EU AI Act Article 99: Penalties, Implemented in Practice
EU AI Act Article 99 sets the fines: up to EUR 35 million or 7 percent of turnover for prohibited practices, 15 million or 3 percent for provider and deployer obligations, 7.5 million or 1 percent for incorrect information. We compute the cap per registry record and build the governance layer so that, in normal operation, it produces the evidence that paragraph 7 treats as mitigating.
Penalties are the chapter every board reads first and every project team reads last. I think both are wrong. Article 99 of the AI Regulation is not a threat, it is a price list with assessment rules: paragraphs 3 to 5 state the caps, paragraph 7 states what the authority has to weigh when it sets the amount. If you know the second list, you know which evidence you need to start collecting today. The article is addressed to the Member States and through them to every operator, from provider to notified body.
The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. Article 99 belongs to a different phase: under Art. 113(b) it has applied since 2 August 2025, but a fine presupposes that the obligation breached itself already applies. The Omnibus moved deadlines; it removed no obligation. The overview is in EU AI Act in practice.
What Article 99 requires
Art. 99(1) and (2): Member States lay down rules on penalties, including administrative fines, warnings and non-monetary measures: effective, proportionate and dissuasive, taking into account the guidelines under Art. 96 and the interests of SMEs and start-ups. They notify them to the Commission by the date of application.
Art. 99(3): Non-compliance with the prohibition of the practices under Art. 5 costs up to EUR 35 million or 7 percent of total worldwide annual turnover of the preceding financial year, whichever is higher.
Art. 99(4): Up to EUR 15 million or 3 percent for non-compliance with the obligations of providers (Art. 16), authorised representatives (Art. 22), importers (Art. 23), distributors (Art. 24), the value chain (Art. 25(2) and (4)), deployers (Art. 26), notified bodies (Art. 31, 33, 34), and with the transparency obligations under Art. 50.
Art. 99(5): Up to EUR 7.5 million or 1 percent for supplying incorrect, incomplete or misleading information to notified bodies or authorities in reply to a request.
Art. 99(6): For SMEs including start-ups, the lower of the two amounts applies. The consolidated text after the Omnibus adds the same rule in paragraph 6a for small mid-cap enterprises, for paragraphs 4 and 5 only.
Art. 99(7): When deciding on a fine and its amount, the authority takes all relevant circumstances into account, in particular: (a) the nature, gravity and duration of the infringement and its consequences, the number of affected persons and the damage; (b) and (c) fines imposed by other authorities for the same activity; (d) size, turnover and market share; (e) other aggravating or mitigating factors such as financial benefits gained; (f) the degree of cooperation with the authorities; (g) the degree of responsibility in view of the technical and organisational measures implemented; (h) how the infringement became known, in particular through self-notification; (i) intent or negligence; (j) action taken to mitigate the harm suffered by the affected persons.
Art. 99(8) to (11): Whether public authorities can be fined is for each Member State to decide (paragraph 8); fines may be imposed by courts or other bodies (paragraph 9), with procedural safeguards (paragraph 10); annual report to the Commission (paragraph 11).
Art. 100 and 101: Union institutions, bodies, offices and agencies are fined by the European Data Protection Supervisor under Art. 100, up to EUR 1.5 million for Art. 5, otherwise EUR 750,000. Providers of GPAI models are fined by the Commission itself under Art. 101, up to EUR 15 million or 3 percent, since August 2026.
The technical example
An insurer has twelve AI systems in its registry and asks me its board’s question: which of them could cost 35 million, which 15, which 7.5? The answer hangs on two fields of the registry record: risk class and role.
Tier 1, paragraph 3: one system. The call centre wanted to run voice and emotion analysis on its own staff as a coaching aid. Emotion recognition in the workplace is prohibited under Art. 5(1)(f), except for medical or safety reasons. That is not a compliance topic, it is a stop: pilot ended, record set to “rejected”, with the reason.
Tier 2, paragraph 4: five systems. The risk assessment for supplementary health cover and the application check in life insurance fall under Annex III point 5(c); the insurer built both itself, so it is provider and deployer at once, and Art. 16 and Art. 26 are both in the catalogue. The applicant screening is Annex III point 4(a), bought in, so Art. 26. The customer chatbot and the generated marketing content run through Art. 50, whose obligations have applied since August 2026, and so have the fines.
No offence under paragraph 4: six systems. Claims routing, document extraction, fraud indicators on claims, marketing propensity, underwriting for commercial property (point 5(c) names only life and health) and the internal RAG assistant. But tier 3 applies to all twelve: whoever answers an authority incorrectly, incompletely or misleadingly risks up to 7.5 million or 1 percent under paragraph 5, regardless of risk class. That is the tier you pick up most easily, through a registry that is out of date.
The SME rule in paragraph 6 does not help the insurer: for it the higher amount applies, and 7 percent of group turnover is more than 35 million. Which leaves the question: what mitigates? Paragraph 7.
How we implement it
At SIFAMO we read paragraph 7 as a specification. Every letter asks for a piece of evidence, and the governance layer produces it in normal operation, not once the authority asks.
The registry as proof that we know our portfolio. Letter (i) distinguishes intent from negligence, letter (g) asks about technical and organisational measures; whoever does not know which models are running can prove neither. The registry with its rule “no registration, no route” proves the opposite: every system has an owner, a purpose, a risk class, a role and the reasoning behind the classification, with date and signature, the rejected emotion analysis included. The registry export is the answer to the authority’s first question, and it is correct because the same dataset controls the routes. That is our insurance against paragraph 5.
The evidence log as proof of monitoring. Letter (a) asks about duration, consequences, affected persons and damage. Without a log the authority estimates, and not in our favour. The log, append-only, per decision with input hash, model and version, policy version, output, signals and reviewer action, allows the counter-calculation: start on day X with version Y, N cases affected, M of them with a different outcome. Monitoring on the log shortens the duration, because drift and out-of-scope attempts land with the business owner, not in a ticket.
The incident runbook as proof of corrective action. Letters (f), (h) and (j) reward cooperation, self-notification and harm mitigation. Our runbook per registry record has four steps with named owners: kill switch set by the business owner, affected-persons analysis from the evidence log, notification to the provider and the market surveillance authority under Art. 26(5) and Art. 73, correction through rollback or a new version via the quality gates, every step with a timestamp. Whoever faces the authority within days with an affected-persons count and a completed correction satisfies three letters at once.
Oversight records as proof of human control. Letter (g) again. Tiered oversight, full, sampled or monitoring only, with blind re-checks and a feedback loop, produces one entry per reviewer action and shows that the human actually changed decisions. A reviewer who has never disagreed in twelve months is, to me, a warning sign, not evidence.
The fine exposure view for the board. From risk class, role and the catalogue in paragraphs 3 to 5, the registry computes the cap per record and places the state of the evidence next to it. Simplified:
{
"model_id": "DE-V-2026-0027",
"purpose": "Risk assessment, supplementary health insurance",
"risk_class": "high (Annex III point 5(c))",
"role": "provider and deployer",
"fine_exposure": {
"tier": "Art. 99(4) via Art. 16 and Art. 26",
"cap": "EUR 15 million or 3 % of group turnover, whichever is higher",
"obligations_apply_from": "2027-12-02"
},
"art99_7_evidence": {
"g_registry_signed": "2026-10-02",
"a_evidence_log_coverage": "100 % of decisions",
"g_oversight": "full review, blind re-check 2 %",
"f_h_j_runbook_last_drill": "2026-09-14",
"open_findings": 0
}
}
That is no legal assessment, but per system the cap and the state of the evidence that lowers it; the auditor sees the same four things: record, log, oversight records, runbook with the date of the last drill.
In Germany, the national AI market surveillance act (KI-MIG) assigns the responsibilities. For the financial sector, BaFin has been the market surveillance authority since July 2026, which Art. 74(6) already provides for financial institutions. So the supervisor that examines the insurer’s business organisation also examines its AI obligations. I am an engineer, not a lawyer. But a supervisor that demands the same evidence under two sets of rules is good news: we build it once.
Typical mistakes
Sorting by risk class and forgetting paragraph 5. The duty to give correct information hits all twelve systems; an out-of-date registry is the most likely source of a fine.
“That only applies from 2027.” Art. 99 has applied since August 2025; Art. 5 and Art. 50 are enforceable. The emotion analysis in the call centre does not wait for Annex III.
Collecting the evidence only once there is an incident. What paragraph 7 rewards must have existed before the infringement. Duration and cooperation do not arise after the fact.
Treating the fine as an IT risk. The cap hangs on group turnover and the business owner carries it. The fine exposure view belongs in that owner’s risk report.
EU AI Act Article 99 checklist
Every registry record has a risk class, a role, the applicable paragraph of Art. 99(3) to (5) and the computed cap; records without an offence are marked as such.
Group turnover and SME or mid-cap status clarified and dated.
The registry is the single source for answers to authorities; the export is reproducible and dated.
The evidence log allows, per system, the reconstruction of the start, duration and affected-persons count of an infringement.
Incident runbook with kill switch, affected-persons analysis, notification path and correction exists and is drilled yearly.
Oversight records show actual interventions by reviewers, not only approvals.
The fine exposure view is part of the risk report to the board; the competent market surveillance authority is named per system.
Open questions
Art. 9 to 15 are not in the catalogue of paragraph 4; the fine runs through Art. 16(a). Whether a breach of a single requirement counts as one infringement or several is open.
How authorities will weigh the letters of paragraph 7 will be shown by practice; I have seen no guidelines under Art. 96. That complete evidence lowers the fine noticeably is my expectation from GDPR practice, not a guarantee.
For groups it is open which turnover counts: that of the company that committed the infringement or that of the group. We calculate conservatively with group turnover.
Frequently asked questions
Does this apply to us as a deployer?
Yes. Art. 26 is expressly listed in paragraph 4, so up to EUR 15 million or 3 percent for breaches of deployer obligations, and paragraph 5 hits anyone who gives authorities incorrect information. What a deployer needs at minimum: AI Governance.
Does the authority use the turnover or the fixed amount?
For undertakings, the higher of the two. Only for SMEs and start-ups does the lower one apply under paragraph 6, and for small mid-cap enterprises under paragraph 6a for paragraphs 4 and 5. A mid-sized insurer falls into neither group.
Who pays, provider or deployer, when a bought-in system is non-compliant?
Each for its own obligations. If the provider breaches Art. 16, the fine hits the provider; if the deployer uses the system outside the instructions for use or without oversight, it hits the deployer under Art. 26. Whoever substantially modifies the purpose becomes a provider itself under Art. 25(1). The series: all articles at a glance.
EU AI Act Article 99: Penalties, Implemented in Practice
EU AI Act Article 99 sets the fines: up to EUR 35 million or 7 percent of turnover for prohibited practices, 15 million or 3 percent for provider and deployer obligations, 7.5 million or 1 percent for incorrect information. We compute the cap per registry record and build the governance layer so that, in normal operation, it produces the evidence that paragraph 7 treats as mitigating.
Penalties are the chapter every board reads first and every project team reads last. I think both are wrong. Article 99 of the AI Regulation is not a threat, it is a price list with assessment rules: paragraphs 3 to 5 state the caps, paragraph 7 states what the authority has to weigh when it sets the amount. If you know the second list, you know which evidence you need to start collecting today. The article is addressed to the Member States and through them to every operator, from provider to notified body.
The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. Article 99 belongs to a different phase: under Art. 113(b) it has applied since 2 August 2025, but a fine presupposes that the obligation breached itself already applies. The Omnibus moved deadlines; it removed no obligation. The overview is in EU AI Act in practice.
What Article 99 requires
The technical example
An insurer has twelve AI systems in its registry and asks me its board’s question: which of them could cost 35 million, which 15, which 7.5? The answer hangs on two fields of the registry record: risk class and role.
Tier 1, paragraph 3: one system. The call centre wanted to run voice and emotion analysis on its own staff as a coaching aid. Emotion recognition in the workplace is prohibited under Art. 5(1)(f), except for medical or safety reasons. That is not a compliance topic, it is a stop: pilot ended, record set to “rejected”, with the reason.
Tier 2, paragraph 4: five systems. The risk assessment for supplementary health cover and the application check in life insurance fall under Annex III point 5(c); the insurer built both itself, so it is provider and deployer at once, and Art. 16 and Art. 26 are both in the catalogue. The applicant screening is Annex III point 4(a), bought in, so Art. 26. The customer chatbot and the generated marketing content run through Art. 50, whose obligations have applied since August 2026, and so have the fines.
No offence under paragraph 4: six systems. Claims routing, document extraction, fraud indicators on claims, marketing propensity, underwriting for commercial property (point 5(c) names only life and health) and the internal RAG assistant. But tier 3 applies to all twelve: whoever answers an authority incorrectly, incompletely or misleadingly risks up to 7.5 million or 1 percent under paragraph 5, regardless of risk class. That is the tier you pick up most easily, through a registry that is out of date.
The SME rule in paragraph 6 does not help the insurer: for it the higher amount applies, and 7 percent of group turnover is more than 35 million. Which leaves the question: what mitigates? Paragraph 7.
How we implement it
At SIFAMO we read paragraph 7 as a specification. Every letter asks for a piece of evidence, and the governance layer produces it in normal operation, not once the authority asks.
The registry as proof that we know our portfolio. Letter (i) distinguishes intent from negligence, letter (g) asks about technical and organisational measures; whoever does not know which models are running can prove neither. The registry with its rule “no registration, no route” proves the opposite: every system has an owner, a purpose, a risk class, a role and the reasoning behind the classification, with date and signature, the rejected emotion analysis included. The registry export is the answer to the authority’s first question, and it is correct because the same dataset controls the routes. That is our insurance against paragraph 5.
The evidence log as proof of monitoring. Letter (a) asks about duration, consequences, affected persons and damage. Without a log the authority estimates, and not in our favour. The log, append-only, per decision with input hash, model and version, policy version, output, signals and reviewer action, allows the counter-calculation: start on day X with version Y, N cases affected, M of them with a different outcome. Monitoring on the log shortens the duration, because drift and out-of-scope attempts land with the business owner, not in a ticket.
The incident runbook as proof of corrective action. Letters (f), (h) and (j) reward cooperation, self-notification and harm mitigation. Our runbook per registry record has four steps with named owners: kill switch set by the business owner, affected-persons analysis from the evidence log, notification to the provider and the market surveillance authority under Art. 26(5) and Art. 73, correction through rollback or a new version via the quality gates, every step with a timestamp. Whoever faces the authority within days with an affected-persons count and a completed correction satisfies three letters at once.
Oversight records as proof of human control. Letter (g) again. Tiered oversight, full, sampled or monitoring only, with blind re-checks and a feedback loop, produces one entry per reviewer action and shows that the human actually changed decisions. A reviewer who has never disagreed in twelve months is, to me, a warning sign, not evidence.
The fine exposure view for the board. From risk class, role and the catalogue in paragraphs 3 to 5, the registry computes the cap per record and places the state of the evidence next to it. Simplified:
That is no legal assessment, but per system the cap and the state of the evidence that lowers it; the auditor sees the same four things: record, log, oversight records, runbook with the date of the last drill.
In Germany, the national AI market surveillance act (KI-MIG) assigns the responsibilities. For the financial sector, BaFin has been the market surveillance authority since July 2026, which Art. 74(6) already provides for financial institutions. So the supervisor that examines the insurer’s business organisation also examines its AI obligations. I am an engineer, not a lawyer. But a supervisor that demands the same evidence under two sets of rules is good news: we build it once.
Typical mistakes
EU AI Act Article 99 checklist
Open questions
Frequently asked questions
Does this apply to us as a deployer?
Yes. Art. 26 is expressly listed in paragraph 4, so up to EUR 15 million or 3 percent for breaches of deployer obligations, and paragraph 5 hits anyone who gives authorities incorrect information. What a deployer needs at minimum: AI Governance.
Does the authority use the turnover or the fixed amount?
For undertakings, the higher of the two. Only for SMEs and start-ups does the lower one apply under paragraph 6, and for small mid-cap enterprises under paragraph 6a for paragraphs 4 and 5. A mid-sized insurer falls into neither group.
Who pays, provider or deployer, when a bought-in system is non-compliant?
Each for its own obligations. If the provider breaches Art. 16, the fine hits the provider; if the deployer uses the system outside the instructions for use or without oversight, it hits the deployer under Art. 26. Whoever substantially modifies the purpose becomes a provider itself under Art. 25(1). The series: all articles at a glance.
Related: EU AI Act · All articles · AI Governance · AI Consultant
Recent Posts
Recent Comments