The EU AI Act, ISO 42001, and your agents: what a credential broker actually covers
11 min read Fullmakt Team Updated
- compliance
- eu-ai-act
- iso-42001
- governance
- agents
- security
The EU AI Act is no longer a future problem. It entered into force in August 2024, the bans on prohibited practices took effect in February 2025, the obligations for general-purpose AI models landed in August 2025, and the transparency rules followed in August 2026. The big one for most teams — the regime for high-risk AI systems — is the next wave: after a 2026 amendment it applies from 2 December 2027 for stand-alone high-risk systems and from 2 August 2028 for AI built into regulated products (more on those dates below). If you build or deploy AI systems into the EU market, the compliance leads are already circulating spreadsheets.
And when those spreadsheets reach the row that says “AI agents,” they hit the same wall an auditor does: what can this agent actually do, who authorized it, and can you prove what it did? This post is about that wall — how the EU AI Act, the ISO/IEC 42001 management standard, and a credential broker line up against it, and, just as importantly, where they don’t.
Three layers, in three paragraphs
The law is the EU AI Act. It is risk-tiered: most systems carry light or no obligations, a defined set of high-risk uses carry heavy ones, and a few practices are banned outright. The heavy obligations fall on providers (who build the system) and deployers (who put it to use), and several of them reduce, in operational terms, to “be able to govern, supervise, and reconstruct what the system does.” That is the part agents make hard.
The management system is ISO/IEC 42001. The Act tells you what outcomes to achieve, not how to run your organization to get there. ISO/IEC 42001:2023 — the first certifiable AI Management System standard, ISO 27001’s younger sibling — is how a growing number of teams build that “how.” We mapped a credential broker against its Annex A controls in detail; this post is the regulatory frame around that mapping.
The enforcement layer is the broker. For agents calling real APIs, the choke point where you can actually control, attribute, and record an action is the place every call flows through. A credential broker is that place. It is not your compliance program; it is where a specific, important slice of it gets enforced and evidenced.
A caveat you have to hear first
It is tempting to draw a straight line — “get ISO 42001 certified, and you’re EU AI Act compliant.” That line does not exist, and saying it does is a mistake.
ISO/IEC 42001 is a voluntary standard. Under the AI Act, the thing that grants a legal presumption of conformity is a harmonised standard cited in the Official Journal of the EU, and those are written by CEN-CENELEC’s JTC 21. The first one arrived in 2026: EN 18286:2026, Artificial intelligence – Quality management system for EU AI Act regulatory purposes, cited in the Official Journal on 14 October 2026. It is not the same document as ISO 42001. The Commission describes it as designed to be layered on top of standards such as ISO/IEC 42001 rather than to replace them, and the presumption it gives is narrow: the Article 17 quality management system requirements that the standard covers, for providers who apply it. It gives no presumption for the requirements this post is about — risk management, logging, human oversight, robustness and cybersecurity.
So an ISO/IEC 42001 certificate remains a serious credibility signal and an excellent way to organize the work, and it is a sensible base to build an EN 18286 quality system on. It is not a legal shield, and it does not cover every requirement the Act imposes.
So the honest thesis of this post is narrow on purpose: the EU AI Act asks you to control, oversee, and keep records of what your AI does; ISO/IEC 42001 is how many teams operationalize that; and Fullmakt turns the specific “what can this agent do / prove what it did” controls into enforced policy. Everything beyond that slice, we’ll name explicitly.
Where agents meet the law
Strip the legal language off the agent-relevant obligations and three demands remain — the same triad the ISO 42001 post is built on:
- Control the action. A high-risk system should only be able to do what you decided it may. The Act’s accuracy, robustness, and security expectations (Art. 15) and its risk-management discipline (Art. 9) live here.
- Attribute the action. Human oversight (Art. 14) is impossible if you can’t say which system did what, on whose behalf, under what authority — and you can’t oversee what you can’t name.
- Record the action. Automatic logging over the system’s lifetime (Art. 12) and the deployer’s duty to monitor operation and keep those logs (Art. 26) are explicit, not implied.
A long-lived API key pasted into an agent’s environment fails all three at once: it authorizes everything, attributes to nobody in particular, and leaves you whatever logging each upstream service happens to do. A broker — one place every call is checked against policy, the real secret injected at the edge, and the call recorded in a tamper-evident log — addresses the triad by construction.
The bridge: obligation → control → mechanism
Here is the whole line at once. The Article references are themes, paraphrased — read the regulation and confirm with your own counsel; the numbering and scope are precisely the things that move. Coverage labels match the ISO 42001 post: Largely addressed / Partially addressed / Evidence input.
| EU AI Act obligation (theme) | ISO/IEC 42001 area | What Fullmakt does | Coverage |
|---|---|---|---|
| Automatic record-keeping over the system’s lifetime (Art. 12) | A.6.2.8 event logging | Every agent call in a hash-chained, tamper-evident trail: agent, credential, policy decision, principal | Largely addressed for agent-to-API activity |
| Human oversight — supervise, intervene, stop (Art. 14) | A.9 responsible use & oversight | Human-in-the-loop approval gates on sensitive operations; instant revocation; hard limits a prompt-injected agent cannot exceed | Partially addressed |
| Accuracy, robustness & cybersecurity (Art. 15) | A.6 / A.9 | Scope ceilings, rate limits, no standing secret to steal, least-privilege blast radius | Partially addressed |
| Risk-management system (Art. 9) | Clauses 6 & 8, A.5 | Logs answer “what can / does this agent do?” as input to assessment; the broker is the operational control point | Evidence input / Partially addressed |
| Data governance (Art. 10) | A.7 data | Resource-level scoping limits which data an agent can reach at all | Partially addressed |
| Deployer duty to monitor and retain logs (Art. 26) | A.6.2.6 operation & monitoring | Monitored operation, anomaly-visible call records, exportable retained logs | Partially addressed |
| Documentation & accountability (Art. 11; roles) | A.3 / A.4 | Per-agent identity — no shared keys; a live registry of agents, tools, credentials and reach | Partially addressed |
| Transparency to affected parties (Art. 13, 50) | A.8 transparency | Mostly product and disclosure work — the broker contributes attribution data only | Evidence input |
A few rows deserve a closer look.
Record-keeping (Art. 12) is the strongest fit. The Act wants high-risk systems to log events automatically across their lifetime so operation can be traced. For the slice where agents touch real systems, a broker produces exactly that: one log, one identity model, every call, with integrity you can verify rather than assert. Logging scattered across the model host and each upstream API gives you fragments; the broker gives you the spine.
Human oversight (Art. 14) needs a place to stand. Oversight is not a sentiment; it is the concrete ability to see what an agent is about to do and stop it. Approval gates on sensitive operations — plus a revocation switch that takes effect on the next call — are that ability, running in production rather than promised in a policy.
The deployer obligation (Art. 26) is the one people forget. Much AI Act attention goes to providers, but if you use a high-risk system you have your own duties — including monitoring its operation and keeping the logs it generates. If your agents call APIs through a broker, “monitor operation and retain logs” is a setting, not a project.
What this does not cover
This list matters as much as the table. Both at the level of the broker and at the level of ISO 42001 itself, the following are out of scope — and a vendor who blurs them is selling you risk:
- The legal machinery of conformity — conformity assessment, the EU declaration of conformity, CE marking, registration in the EU database. That is process and law, not tooling.
- Fundamental-rights impact assessments (Art. 27, for certain deployers).
- General-purpose / foundation-model provider obligations — model documentation, copyright policy, systemic-risk duties. A different regime, not shaped like agent access.
- Bias and fairness, accuracy testing, data quality and provenance, the model life cycle — the substance that lives inside the model, not in the call it makes.
- Writing policy, leadership commitment, training, management review — the human core of ISO 42001 (clauses 5–7 and 10).
A broker helps you run an AI management system and feed an audit. It does not give you a certificate — and a certificate is not the same thing as legal conformity.
A practical sequence
If the AI Act is on your roadmap and agents are in production, start where they touch real systems — it is the highest-leverage move and it doubles as evidence:
- Inventory the credentials your agents hold today. This is the start of your resource documentation, and it usually surfaces the scary ones.
- Move them behind a broker. Agents get references, never secrets. One control point replaces N copies of a key.
- Encode policy. Translate what an agent may do into rules the broker enforces, and ratchet scope down using what the logs show it actually needs.
- Turn on approval gates for the operations your risk assessment flags as sensitive. That is your Article 14 human-oversight story, live.
- Hand the auditor the log. One verifiable trail of every agent action, identity and policy decision attached, is your Article 12 record-keeping story in one export.
A note on the timeline
Plan against dates, but verify them close to the day — they have already moved once. The high-risk regime was originally set to apply from 2 August 2026 for stand-alone (Annex III) systems and from 2 August 2027 for AI built into regulated products. The “Digital Omnibus on AI”, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026, and it replaced those dates with two later, fixed ones:
- 2 December 2027 for stand-alone high-risk systems (Article 6(2) and Annex III).
- 2 August 2028 for high-risk AI that is part of a regulated product (Article 6(1) and Annex I).
What did not move: the prohibitions and the AI-literacy duty have applied since 2 February 2025 (the amendment softened the literacy wording and added new prohibitions that apply from 2 December 2026), the general-purpose AI obligations since 2 August 2025, and the transparency rules since 2 August 2026, with providers of generative systems already on the market before that day given until 2 December 2026 to add machine-readable marking of their output.
The amendment changed some of the high-risk requirements at the edges, but the ones this post leans on — risk management (Art. 9), record-keeping (Art. 12), human oversight (Art. 14), robustness and cybersecurity (Art. 15) and the deployer duties (Art. 26) — were not amended, although for product-embedded systems the Commission may now narrow how some of them apply. Deadlines shifted; the direction did not. The work the Act asks for around agents — control, oversight, record-keeping — is worth doing whether your deadline is next year or the year after, because it is also just what safe agent infrastructure looks like.
Conformity is downstream of architecture
None of this exists for the auditor. Scoped, short-lived credentials, per-agent identity, policy enforced at one choke point, and a tamper-evident log are simply what running AI agents safely looks like. The EU AI Act, in large part, is the law asking whether you operate that way — so when you do, a real block of the agent-shaped work stops being a project and becomes an export.
Build the architecture for safety. The compliance conversation gets shorter on its own.
Fullmakt is a credential broker, not a certification body, a law firm, or your compliance program. This post is general information, not legal advice. ISO/IEC 42001 certification is not, by itself, conformity with the EU AI Act, and it does not confer a legal presumption of conformity under it; the first harmonised standard cited for the Act, EN 18286:2026, covers the provider’s quality management system only. Dates and citations were checked against the Official Journal in November 2026. Validate every control and Article mapping against the current regulation, the applicable standards, and your own counsel and auditor.