AI systems and agents

What does AI literacy mean for a small business?

How to assign knowledge, behaviour and responsibility according to role and system

Electronic letter tiles with AI highlighted on a dark circuit-board surface.

AI literacy for a small business is not one generic course for everyone. It means giving each person enough knowledge and practical capability for the AI system they use, the authority they hold and the consequences of getting it wrong. EU AI Act Article 4 requires providers and deployers to take measures supporting AI literacy for staff and other people operating or using AI on their behalf. It does not prescribe a universal curriculum, certificate, training frequency or guaranteed individual level.

AI literacy for small business: the practical answer

A person is AI-literate in relation to a system when they can perform their role responsibly: understand the intended use and limits, follow data rules, verify what matters, recognise when to stop and know where to escalate.

The required capability changes with the work. An occasional user of an approved tool does not need the same depth as an integration owner, procurement lead or manager defining an AI-supported process. But a less technical role is not consequence-free. An occasional user can still expose confidential data or rely on a persuasive but wrong output.

M4TIC recommends starting with role, system and consequence, not a training catalogue:

  1. who interacts with the system;
  2. what task they perform and with what authority;
  3. which data and affected people are involved;
  4. which decision or action depends on the output;
  5. what happens when the system or person is wrong.

What EU AI Act Article 4 currently requires

Under the consolidated text following Regulation (EU) 2026/1744, Article 4 requires providers and deployers of AI systems to take measures supporting the development of AI literacy among staff and other people dealing with system operation and use on their behalf.

Those measures must take account of:

  • technical knowledge, experience, education and training;
  • the context in which the AI systems will be used;
  • the people or groups on whom the systems will be used.

Article 4 now expressly says that the obligation does not require a provider or deployer to guarantee a specific AI-literacy level for any individual. It also requires the Commission and Member States to support providers and deployers, particularly SMEs, and requires the AI Board to develop recommendations and common objectives.

This is practical information about the current text, not legal advice. Determining whether an organisation is a provider, deployer or both—and what measures are appropriate in a specific case—may require qualified review.

What Article 4 does not prescribe

The current obligation should not be expanded into requirements that the law does not state. Article 4 does not prescribe:

  • the same curriculum for every organisation;
  • a universal number of hours or annual frequency;
  • an individual certificate;
  • an exam or guaranteed individual level;
  • appointment of an “AI officer”;
  • a particular governance structure;
  • one mandatory evidence format.

The Commission’s Q&A—informative guidance rather than legislation—says no certificate is needed and different learning approaches may be appropriate. It also says organisations can keep an internal record of training or other guiding initiatives. “Can” does not turn a register, test or timetable into a universal statutory requirement.

The AI Office’s living repository offers examples for learning and exchange. The Commission states that copying a listed practice does not automatically create a presumption of Article 4 compliance or amount to endorsement.

No automatic shortcut for SMEs, occasional users or contractors

Article 4’s particular support for SMEs concerns help with fulfilling the obligation; it is not a general SME exemption. Nor should a business infer that a low-consequence use, occasional user, contractor or service provider is automatically outside scope.

The operational questions are more specific. Is the person operating or using the system on behalf of a provider or deployer? What work are they doing? Which data and people are affected? How visible, reversible and containable is an error?

For external workers and suppliers, contracts and instructions should reflect the real role. A general clause does not create the ability to apply boundaries, verification and escalation during the task.

Before choosing AI training, inventory systems, roles and consequences

A small business can begin with a compact inventory. For each AI use, record at least:

Scroll horizontally to compare both columns

A compact inventory for each AI use
FieldOperating question
Person and roleWho uses, reviews, configures, buys or governs the system?
System and versionWhich product, model or configuration is actually in use?
TaskWhat is it allowed and prohibited from doing?
Decision or actionWhich outcome depends on the output?
DataWhat enters, where does it come from and what is prohibited?
Affected peopleWho may receive the benefit, error or consequence?
AuthorityDoes the system suggest, prepare, decide or act?
Error consequenceHow visible, reversible and containable is it?
MeasureInstruction, briefing, scenario, policy, control or review?
Owner and refreshWho checks adequacy, and what triggers another review?

This inventory is not certification. It prevents generic training from drifting away from the real systems and helps identify when privacy, security, employment, intellectual-property or other specialist review is needed.

Five AI-literacy roles in a small business

The following matrix is a M4TIC implementation aid informed by Article 4. It is not statutory text, a compliance scale or certification.

Scroll horizontally to compare every responsibility

Role-proportionate AI literacy in a small business
RolePurpose and authorityCapability, limits and verificationData, rights and peopleOversight, incidents and change
Occasional user of an approved toolKnows allowed and prohibited tasks; does not expand system authorityRecognises plausible but wrong output and verifies before useKnows prohibited or confidential data and privacy, confidentiality and IP rulesStops and reports anomalies, exposure or harmful output; knows where to ask for help
Reviewer or person relying on AI outputUnderstands which decision or action depends on the output and retains judgementApplies acceptance criteria; checks sources, completeness and relevanceConsiders bias, affected people and severity of consequenceCorrects, rejects, overrides or escalates; records the decision when required
Process owner or managerDefines intended use, roles, approvals, limits and fallbackSets proportionate tests and quality standards; separates a demo from operating evidenceMaps data, affected people and specialist-review needsAssigns oversight, logging, incident handling, review and refresh triggers
Builder, configurator or integration ownerUnderstands components, versions, permissions, connected tools and authority boundariesTests behaviour, failure modes, security, robustness and monitoring in contextProtects data and credentials; communicates limits and impactsManages change, rollback, dependencies, incidents and useful technical documentation
Leadership, procurement, privacy, security and governanceDecides whether to adopt, fund, buy, constrain or retire; assigns accountabilityRequests evidence, contractual terms, controls and risk-appropriate competenceAssesses affected people, rights, sector, vendor, transfers and qualified-review needsProvides resources, accountability, escalation and review after changes or incidents

One person may hold several roles in a small business. In that case, do not simply select the “highest” row. Combine the capabilities required by the real tasks and separate execution, approval and control where practical.

From AI awareness to accountable practice

General awareness is useful, but it is not enough for someone who makes decisions or changes the system. Capability develops across five practical responsibilities:

  1. Awareness: recognise AI use, intended purpose, limits and prohibited use.
  2. Competent use: apply data rules, verify outputs and activate escalation.
  3. Meaningful review: assess sources, relevance, bias, consequence and acceptance criteria.
  4. Process ownership: define authority, controls, fallback, incident handling and change.
  5. Technical/governance responsibility: test, integrate, procure or govern with evidence, security and accountability.

This progression is an M4TIC recommendation, not a legal ranking. Article 4 requires context-sensitive measures; it does not create these five levels or state that a person passes or fails a threshold.

Proportionate evidence without compliance theatre

Evidence should help the business remember what it decided, communicated, tested and updated. Depending on the context, it may include:

  • contextual policies and instructions;
  • briefings, training or role-based micro-modules;
  • exercises using plausible scenarios and failures;
  • review criteria and escalation routes;
  • notes on system, version, data and limits;
  • an internal record of measures;
  • incidents, corrections and material changes.

Course attendance shows participation, not necessarily competent use. At the other extreme, a small business should not create documents nobody applies. Evidence is useful when it connects a measure to a role, system and real consequence.

The adequate form remains case-specific. This article does not establish a sufficient evidence format or give a compliance conclusion.

When to refresh AI literacy, instructions and controls

AI literacy is not a one-off event. Review the measures when at least one of these changes:

  • system, model, version or configuration;
  • task, output or influenced decision;
  • allowed data, source, sensitivity or affected people;
  • connected tools, permissions or authority to act;
  • testing, monitoring or human-review results;
  • an incident, near miss or unexpected use;
  • a contract, policy, official guidance or applicable law.

Do not choose frequency by habit. Let the pace of change and potential consequence determine it. A minor interface change may not justify a new programme; a system that begins sending messages or changing records may require immediate review.

The first step: map one real workflow, not a generic course

Choose one AI use already present or under consideration. Identify the people who use, review, configure and govern it. Assign each person the purpose, allowed data, controls, authority, escalation route and evidence appropriate to their role.

Then test the arrangement with a realistic scenario: a persuasive wrong output, confidential input, potentially affected person, unavailable tool or version change. If nobody knows who decides or how to stop, the gap is not merely educational. It is a process and governance problem.

The aim is not to declare the company “AI-literate.” It is to keep people and systems prepared for the responsibilities they hold, updating the measures when the context changes.

Sources and further reading

Primary guidance behind this article.

  1. Regulation (EU) 2024/1689 — consolidated text of 27 July 2026European Union

    Authoritative legal source for the current Article 4 and AI-literacy definition.

  2. Regulation (EU) 2026/1744 — Digital Omnibus on AIEuropean Union

    Amending act that replaced Article 4 in 2026.

  3. AI Literacy — Questions & AnswersEuropean Commission / AI Office

    Informative guidance on scope, proportionality, formats, certificates, records, governance and enforcement; not legislation.

  4. New survey to gather practices for the AI literacy living repositoryEuropean Commission / AI Office

    Explains the repository’s purpose, limits and lack of automatic presumption of compliance.

  5. The enforcement framework of the AI ActEuropean Commission

    Non-binding overview of enforcement authorities and timeline.