Definition: Systemic Risk (EU AI Act)
Systemic risk is the EU AI Act’s classification tier for GPAI models whose scale or reach could cause large-scale harm across the EU.
Core characteristics of systemic risk (EU AI Act)
A model enters this tier through a compute threshold or a Commission decision, attaching to the model, not any application built on it.
- Presumed once training compute exceeds 10^25 FLOPs
- Can also be designated by the Commission regardless of compute, under Annex XIII
- Providers must notify the AI Office within two weeks of meeting the threshold
- Obligations sit with the provider, not any downstream use case
Systemic risk vs. General-Purpose AI Model (GPAI)
Systemic risk is a sub-tier inside General-Purpose AI Model (GPAI), not a separate AI system type. Only a handful of models cross into it; the rest carry baseline duties only. It differs from a High-Risk AI System, which labels a use case, not scale.
Importance of systemic risk in enterprise AI
Nearly every enterprise AI deployment in Germany runs on a GPAI model from a handful of providers, several past the threshold. Bitkom’s 2024 survey found close to three in four German firms use tools such as ChatGPT.
Methods and procedures for systemic risk classification
Three checks determine how a GPAI model lands in the systemic-risk tier.
Compute threshold presumption
The EU AI Act creates a rebuttable presumption of systemic risk once training compute exceeds 10^25 FLOPs, a level frontier providers already cross.
- Check provider documentation for disclosed training compute
- Treat the threshold as a presumption, not an automatic classification
- Monitor fine-tuning that could push a model past the line
Commission designation pathway
The Commission can also designate a model as systemic risk independent of compute, using Annex XIII criteria such as reach.
Notification and rebuttal process
Providers must notify the AI Office within two weeks of meeting the threshold, and may argue the model poses no systemic risk.
Important KPIs for systemic risk
Tracking exposure needs operational, strategic, and quality indicators.
Model evaluation and testing cadence
- Adversarial testing frequency: red-teaming cycles per model version
- Documentation coverage: vendors with published evaluation reports
- Notification lag: days between threshold crossing and notice
- Mitigation backlog: open findings awaiting a fix
Strategic exposure
Gartner projects over 80% of enterprises will use generative AI APIs by 2026, up from under 5% in 2023, making one-provider dependence a board-level question.
Incident reporting readiness
Article 55 requires providers to report serious incidents without undue delay, so deployers benefit from knowing a vendor’s process in advance.
Risk factors and controls for systemic risk
Systemic-risk classification creates exposure for the provider and anyone building on the model.
Downstream dependency on a systemic-risk model
Relying on one systemic-risk provider concentrates exposure.
- Enforcement action could restrict or withdraw a model version
- A serious incident at the provider can cascade downstream
- Contracts rarely guarantee advance notice of a classification change
Cybersecurity exposure
Article 55 requires providers to secure the model and its infrastructure, but a company building on top still owns its own stack.
Reputational and regulatory exposure
If a provider faces action by the AI Office or a Market Surveillance Authority, downstream companies face questions too.
Practical example
A 95-employee industrial software vendor in North Rhine-Westphalia built a technical-support assistant on a systemic-risk foundation model. Its compliance lead reviewed the provider’s evaluation reports, documented why the assistant sits outside the high-risk categories, and folded the check into a quarterly review.
- Documented evidence the use case falls outside high-risk categories
- Quarterly tracking of the provider’s systemic-risk notifications
- A fallback provider identified in case access changes
- Internal escalation path for a provider-reported incident
Current developments and effects
Systemic-risk enforcement is moving from paper rules to active oversight.
AI Office enforcement ramp-up
The AI Office gained full investigative and fining powers over systemic-risk providers on 2 August 2026, a year after the obligations took effect.
- First enforcement actions and guidance are expected
- Fines for non-compliance can reach 15 million euros or 3% of turnover
- National authorities coordinate with the AI Office
Code of Practice adoption
Major providers signed the voluntary GPAI Code of Practice, spelling out how to meet the Article 55 duties.
Digital Omnibus adjustments
The Digital Omnibus proposal under review in 2026 touches several timelines, but leaves the systemic-risk threshold unchanged.
Conclusion
Systemic risk sits one level above standard GPAI duties and applies to only a handful of models, but its effects reach every company that builds on one. The compute presumption, Commission designation, and Article 55 duties are now in active enforcement, not draft form. Mittelstand companies rarely carry these obligations directly, yet still need to know whether their model sits in this tier. Treating that check as routine diligence avoids surprises later.
Frequently Asked Questions
What makes a GPAI model count as “systemic risk”?
A model is presumed systemic risk once training compute exceeds 10^25 FLOPs, or when the Commission designates it under Annex XIII.
Does a Mittelstand company using a systemic-risk model have direct obligations?
Usually not. Most access the model via a vendor API, so duties sit with the provider.
How is systemic risk different from a high-risk AI system?
Systemic risk classifies a model by scale, with obligations on the provider. A high-risk AI system classifies a use case, such as hiring.
What does systemic-risk compliance cost a company with under 250 employees?
Mainly staff time reviewing vendor documentation, typically a few days per vendor.
How long does it take to check whether a vendor’s model carries systemic risk?
A basic model-card check takes a few hours. A standing quarterly review takes two to four weeks to set up.
Why does this classification matter for companies building AI employees on third-party models?
An AI-Mitarbeiter inherits its foundation model’s regulatory status. Superkind builds on a sovereign, EU-compliant AI foundation to reduce that inherited exposure.
