Governance requirements
Future tenants will arrive through different standards universes, but the control objective is remarkably consistent: an agent should not turn its own proposal into authority. Its actions need accountable ownership, bounded access, human intervention where appropriate, and evidence that can be inspected after the event.
This is a practical control mapping, not legal advice, a certification claim or a statement that Agent Control Room makes an organisation compliant. Applicability depends on your role, use case, risk classification and wider control environment.
Read the implementation boundary on Security and the complete decision lifecycle on How it works.
Scope
The frameworks have different legal and assurance roles. The mapping below uses their current published editions and keeps those differences explicit.
European Union law
The mapping focuses on high-risk-system risk management, records, transparency and human oversight, plus deployer duties. Whether the Act applies — and whether an organisation is a provider or deployer — requires a separate legal classification.
Current consolidated Regulation (EU) 2024/1689 (opens in a new tab)Certifiable AI management-system standard
Covers an organisation-wide AI management system: policy, responsibilities, risk and impact processes, lifecycle controls, operation, monitoring, suppliers and continual improvement.
ISO overview of ISO/IEC 42001 (opens in a new tab)Voluntary, use-case-agnostic framework
Organises AI risk work into Govern, Map, Measure and Manage. It is not a checklist, and NIST is currently revising version 1.0; the references here identify relevant outcomes rather than claim conformance.
NIST AI RMF Core (opens in a new tab)Certifiable information-security management-system standard
Provides the security-management foundation around governed agents, including access, identity, logging, monitoring, cryptography, cloud services, suppliers and incident management.
ISO overview of ISO/IEC 27001 (opens in a new tab)Assurance criteria used in CPA examinations
This page concentrates on CC6 logical access and CC7 system operations and monitoring. These are criteria for evaluating controls, not a product certification that a feature can confer.
AICPA Trust Services Criteria (opens in a new tab)Control map
Rather than repeat the same product feature under five headings, the sections below group the recurring requirements. Each separates the framework expectation, the contribution Agent Control Room can make, and the work the customer still owns.
01 · Accountability
Organisations need defined accountability, policies and roles, an inventory of relevant AI systems, and a risk-based process for deciding how those systems may be used.
02 · Risk control
Risk decisions should be translated into repeatable controls appropriate to the purpose and context of the system, with unsafe or unsupported conditions handled deliberately.
03 · Human oversight
People assigned to oversight need authority, competence and a practical means to understand, intervene in or stop relevant operation rather than merely review it afterwards.
04 · Access
Human and machine identities should be authenticated, authorised to the minimum necessary scope, isolated from other tenants, and removable when access is no longer justified.
05 · Evidence
Relevant operation should be recorded so authorised reviewers can understand what happened, monitor controls, investigate anomalies and demonstrate that oversight operated as designed.
06 · Operations
Controls should be monitored, deviations investigated and risks treated throughout operation, with evidence supporting incident response, correction and continual improvement.
07 · Trusted context
Operators need enough information about purpose, inputs, limits and expected use to apply controls correctly and judge whether the evidence supporting an action is suitable.
08 · Third parties
Third-party services, dependencies and evidence should be governed without transferring accountability or silently weakening the organisation's own controls.
Execution safeguards
Most frameworks state control outcomes rather than prescribing one software design. Agent Control Room adds a set of execution-level safeguards that are more specific than many baseline formulations.
A grant covers one registered agent, one declared action and one exact payload fingerprint. It is not a general permission for a workflow or session.
The model can propose work but cannot issue, approve or verify the grant that permits execution.
Signed grants are short-lived and bound to the payload fingerprint. Approval grants are single-use; changed work needs a new decision.
An undeclared action, unknown agent or incomplete enforced context cannot inherit a broad default allow.
Approval state, idempotency, revocation, rate windows and cooldown claims are recorded around deterministic rule evaluation rather than hidden inside model behaviour.
Optional Complyee evidence can support, hold or refuse a locally allowed step. It can never become a second source of permission.
Application
Treat the page as a starting point for a control owner, security reviewer or auditor — not as a completed assessment.
We can walk your governance, security or audit team through the decision path and the evidence it produces. Your advisers remain responsible for deciding which obligations apply and whether the complete system meets them. Write to hello@agent-cr.com.