What are the signs that an organisation is not meeting Article 4 expectations in practice?

Common warning signs include a generic annual training deck, no inventory of active AI agents, unclear policies on what those agents may do, and no runtime record of actual actions. Another red flag is when only IT understands the deployment while compliance, legal, and executives lack visibility. Those gaps make it hard to show sufficient literacy on demand.

What an organisation usually gets wrong when Article 4 becomes “training only”

Article 4 expectations are not satisfied by a one-time awareness exercise. In practice, the first sign of trouble is that the organisation treats literacy as a presentation asset instead of an operating discipline, so there is no evidence that the people deploying, approving, or supervising AI systems understand their actual obligations and limits. That gap becomes visible when governance is detached from day-to-day use.

A second warning sign is that accountability lives in a single function, usually IT or a pilot team, while legal, compliance, risk, and leadership cannot describe what the system does, who approved it, or which controls were expected. When only the builders understand the deployment, the organisation cannot reliably demonstrate that knowledge has been distributed to the people who need it most.

Another common pattern is inconsistency between policy and practice. The policy may describe responsible use, but there is no inventory of active systems, no clear owner for each deployment, and no way to confirm whether staff or contractors received role-appropriate instruction before access was granted. That is a practical sign that literacy is not being managed as part of governance.

Where the evidence of non-compliance tends to show up

The most useful evidence is operational, not rhetorical. If the organisation cannot produce a current inventory of AI agents or systems, cannot show who is allowed to do what, and cannot explain how training maps to actual tasks, Article 4 readiness is weak. The same is true when there is no runtime record of actions, decisions, or exceptions, because the organisation then lacks a basis for showing what was actually understood and applied.

That evidence gap often appears in reviews, incident handling, or procurement. Teams may be able to point to a policy deck, but not to attendance records, role-based instruction, control attestations, or records showing that business owners understood the operational boundaries of the system. If an auditor or regulator asks for proof, the organisation is left with general statements instead of concrete artefacts.

Practitioners should also watch for silence from the business side. When executives, legal, procurement, or compliance cannot explain the deployment in plain terms, it usually means literacy has not reached the decision-makers who set scope, approve use, and accept risk. That is a stronger signal than any single training metric because it reflects whether understanding is embedded where decisions are made.

What good practice looks like before the gap becomes visible

Good practice is measurable and specific to the system. The organisation should be able to show that AI literacy is tied to roles, updated as the deployment changes, and reinforced where actual decisions happen. That means training, policy, oversight, and usage records should align, rather than existing as separate documents that never meet in practice.

It also means the organisation can answer basic questions without hesitation: which AI systems are live, who owns them, what users are allowed to do, which decisions remain human, and where exceptions are recorded. If those answers depend on informal knowledge from a small technical group, the organisation has not yet reached a defensible level of readiness.

Where the operating model is mature, there is usually a clear trail from policy to role assignment to review evidence. That trail does not need to be elaborate, but it should be strong enough that an informed outsider can see how literacy was delivered, by whom, to which audience, and for what purpose.

Risk and Threat Considerations

Weak Article 4 practice creates more than a documentation problem. It increases the chance that AI systems are deployed by people who do not understand their limits, which raises governance failure, unsafe use, and uncontrolled escalation of authority. It also makes it easier for poor decisions to persist because the organisation cannot prove who understood what, or when.

Failure mechanism: training is treated as a compliance artefact instead of a role-based operating control, so the organisation cannot show that the right people understood the right system at the right time.

Impact: the deployment becomes harder to govern, harder to defend in review, and more likely to produce unmanaged decisions, missed responsibilities, and weak accountability when something goes wrong.

Standards & Framework Alignment

This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.

NIST CSF 2.0 sets the technical controls, while EU AI Act, ISO/IEC 27001:2022 and ISO/IEC 42001:2023 define the regulatory obligations.

Framework	Control / Reference	Relevance
EU AI Act	Article 4 AI literacy	Article 4 directly governs AI literacy obligations for deployers and providers.
Recommendation — Build role-based AI literacy evidence and keep it current as systems and responsibilities change.
ISO/IEC 27001:2022	A.5.31 — Legal, statutory, regulatory and contractual requirements	Article 4 implementation depends on meeting a regulatory obligation in practice.
Recommendation — Map Article 4 duties into governance evidence and review them as a regulatory requirement.
NIST CSF 2.0	GV.OC-01 — Organizational Context	The answer depends on whether AI use, ownership, and decision paths are understood across the organisation.
Recommendation — Document AI system ownership, scope, and decision accountability in organisational context.
ISO/IEC 42001:2023	7.2 — Competence	AI literacy failures are competence and awareness gaps in the management system.
Recommendation — Define competence expectations and retain evidence that relevant roles are trained.

Practitioner Guidance

What to verify: Confirm that literacy evidence is role-specific, current, and tied to live AI use. A generic deck is not enough if the same people approving, deploying, and supervising the system cannot explain the system’s purpose, scope, and limits.

Common mistake: treating completion of a training module as proof of readiness. For this topic, the real test is whether the organisation can demonstrate who knows what, who is responsible, and how that knowledge is reflected in actual operations.

What good looks like: executives, legal, compliance, and operational owners can each describe the deployment in their own terms, and the organisation can produce an inventory, ownership record, and usage evidence on demand.

Practitioner takeaway: If the organisation cannot connect literacy to real ownership and real system use, it is probably managing Article 4 as awareness theatre rather than as a control it can defend.