Limited Risk vs Minimal Risk Under the EU AI Act

Most business AI is not high-risk. It lands in one of the two lightest tiers — limited risk or minimal risk — and the difference between them is one short statutory provision: Article 50. If a system is caught by Article 50, it owes a transparency duty (telling people they are dealing with AI, marking synthetic content, disclosing a deepfake). If it is not, it owes no AI-Act-specific obligation at all. That is the whole distinction. This page draws the line precisely, with a comparison table and a worked example.

The quick test: does the system interact with people, generate synthetic media, do emotion recognition or biometric categorisation, or produce deepfakes? If yes → limited risk (Article 50). If no → minimal risk.

Limited risk = Article 50 transparency

"Limited risk" is not a label the Act uses for a tier of systems — it is shorthand for systems that trigger the transparency duties in Article 50 of Regulation (EU) 2024/1689. The duties are specific and fall on whichever actor the sub-paragraph names:

• Article 50(1) — providers must design systems that interact with people so those people are informed they are dealing with AI, unless it is obvious. (A provider design duty — the chatbot disclosure.)
• Article 50(2) — providers of systems that generate synthetic audio, image, video or text must mark the output as artificially generated in a machine-readable form.
• Article 50(3) — deployers of emotion-recognition or biometric-categorisation systems must inform the people exposed to them.
• Article 50(4) — deployers who produce or manipulate deepfakes (and certain public-interest text) must disclose that the content is artificially generated or manipulated.

All of Article 50 — including the 50(2) marking duty — has applied since 2 August 2026. (Do not confuse it with 2 December 2026, which is only the Article 111(4) transitional deadline for generative systems already on the market before 2 August 2026 to meet the marking duty. The Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026, also adds two new Article 5 bans, on non-consensual intimate imagery of an identifiable person (point (ba)) and CSAM (point (bb)), that apply from 2 December 2026.) A breach of Article 50 sits in the €15,000,000 or 3% penalty tier under Article 99(4).

Limited risk does not drag in the high-risk machinery — no conformity assessment, no technical file, no risk-management system. It is a disclosure obligation, nothing more.

Minimal risk = no AI-Act-specific duty

If a system is not prohibited, not high-risk, and not caught by Article 50, it is minimal risk. The Act imposes no mandatory obligation on minimal-risk systems. Spam filters, inventory forecasting, recommendation engines, internal analytics, most productivity tooling — the vast majority of business AI lands here.

Two things still apply, but neither is tier-specific:

• Article 4 AI literacy applies to all AI, in force since 2 February 2025 — providers and deployers must take measures to support the AI literacy of staff who use or operate AI. Since the Digital Omnibus (27 July 2026) no specific individual level has to be guaranteed, but the duty itself remains. It is not a minimal-risk obligation; it is universal.
• Voluntary codes of conduct under Article 95. The Act encourages providers and deployers of non-high-risk systems to adopt voluntary codes — for instance applying some high-risk-style safeguards by choice. These are voluntary and create no legal duty; adopting one is a trust signal, not a compliance requirement. (Do not confuse Article 95 codes of conduct with the Article 56 codes of practice, which are for GPAI model providers.)

So the practical content of "minimal risk" is: classify it, record why it is minimal, train your staff under Article 4, and you are done.

Limited vs minimal — side by side

How to tell them apart in practice

Run the same short check on each system:

1. Does a human interact with it and might not realise it's AI? → 50(1) disclosure → limited.
2. Does it generate audio, image, video, or text output? → 50(2) marking (provider) → limited.
3. Does it infer emotions or sort people into biometric categories? → 50(3) notice → limited. (Watch the line: emotion recognition in the workplace or education is prohibited under Article 5(1)(f), not limited.)
4. Does it create or alter realistic media of real people or events? → 50(4) deepfake disclosure → limited.
5. None of the above, and not high-risk or prohibited? → minimal.

The most common mistake is parking a customer-facing chatbot in "minimal" because it feels low-stakes. If it talks to people, it is limited-risk and owes the 50(1) disclosure. The second most common is treating an internal analytics tool as limited because it is "AI" — if it doesn't hit a 50 trigger, it is minimal.

Worked example: Tirol Reisen, ~45 employees, ~€7M turnover

Tirol Reisen, an Innsbruck travel firm, runs four AI features:

• A booking chatbot on its website. Interacts with customers → Article 50(1) disclosure → limited risk. Tirol uses it as supplied, so it is the deployer; the disclosure is a design duty on the chatbot's provider, which Tirol checks is switched on.
• An AI tool that generates marketing images for campaigns. Produces synthetic images → Article 50(2) marking applies to the provider of that generator; Tirol, as deployer, must not strip the marking and discloses any visual that counts as a deepfake under Article 50(4) → limited risk in the chain.
• A demand-forecasting model for hotel inventory. No human interaction, no synthetic media, no biometrics → minimal risk; only Article 4 literacy applies.
• An internal expense-categorisation tool. Same — minimal risk.

Verdict: two limited-risk systems carrying Article 50 disclosures, two minimal-risk systems carrying nothing AI-Act-specific. Tirol documents each classification, confirms the chatbot and image disclosures are in place — Article 50 has applied since 2 August 2026 — and supports staff AI literacy under Article 4. No conformity assessment anywhere — none of these is high-risk.

How Confir helps

Confir's rule-based engine runs the Article 50 trigger checks for each system and returns "limited" with the exact sub-paragraph (50(1)–(4)) and the disclosure owed, or "minimal" with the reason recorded. It is deterministic: the same intake always yields the same tier, so a chatbot never quietly lands in "minimal" and a plain analytics tool never gets over-scoped. The rule that fired is shown in plain English for your audit trail.

Frequently Asked Questions

What is the difference between limited risk and minimal risk?

Limited risk means the system triggers an Article 50 transparency duty — disclosing AI interaction, marking synthetic content, notifying emotion/biometric recognition, or disclosing a deepfake. Minimal risk means none of those triggers apply and the system is not high-risk or prohibited, so it owes no AI-Act-specific obligation. The dividing line is simply whether Article 50 is engaged.

When do limited-risk transparency duties start?

They already have. Article 50 transparency duties, including the 50(2) synthetic-content marking duty, have applied since 2 August 2026. Do not confuse this with 2 December 2026, which is only the Article 111(4) transitional deadline for generative systems already on the market before 2 August 2026; the Digital Omnibus (Regulation (EU) 2026/1744) also brings two new Article 5 bans from that date: non-consensual intimate imagery of an identifiable person (point (ba)) and CSAM (point (bb)).

Does a minimal-risk system have any obligations at all?

No mandatory AI-Act-specific obligations. Two non-tier-specific things still apply: Article 4 AI literacy, which covers all AI and has been in force since 2 February 2025, and the option to adopt voluntary codes of conduct under Article 95. Adopting a voluntary code is a trust signal, not a legal duty, so a minimal-risk system that simply classifies, documents and trains staff is compliant.

Is a customer chatbot limited risk or minimal risk?

Limited risk. Article 50(1) requires that people be informed they are interacting with an AI system, unless that is obvious from the context — a design duty on the chatbot's provider, which the deploying company should verify. Parking a customer chatbot in "minimal" is a common misclassification — if it talks to people, the disclosure duty applies and the system is limited-risk, even though it owes nothing beyond that transparency notice.

What's the penalty for breaching an Article 50 transparency duty?

A breach of Article 50 falls in the €15,000,000 or 3% of total worldwide annual turnover tier, whichever is higher, under Article 99(4) — the same tier as most provider and deployer duties, not the top €35M/7% prohibited-practice tier. For SMEs, Article 99(6) applies the lower of the percentage or the fixed amount, and Article 99(6a) extends that lower-of cap to small mid-caps (SMCs) for Article 99(4) and (5) fines, though not for prohibited practices.

Is emotion recognition always limited risk?

No. Emotion recognition in the workplace and in education is prohibited under Article 5(1)(f), not limited-risk. Elsewhere it can be lawful and then carries the Article 50(3) notice, and may also be high-risk biometrics under Annex III point 1. Always check the prohibited tier before concluding that an emotion-recognition use is merely limited-risk.

Related guides

• Limited risk explained
• Minimal risk explained
• How to classify your AI system
• Limited risk (glossary)
• Deepfake
• Article 50 — transparency obligations
• Codes of practice