By the time most organizations finish their planning cycles and stand up a cross-functional governance working group, the EU AI Act’s core enforcement provisions will already be in force (August 2026 deadline!) The question for regulated firms – banks, insurers, payment processors, professional services firms with EU exposure – is not whether to comply. It is what compliance actually requires, and what evidence they need to produce.
What Article 99 actually says
The EU AI Act’s penalty provisions under Article 99 are among the most significant in any technology regulation. For prohibited AI practices, penalties reach €35 million or 7 percent of global annual turnover, whichever is higher. For high-risk AI system violations – the tier covering most enterprise AI in regulated industries – the ceiling is €15 million or 3 percent of global annual turnover.
For a firm with €500 million in annual revenue, a 3 percent turnover penalty is €15 million – before remediation costs, reputational damage, or the cost of demonstrating compliance after the fact.
The Act’s high-risk category covers AI systems used in credit scoring, access to financial services, employment management, and critical infrastructure. A material portion of enterprise AI deployment sits in those domains. Organizations that have not mapped their AI systems against the Act’s risk classification framework are operating blind.
The evidence standard the Act implies
The EU AI Act does not simply require that organizations *have* governance; it requires that they can *demonstrate* it. Obligations for high-risk AI systems include:
- Technical documentation describing system purpose, design, and performance.
- Logging and monitoring sufficient for post-hoc traceability of behavior and outcomes.
- Human oversight measures ensuring qualified persons can monitor and override system outputs.
- Accuracy, robustness, and cybersecurity with ongoing monitoring.
For AI agents – systems that take autonomous, multi-step actions – the logging and human oversight requirements are particularly demanding. An agent that executes dozens of actions per minute creates an audit trail problem that manual logging cannot solve. The evidence standard implied by the Act is continuous, automated, and tamper-evident.
This is the critical gap. Organizations that have governance policies written down, but no technical control enforcing those policies in real time and producing verifiable evidence, are unlikely to survive a competent audit.
The gap IBM’s data reveals
The IBM Cost of a Data Breach 2025 found that 63 percent of organizations had no AI governance policy in place at all, and 97 percent of organizations that experienced an AI-related breach had lacked AI access controls. For organizations with EU market exposure, those averages are now a legal problem. The Act does not grade on a curve relative to industry practice; it sets an objective standard.
What a defensible compliance posture looks like
Practical compliance with the EU AI Act for agentic AI systems requires four demonstrable capabilities:
- AI system inventory and risk classification. A current, maintained register of AI systems in use, with documented risk classification against the Act’s criteria.
- Policy-driven access control at the action level. Runtime enforcement of what each agent is permitted to do, evaluated against a documented, versioned policy – not just authentication.
- Continuous, automated audit logging. Append-only records of agent actions, policy evaluations, and decisions, structured in a format compliance staff and regulators can work with. SARIF (Static Analysis Results Interchange Format) is an emerging standard for structured findings.
- Human oversight integration. A defined escalation path for out-of-scope actions, with evidence that the path functions.
Ethosure’s enforcement architecture – built on the `coding-ethos` open-source project – delivers all four. The CEL-based policy engine produces deterministic decisions linked to specific policy provisions. The append-only evidence ledger generates SARIF-formatted findings alongside JSONL event records. The four-disposition model (allow, transform, escalate, block) includes a formal escalation path. Deployment is local-first – no action data leaves the customer’s environment – directly relevant for organizations with EU data residency obligations under GDPR.
The timeline problem
Building compliance infrastructure takes time most organizations have not budgeted. A cross-functional AI classification exercise, a procurement and integration process for enforcement tooling, and a validation phase – that is a six-to-twelve month effort, and the clock is already running.
The organizations in the strongest position are those that start with a concrete, bounded pilot: one agent workflow, one policy bundle, one evidence pack. That exercise produces both technical foundation and organizational learning. It is also the artifact that demonstrates to a regulator that governance is an active practice, not a future intention.
The EU AI Act is not a surprise. For regulated firms with EU exposure, August 2026 is the moment consequences become real. The gap between policy and enforcement is exactly the gap the Act is designed to close.
