Advanced9 min readAI Safety & Data Privacy

EU AI Act for SMEs: a practical governance plan

The EU AI Act is not just a legal problem for large vendors. A practical SME plan for inventory, risk classification, human oversight, transparency, vendor records, and rollout discipline.

What you should be able to do

For most SMEs, AI Act readiness starts with an inventory, risk classification, vendor evidence, human oversight, and disclosure rules. Do that before buying tooling or writing a 60-page policy.

Saved only in this browser.
In this article

Start with one row per AI system and actual use. Record its supplier, purpose, data, affected people, decision impact and accountable owner. That is the evidence needed to assess the company’s role and obligations; buying an AI tool does not answer either question.

This guide prepares an SME for qualified classification and operational decisions. It does not determine that a system is compliant. The worked example below deliberately leaves legal questions open until the relevant facts and review are available.

The timeline that matters

The EU AI Act (Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744) applies in phases. The timetable below was checked on 28 September 2026 against the Commission’s AI Act page, AI Omnibus entry-into-force notice, and AI Act Service Desk enforcement FAQ. The Act entered into force on 1 August 2024. Prohibitions, definitions, and the then-applicable AI-literacy provisions started applying from 2 February 2025; governance and general-purpose AI model obligations became applicable on 2 August 2025. The AI Omnibus entered into force on 27 July 2026 and amended parts of the timetable and literacy framework. Article 50 transparency duties have applied since 2 August 2026. The Service Desk describes a limited transition to 2 December 2026 for providers of systems placed on the market before 2 August 2026 to meet the Article 50(2) marking and detection obligation.

The final AI Omnibus timeline has rules for systems used in certain high-risk areas such as biometrics, critical infrastructure, education, employment, migration, asylum, and border control applying from 2 December 2027, while high-risk systems embedded into regulated products apply from 2 August 2028. Because these dates have already shifted once, check the Commission page before committing budget to a compliance deadline.

What non-compliance actually costs

The penalty framework is one input, not a substitute for determining scope. The tiers below are based on the Commission Service Desk’s Article 99 text and require counsel to confirm how the amended framework and national rules apply:

  • Prohibited practices under Article 5: up to €35 million or 7% of worldwide annual turnover, whichever is higher. The definitions and exceptions are detailed; labels such as “manipulation” or “biometrics” are not a complete legal classification.
  • The operator and notified-body obligations listed in Article 99(4), including specified provider, deployer, and Article 50 transparency duties: up to €15 million or 3% of turnover. Do not treat this as a cap for every other AI Act obligation.
  • Supplying incorrect, incomplete, or misleading information to authorities in reply to a request: up to €7.5 million or 1% of turnover.

Two SME-relevant provisions affect the framework: for SMEs and startups each cap applies as the lower of the fixed amount and the percentage, and authorities must apply the Regulation’s penalty factors and proportionality rules. Do not treat an informal label as proof that a practice is prohibited or permitted: stop any possible Article 5 use and obtain qualified classification. An inventory helps identify scope; it does not keep an organization out of it.

Who supervises this in Estonia

The Consumer Protection and Technical Regulatory Authority (TTJA) describes itself as an AI-system supervisory authority and says some product and high-risk-sector oversight is forthcoming. Its page was updated 10 September 2026 and checked again 28 September 2026. Its summary table gives a general 2 December 2026 transparency date, while the Commission’s Service Desk says Article 50 duties generally apply from 2 August 2026 and describes a narrower provider transition. Do not rely on the summary table alone to set a deadline. The exact authority and duties for a particular system still require qualified confirmation. For an Estonian SME, keep the system inventory and vendor evidence ready and check which Article 50 duties already apply, including whether the limited Article 50(2) transition applies.

Use those dates as planning inputs, not as a substitute for legal confirmation. The practical point for SMEs is simpler: start now, because inventory, ownership, documentation, and human oversight take time to build.

Provider, deployer, or buyer?

“Buyer” is a useful procurement label but not a substitute for an AI Act role. An SME can be a provider, deployer, importer, distributor, product manufacturer, authorized representative, or affected person depending on the facts. The simplified table below is triage only:

RoleWhat it meansSME examplePractical duty
Buyer (informal)You procure a tool with AI featuresCRM assistant, meeting summarizer, coding assistantDetermine the actual legal role; perform vendor due diligence and set internal use rules
DeployerYou put an AI system into use in your businessSupport triage, lead scoring, HR screening workflowOversight, monitoring, disclosure, records
ProviderYou place an AI system on the market under your nameAI chatbot product, scoring API, industry toolProduct compliance, technical documentation, risk management

An organization can hold more than one role. A company that buys a model API, wraps it in an industry-specific product, and sells it to customers may have provider and other operator roles. A company that uses a SaaS chatbot internally may be a deployer. Counsel must classify the actual system, substantial modifications, name/brand, purpose, and supply chain.

Do not guess this in a meeting. Put every AI system in an inventory and classify the role.

Build the AI inventory

Start with a spreadsheet. Every AI system gets one row:

FieldWhy it matters
System namePeople need a shared label
Vendor or ownerSomeone must answer questions
Business purposeRisk depends on intended use
UsersInternal staff, customers, applicants, public
Data categoriesPublic, internal, personal, confidential, restricted
Output useDraft, recommendation, automated decision, customer-facing answer
Human oversightWho checks it and when
DisclosureWhether people are told they are interacting with AI
LogsWhat evidence exists after use
Classification questionsPossible Article 5, Annex I/III or Article 50 issue, with facts and review status

Keep evidence and unresolved questions beside the label. A blank classification cell is not clearance.

Classify practical risk

Do not start by asking “is this high-risk under Annex III?” Start with operational impact:

Draft assistance. Editing or summarising for a reviewer may limit some output consequences, but the input can still contain restricted information. Assess purpose and data rather than assigning a low-risk legal category from the tool name.

Interaction and publication. Chatbots, voice agents, AI-generated media, public text or support responses. Disclosure and user clarity matter.

Decision-support workflows. Lead scoring, support routing, invoice handling, quality review, fraud flags. Human oversight, monitoring, and appeal paths matter.

Potential high-risk areas. Employment, education, credit, essential services, healthcare, law enforcement, migration, critical infrastructure, biometric categorization. Legal review required before deployment.

Possible prohibited practice: stop and escalate. Do not deploy a use that may fall within Article 5 while waiting for an internal “approval.” Qualified counsel must determine scope; an internal approval cannot make a prohibited practice lawful.

This is not a final legal classification. It is the triage that tells you where expert review is needed.

Minimum SME governance controls

For each non-trivial AI system, require six controls:

  1. Owner. One named person or team accountable for the system.
  2. Use boundary. What the system may and may not be used for.
  3. Data rule. What data can enter the system.
  4. Human oversight. Which outputs need review before action.
  5. Monitoring. How errors, complaints, drift, and vendor changes are noticed.
  6. Record. What evidence is kept: vendor docs, prompts, settings, approvals, logs, test results.

These controls address concrete failure modes: missing ownership, unknown data flows, ineffective oversight, and inability to reconstruct why an output was used. Their presence does not itself establish compliance or effectiveness; test and audit them.

A fictional repair-business FAQ/callback assistant can prepare content, synthetic tests and a notice, but collection and launch remain held while vendor retention and legal role decisions are missing.
The fictional repair business can prepare and test on synthetic requests while retention and role decisions remain open. The inventory supports qualified assessment; it does not establish compliance or authorise data collection. Open the full-size diagram.

Vendor due diligence

For vendor tools, ask for evidence rather than promises:

  • Is customer data used for training by default?
  • Where is data processed and stored?
  • What retention controls exist?
  • Are enterprise settings available for training opt-out, logging, SSO, and access control?
  • Does the vendor provide AI Act, GDPR, security, and subprocessors documentation?
  • Can the AI feature be disabled or scoped?
  • Does the vendor disclose model providers and major architecture changes?
  • What happens if the vendor changes model, prompt, or retrieval behavior?

If required evidence is missing, do not send the affected customer, employee or confidential data while the gap remains. A “low-risk” label does not repair an unknown data flow. Use public/synthetic preparation or a separately approved alternative.

Disclosure and human oversight

For customer-facing AI, disclosure should be simple and visible. If a customer is talking to an AI chatbot or voice agent, say so. For AI-assisted text, check applicable law, contract and policy together; internal policy cannot waive another obligation. Public-interest text has specific Article 50 qualifications. See the workplace disclosure decision guide.

Human oversight must be specific. “A human is in the loop” is not enough. Define:

  • What output the human sees.
  • What source evidence they can inspect.
  • Whether they can override or reject.
  • How much time they have.
  • Whether approval is logged.
  • What happens when the human disagrees with the system.

Oversight without authority is theatre. If the human cannot stop the action, they are not meaningful oversight.

A worked SME decision record

This fictional repair business wants a website assistant to answer opening-hours questions and collect callback requests. It does not diagnose hazards, quote jobs, rank employees or commit appointments.

Record fieldInitial entryEvidence or unresolved decision
Purpose/outputApproved FAQ answers and callback intakeTest that a request for electrical repair instructions goes to the published human route
DataContact detail and customer-written requestConfirm necessity, notice, processor, access and retention before collecting it
RoleBusiness integrates a vendor model into its branded serviceQualified reviewer assesses provider/deployer facts; “we bought an API” is insufficient
TransparencyDirect public interactionRecord applicable notice/marking duties and exceptions, not just a generic vendor badge
OversightNamed owner can disable intake and inspect approved evidenceTest access, stop behaviour and pending-request handling
Launch statePreparation onlyVendor retention answer and legal role decision still missing

The team can write the FAQ, test synthetic requests and prepare the notice while those questions remain open. It cannot claim the inventory itself authorised personal-data collection. A later expansion into appointment commitments or applicant screening triggers a new assessment.

Use an evidence sequence rather than a compliance calendar: inventory the use, resolve role and risk questions, establish controls, test them, then make the documented release decision. A month may be enough to assemble a baseline or may not; elapsed time does not establish compliance.

Keep the decision current

Use the SME governance checklist to retain the exact use, provision, source date, reviewer decision, controls and unresolved gaps. Revisit after a change in purpose, supplier, data, output authority or applicable rules. Where a national summary and Commission guidance differ, record the discrepancy and obtain a resolved deadline for the actual duty.

Read next

Continue through the same learning path with the next practical articles.

Take it further

Hand-picked external courses that go deeper on this topic.

AWS Skill Builder

AWS Security: Securing Generative AI on AWS

AWS Training and Certification

A cloud-vendor-specific complement to the Macquarie specialization: AWS's own Generative AI Security Scoping Matrix, OWASP Top 10 for LLMs, and MITRE ATLAS, walked through governance, legal, and compliance controls for five different AI deployment scopes — from consumer apps to self-trained models. Not GDPR-specific, but a genuinely practical advanced pick for teams whose AI workloads actually run on AWS and need concrete data-governance and compliance controls, not just theory.

Advanced~2 hours · self-paced (9 modules)Verified 8 days ago
Coursera · Macquarie University

Cyber Security: Data, Privacy and AI Security

Macquarie University Cyber Security Hub faculty

An advanced pick for professionals responsible for both GDPR compliance and AI security: a three-course specialization from Macquarie University's Cyber Security Hub that goes from GDPR/CCPA fundamentals and privacy-by-design, through privacy impact assessments, to a dedicated third course on securing AI systems against adversarial attacks and model leakage. It genuinely bridges the two rather than treating them as separate topics.

Advanced~47 hours · self-paced (3-course specialization)Verified 8 days ago
EU Digital Skills & Jobs Platform · CyberSuite

Secure AI Adoption for SMEs: Cybersecurity and the EU AI Act

CyberSuite

The rare AI Act course written for the companies the Act actually reaches: SMEs adopting AI, not the labs building it. Hosted on the European Commission's own skills platform, it pairs the legal side — roles, obligations, risk classification — with the security side (prompt injection, data leakage, supplier due diligence) that most compliance courses skip. For an Estonian SME deploying AI, this is the practical starting point.

Advanced~15 hours · self-pacedVerified 8 days ago

See all courses for AI Safety & Data Privacy