The EU AI Act read covered the regime with a statutory deadline: Annex III obligations take effect August 2, 2026. Two other frameworks show up in reviews that have nothing to do with the EU. US customer security questionnaires cite the NIST AI Risk Management Framework. Procurement teams that already ask for ISO 27001 certificates have started asking for ISO/IEC 42001. Neither is a law. Both generate the same operational question the Act does: what records exist, and can you produce them.

This post maps both to the runtime layer — the part that registers agents, gates their actions, and keeps the trail. The organizational parts (risk committees, management review, certification audits) belong to the deploying company; the runtime's job is to supply the evidence those processes consume.

NIST AI RMF: four functions, one evidence question#

NIST published the AI Risk Management Framework 1.0 in January 2023, with a Generative AI Profile following in July 2024. It is voluntary — there is no NIST certificate — but questionnaire authors use its four functions as section headings, so answers organized the same way transfer directly.

GOVERN asks for accountability structures: policies for AI use, defined roles, and an inventory of AI systems with assigned responsibility. The runtime artifact is the agent registry — each agent registered with a named owner, and a record of which connections it holds. Because the agents run through the registry, there is no separate inventory document to keep current.

MAP asks the organization to establish the context of each system: what it does, what it can reach, who is affected. For an agent, the answer is its connected tools and the actions it has actually taken. The connection list bounds what the agent can reach; the activity trail shows what it did with that reach.

MEASURE asks for tracking of identified risks using defined metrics. For agent fleets the measurable quantities today are spend (dollars and tokens per agent, with gaps flagged where cost evidence is missing), action volume, and the count of actions held for human decision. These come out of the same records the Spend and Activity screens read.

MANAGE asks for the ability to respond: prioritize risks, act on them, and document the response. The runtime artifacts are the gate and the pause — a gated action waits for a named person's approval or denial, both outcomes are recorded, and an agent that needs follow-up can be paused. The decision record, with approver and evidence attached, is the documentation MANAGE asks for.

A questionnaire section headed "describe your AI risk management practices" can be answered function by function with those four sets of records.

ISO/IEC 42001: the certifiable one#

ISO/IEC 42001:2023 defines an AI management system (AIMS), the way ISO 27001 defines an information security management system. It is certifiable, and it follows the same harmonized clause structure: context (clause 4), leadership (5), planning (6), support (7), operation (8), performance evaluation (9), improvement (10), plus an Annex A control catalog.

A certification audit spends most of its time on clauses 8 and 9: show that the controls you documented actually operate, and show that you evaluate them. That is an evidence request, and the runtime's records are what answer it:

AI system inventory and life cycle controls (Annex A's life-cycle objective) — the registry answers which agents exist, who owns each, and whether each is connected and ready for gated work.
Impact assessment inputs (the impact-assessment objective) — an assessment of what an agent can affect starts from its actual connections and its actual action history, not from a design document.
Operational control (clause 8) — the gate is a documented control an auditor can watch operate: a live queue of held actions, each waiting on a named approver, with the decision landing on the trail.
Performance evaluation (clause 9) — spend by agent, flagged evidence gaps, and the activity timeline are the monitoring records the clause asks the organization to review.

The certificate itself attests to the management system, and no runtime purchase produces one. What a runtime changes is the cost of the audit: the difference between reconstructing agent behavior from scattered framework logs and pulling rows from one trail.

Where the three regimes overlap#

Laid side by side, the EU AI Act, the RMF, and 42001 converge on the same short list of runtime records:

Record	EU AI Act	NIST AI RMF	ISO/IEC 42001
Agent inventory with owners	Art. 9 risk process input	GOVERN	Life-cycle controls
Automatic action records	Art. 12	MEASURE	Clause 9 evaluation
Human approval with evidence	Art. 14	MANAGE	Clause 8 operation
Stop / pause capability	Art. 14 safe halt	MANAGE	Clause 8 operation
Spend and usage by system	—	MEASURE	Clause 9 evaluation

One set of records serves all three, provided the records are produced automatically by the layer the agents run through, survive team-by-team framework choices, and carry named humans on every decision. That is the argument for keeping enforcement and record-keeping outside the agent frameworks — the gate post makes it in full.

The runtime side of all three regimes is shipped product: registration, gates, pause, spend, and the trail. The organizational side — risk process, management review, conformity assessment — is the deploying company's, and the records above are its inputs. See Trust for the compliance posture, including SOC 2 status.