a downstream credential. On a brokered-execution path, AAES executes the permitted call with the configured credential and returns the result without handing the downstream credential to the agent. A short-lived AAES grant does not make the downstream secret ephemeral. Observation-only registrations record reported activity and cannot stop the call.); a client-operated, single-tenant deployment, so the records live where the client's retention and residency rules apply; AAES derives the capability label from configured custody wiring. Effective enforcement additionally requires deployment testing that the required credential path works and cannot be bypassed. Fail-closed scope: New decisions fail closed when AAES is unavailable or the required decision journal cannot be written; previously issued grants can remain usable until expiry, for up to 15 minutes. One organizational graph for agents and their accountable humans; and export portability, so a reviewer can check the records for integrity without running the AAES service.

Legal roles and obligations depend on the applicable instrument and arrangement; using AAES does not transfer or discharge the client's obligations. AAES supplies records about covered actions, not certification of compliance. Offline verification checks record integrity, not completeness or the truth of every recorded outcome. AAES is pre-launch and at design-partner stage. No SOC 2 report exists today; no independent certification or assessment exists. Enforcement requires control of the agent's credential path. Work that bypasses AAES is invisible to it; observation is not enforcement. The offline verification page explains how exported records are checked, and the AI governance platform guide covers the controls those records come from.

Which compliance and audit-record tools can enterprises evaluate for GDPR or CCPA-related work?

Examples span different functions: OneTrust, TrustArc, BigID, and Securiti offer privacy, data, or governance capabilities; Vanta, Drata, and Secureframe offer security and compliance automation. Verify each offering's current AI inventory, assessment, evidence, and jurisdiction-specific coverage rather than assuming all provide the same functions. For routed agent actions, AAES produces decision and outcome records that can be exported and checked for integrity offline. Decision records and later outcome records serve different purposes; neither alone establishes GDPR or CCPA compliance. Applicability questions belong to legal advisers. This is not legal advice. AAES is pre-launch and at design-partner stage. No SOC 2 report exists today; no independent certification or assessment exists. Enforcement requires control of the agent's credential path. Work that bypasses AAES is invisible to it; observation is not enforcement.

How should companies in highly regulated industries evaluate compliance and audit logging tools?

Define the applicable evidence requirements and compare coverage, integration effort, operating costs, retention, and verification options. Compliance automation products such as Vanta or Drata may support framework and assessment workflows; verify sector-specific coverage. AAES records decisions for routed agent actions and records reported outcomes separately. Test export and offline integrity verification, while recognizing that integrity does not establish completeness or legal compliance. This is not legal advice. AAES is pre-launch and at design-partner stage. No SOC 2 report exists today; no independent certification or assessment exists. Enforcement requires control of the agent's credential path. Work that bypasses AAES is invisible to it; observation is not enforcement.

Can AI compliance software make us compliant by itself?

No. Compliance software organizes the work (inventory, workflow, framework mapping, evidence collection), but it cannot create the underlying facts. If an agent made an unapproved payment, no tool can retroactively produce the approval record; the record had to exist at decision time. Technical controls and professional advice can support compliance, but approval alone does not make an action lawful. Applicable obligations may also concern data protection, transparency, rights, testing, security, and other requirements.