EU AI Act fines & penalties: A 2026 Compliance Guide

A tiered penalty structure governs the EU AI Act, where fines are calculated as a percentage of your global annual turnover.

This model, which applies even to organizations that automate workflows using Activepieces, scales the cost of non-compliance with the size and revenue of your enterprise.

If you struggle to inventory legacy "shadow" automations, you'll leave your organization vulnerable to these top-line revenue hits. An unmanaged workflow is an unmanaged risk.

Scale fines by violation severity

Prohibited AI practices: The 7% ceiling

7% of global annual turnover or €35 million, whichever is higher, is the maximum fine triggered by deploying AI for forbidden activities like biometric categorization or social scoring.

An unmanaged workflow is an unmanaged risk.

This means your multinational corporation could face a multi-billion dollar penalty for a single systemic violation of fundamental rights.

The nature of the AI application is viewed by the European Commission as the primary driver of financial liability, as this hierarchy demonstrates.

High-risk obligations: The 3% tier

3% of global annual turnover or €15 million is the maximum fine for failing to comply with requirements for high-risk systems, such as those used in recruitment, which means companies face significant financial exposure for even minor procedural oversight.

This 3% threshold applies to technical documentation and risk management failures, which means companies must prioritize precise record-keeping to avoid triggering penalties. Even if an AI is safe, a lack of auditable logs creates a direct financial liability for your Chief Technology Officer.

Transparency and data governance: the 1% tier

Providing incorrect, incomplete, or misleading information to notified bodies results in a fine of up to 1% of global annual turnover or €7.5 million, meaning that transparency during the audit process is a critical compliance necessity.

A single inaccurate disclosure during an audit can cost your firm tens of millions of euros regardless of whether the AI caused actual harm, so you must prioritize rigorous internal documentation.

Small businesses and the SME cap

For startups and small-to-medium enterprises (SMEs), the Act mandates that fines must be the lower of the two amounts.

This protection ensures that a 7% fine doesn't automatically bankrupt a small innovator, thereby preserving the viability of emerging businesses within the competitive landscape, which means these companies can continue to invest in research despite occasional regulatory setbacks.

It shifts the accountability for survival from the regulator to your own ability to demonstrate a "good faith" effort toward compliance by allowing you to mitigate the most severe existential threats.

Specific euro caps for smaller entities

The "lower of" rule applies to each tier, meaning an SME facing a 7% violation is capped at the lower of 7% or €35 million.

For the 3% tier, the cap is the lower of 3% or €15 million, while the 1% tier is capped at the lower of 1% or €7.5 million, meaning that high-value transactions will inevitably hit these ceilings and limit the total applicable benefit.

This distinction is vital for startups with low revenue but high valuations, as it prevents a massive fine based on a theoretical percentage from exceeding the fixed Euro ceiling.

Without this cap, a pre-revenue startup could be wiped out by a percentage-based penalty that ignores its actual liquidity.

Step 1: Audit your systems for prohibited AI practices

You must categorize every model against the EU AI Act’s prohibited list to avoid the maximum penalty tier of €35 million or 7% of global turnover.

Without this cap, a pre-revenue startup could be wiped out by a percentage-based penalty that ignores its actual liquidity.

Relying on a developer’s verbal assurance that a system is "safe" creates a documentation gap. This gap leaves your Chief Risk Officer personally exposed during a regulatory inquiry.

The following audit sequence provides the technical trail required to prove a system doesn't employ forbidden heuristics. You must:

• Identify biometric categorization based on sensitive traits.
• Flag untargeted scraping of facial images.
• Detect emotion recognition in the workplace or educational institutions.

Identifying social scoring and biometric categorization

Social scoring is defined by the Act as the evaluation of natural persons over time based on social behavior. Consequently, you must scrutinize any loyalty program or credit-risk model using non-financial data for discriminatory outcomes.

If a system uses the GitHub repository hosting your model weights to track sensitive traits like political leanings or religious beliefs, it triggers an immediate ban. Your Steerco must decide whether to strip these variables entirely or face the total decommissioning of the tool.

Detecting emotion recognition in workplaces or schools

The regulation forbids AI from inferring the emotional state of a person in professional or educational settings. Your HR department’s "engagement sentiment" tools are likely non-compliant if they analyze individual facial expressions or vocal tones.

Because these systems often run as third-party SaaS integrations, your IT Director is responsible for auditing the API responses. This ensures the system doesn't capture or store emotional metadata.

Verifying the absence of untargeted facial scraping

Untargeted scraping of facial images from the internet or CCTV footage to create facial recognition databases is strictly prohibited. This practice violates the fundamental right to privacy.

To satisfy an auditor, your Data Protection Officer must provide a verified lineage of all training data. The DPO proves that explicit consent or legitimate public datasets sourced every image rather than automated web-crawling.

Step 2: Classify high-risk systems and document compliance

Applying Annex III criteria to your use case

Determining if an AI system is High-Risk requires your Product Owner to map the tool against specific categories of harm rather than general utility.

Under the EU AI Act, a system is High-Risk if it is a safety component of a product already covered by Union harmonization legislation. Examples include medical devices or machinery.

It is also High-Risk if the use case is explicitly listed in Annex III, which covers sensitive domains like biometrics, critical infrastructure, employment, or the administration of justice.

By visualizing the logic this way, you can pinpoint exactly which regulatory obligations trigger. Once you flag a system as High-Risk, you must pivot from voluntary guidelines to the mandatory requirements of a formal Risk Management System.

Setting up the Risk Management System (RMS)

The Risk Management System is a continuous, iterative process that your Engineering Lead must integrate into the software development lifecycle to identify and mitigate known and foreseeable risks. This framework requires:

1. Identification of risks associated with the AI system when used as intended or under conditions of reasonable misuse.
2. Estimation and evaluation of risks that emerge during post-market monitoring.
3. Adoption of suitable measures to eliminate or minimize identified risks through targeted design and control.

Drafting the Technical Documentation and Instructions for Use

Your Legal Counsel must ensure that technical documentation is final before the system is placed on the market. Your Legal Counsel provides the "source of truth" for how the AI functions.

The documentation must demonstrate that the system complies with the Act and be kept up to date to reflect any substantial changes.

Accompanying this is the Instructions for Use. This document provides the deployer with clear information on the system’s capabilities, its level of accuracy, and the human oversight measures required to prevent algorithmic bias or output errors.

Step 3: Implement automated transparency and record-keeping

Disclosing AI-generated content to end-users

The EU AI Act mandates that AI systems interacting with humans must be designed so that users are informed they are engaging with an AI.

The gap usually lies in the lack of a "disclosure layer" within the automation logic itself. To solve this, your Product Owner must ensure that the notification trigger is a hard-coded step in the interaction flow.

This creates a functional lock: if the disclosure step fails to execute, the rest of the automation sequence is programmed to abort.

Maintaining automatically generated event logs

Transparency is only as robust as the logs that prove it, requiring technical documentation that records every automated decision-point from input to output.

Compliance requires that logic changes are treated with the same rigor as production code. Activepieces facilitates this by syncing automation flows to Git and using Release Management to promote logic through separate projects and environments, ensuring every modification is versioned and reviewed.

You can check the Activepieces documentation for Git Sync and Release Management to see how this makes promotion a deliberate governance step instead of an accidental click of a publish button.

By capturing these "Success" states in a version-controlled repository, your Compliance Officer can reconstruct the system's state at the moment of any disputed output. This shift from manual oversight to automated logging ensures that the "Instructions for Use" are a live, auditable reality.

Flagging deepfakes and synthetic media

Deployers of AI systems that generate or manipulate image, audio, or video content (specifically deepfakes) must label the output in a machine-readable format.

While you may choose to add visible watermarks for brand safety, the regulation specifically requires the underlying metadata to be persistent. Persistent metadata allows downstream platforms to detect the synthetic origin.

Your Chief Technology Officer is responsible for selecting model providers that support C2PA (Coalition for Content Provenance and Authenticity) standards or similar digital signatures.

Without this automated metadata injection, you risk being held liable for the spread of misinformation.

Establish a compliance audit trail with Activepieces

Activepieces provides an MIT-licensed AI automation platform that bridges the gap between disparate AI services and a central version-controlled repository.

Every agent decision, the data it processed, and the resulting actions are traced step-by-step alongside deterministic workflow stages.

These traces export as event streams into the SIEM your security team already operates, ensuring that an agent's choices are audited with the same rigor as standard code.

Check the Run Details and Debugging UI for the per-step agent decision trace that enables this granular oversight.

Automating the Technical Documentation workflow

The platform generates an immutable record of how data moves between LLMs and internal databases. This satisfies the EU AI Act’s requirement for detailed technical documentation without relying on manual developer logs.

Because Activepieces can be deployed on-premise or within a private cloud, your Chief Information Security Officer retains total control over the metadata. This ensures that sensitive workflow logic never leaves your secured perimeter.

This systematic capture transforms documentation from a reactive, error-prone task into a continuous background process.

Enforcing human-in-the-loop approvals for high-risk outputs

Activepieces mitigates the risk of autonomous "hallucinations" by inserting mandatory approval steps into any workflow designated as high-risk by your internal steering committee.

By using the "Wait for Approval" integration, the system halts the execution of a flow until a designated human reviewer signs off on the generated content, creating a timestamped record of human oversight.

This mechanism ensures that the accountability for a specific output rests with a named individual rather than an untraceable algorithm.

Centralizing logs for regulatory requests

The platform aggregates execution data from every connected service into a unified stream. This allows your Legal Team to respond to transparency inquiries with a single, comprehensive report.

MoneyGram and FundingSocieties run Activepieces in production to manage these complex environments where auditability is a requirement. Instead of hunting for fragmented logs across different AI provider dashboards, you can export a consolidated history of every decision the automation engine facilitated.

This centralization means you can demonstrate compliance during a snap audit without the weeks of forensic data reconstruction that typically follow manual oversight failures.

EU AI Act readiness checklist

Immediate compliance readiness requires shifting from passive observation to a structured governance framework that treats AI logic as auditable code.

While engineering teams often focus on model performance, the EU AI Act demands proof of systemic oversight, which you can only achieve if the following operational foundations are established before the enforcement deadlines arrive.

Phased implementation of enforcement deadlines

The Act follows a tiered timeline where the most dangerous practices are banned first. At the 6-month mark, the prohibitions on social scoring and biometric scraping become fully enforceable, meaning any organization still running these systems faces immediate 7% turnover fines.

The 12-month mark introduces obligations for General-Purpose AI (GPAI) models, requiring providers to maintain technical documentation and transparency records. Failure to meet these marks results in regulatory action, as the grace period for basic governance expires.

Appoint a dedicated AI Compliance Officer

You must designate a single point of accountability to bridge the gap between technical automation and legal obligations. Without a specific AI Compliance Officer, the responsibility for regulatory adherence fragments across IT and Legal departments.

This individual holds the authority to veto non-compliant deployments and ensures that the technical documentation generated by version-controlled workflows matches the legal transparency reports required by regulators.

Inventory all third-party AI tools in use

Under the Act, a comprehensive audit of every large language model and automated decision system currently in operation is required.

You often discover dozens of "shadow AI" instances where employees have fed proprietary data into consumer-grade tools. This means your firm is unknowingly leaking intellectual property into models it doesn't control.

This inventory must distinguish between general-purpose AI and systems integrated into critical infrastructure, as the latter triggers more stringent reporting requirements under the Act.

Review contracts for 'High-Risk' liability shifting

Legal teams must re-evaluate service level agreements to ensure that third-party vendors aren't contractually offloading their compliance burdens onto you.

If a vendor classifies their tool as a general-purpose model but you use it for human resources or credit scoring, you may be legally redefined as a "provider" of a High-Risk system. This increases the cost of regulatory insurance and technical audits.

The following checklist provides a sequence for establishing this governance structure.

Step	Action
1	Appoint a compliance lead.
2	Inventory all internal LLM usage.
3	Map systems to the 6 and 12-month enforcement deadlines.
4	Set up an automated evidence collection process.

Executing these steps ensures that the technical evidence gathered by Git-based version control is funneled directly to the individuals responsible for signing off on regulatory filings. Once these roles and inventories are solidified, you can move from reactive troubleshooting to proactive risk management.

Frequently asked questions about EU AI Act penalties

Do these fines apply to companies outside the EU?

Extraterritorial reach ensures that any entity placing an AI system on the market is subject to the same penalty regime as local firms. This applies to any entity putting a system into service within the Union.

Leadership teams often mistakenly assume that offshore hosting provides a compliance shield. However, the regulation explicitly targets the point of impact rather than the point of origin.

If the output produced by the system is used within the EU, the provider or deployer falls under the jurisdiction of the Act.

This means a US-based developer using a version-control platform like GitLab to push updates to a European client must maintain an auditable trail that satisfies EU regulators. Otherwise, they risk the same turnover-based fines as a Parisian startup.

Can a company be fined for AI models deployed before 2024?

Any significant change to a legacy system's code or purpose triggers immediate compliance obligations under the new framework, though retroactive penalties aren't permitted.

While a model deployed years ago is generally "grandfathered" in, the moment a developer modifies the weights or changes the data ingestion flow, the system is viewed as a new iteration.

For your Chief Risk Officer, this creates a specific accountability gap. Without a Git-based history showing exactly when a logic change occurred, you can't prove to an auditor that a system remains in its exempt, pre-2024 state.

Who actually issues and collects the fine money?

National supervisory authorities in each Member State are responsible for enforcement, while the centralized AI Office oversees the most complex "GPAI" models with systemic risks.

This dual-layer approach means you might answer to the Spanish Data Protection Agency for a local HR tool while simultaneously reporting to Brussels for a foundational model.

Because these national bodies retain the collected fines to fund their own market surveillance activities, they have a direct structural incentive to maintain rigorous inspection schedules.

Related reading

• Top 6 Finance and Insurance Compliance Automation Tools

References

• Lawve AI