Insights — AI Act Implementation
AI Act implementation for financial institutions
Checklist for Compliance Leaders
For Compliance Leaders, Heads of Compliance and Regulatory Affairs at financial institutions.
The deadline is closer than you think
August 2026. That is when requirements for high-risk AI systems under the AI Act become fully applicable.
Sounds far away. It isn't.
Building governance structures, classifying systems, establishing controls, making evidence reproducible — this takes time. Institutions starting in Q2 2026 will not meet the deadline.
The question is not whether you should start. The question is where you stand now and what is still missing.
Inventory your AI systems completely
Before you can classify anything, you need to know what exists.
- Overview of all systems making automated or assisted decisions
- Including systems from external vendors
- Including internally developed systems
- Per system: purpose, owner, using department, decisions it influences
- Including systems in pilot or test
An incomplete register is the most common reason institutions have governance gaps.
Classify each system
Classification determines which obligations apply.
- Credit assessment systems
- Risk assessment and pricing for insurance
- AML analysis and transaction monitoring
- Critical infrastructure systems
- Personnel management systems (internal)
Classification is not a one-time event. With significant system changes, classification must be reviewed.
Establish a risk management system per high-risk system
The AI Act requires a continuous risk management system — not a one-time risk assessment.
- Identification of known and foreseeable risks
- Analysis for correct use and foreseeable misuse
- Mitigation measures per identified risk
- Residual risks evaluated after mitigation
- Review cycle established
- Owner designated per system
The risk management system is not a document. It is a process.
Ensure technical documentation before deployment
For each high-risk AI system, documentation must be available before use.
- General description and intended purpose
- Design and architecture description
- Training data: origin, scope, quality measures
- Validation and test results
- Description of human oversight
- Cybersecurity measures
- For external systems: documentation from the provider
For externally procured models, obtaining sufficient documentation is a practical bottleneck.
Establish human oversight
The AI Act requires effective human oversight — not a checkbox, but a structural arrangement.
- Per system: documented who fulfills the oversight role
- Criteria for when human intervention is required
- Overseers have sufficient insight
- Override capability is guaranteed
- Overseers are demonstrably trained
- Escalation path documented
'Human oversight' does not mean an employee reviews the output. It means that employee can understand, assess, and intervene.
Ensure automatic logging
High-risk AI systems must automatically log events.
- System logs activation periods
- Reference databases are logged
- Input data is stored
- Identification of involved persons is recorded
- Logs are tamper-proof
- Retention period is established
For external systems: verify the provider offers logging that meets requirements.
Register high-risk systems in the EU database
Registration obligation applies to providers and deployers in certain categories.
- Determine per system whether registration obligation applies
- Required information is available
- Registration planned before deployment
- Changes are communicated timely
Ensure reproducible evidence
Reproducible evidence is generated from a defined governance state — not assembled from loose documents.
- Governance state is retrievable at any moment
- Changes are dated and traced
- Reviews are documented with date and outcome
- Incidents are logged including context
- Evidence is demonstrably not assembled after the fact
The difference between 'we have this arranged' and 'we can prove we have this arranged' is the difference supervisors assess.
Most common mistakes
Treating classification as a one-time project
Systems evolve. Classification must keep pace.
Placing governance with IT
AI governance is a risk and compliance responsibility.
Considering external vendors 'their problem'
As deployer, you are responsible.
Waiting for definitive guidance
The AI Act is in force. Core requirements are fixed.
Building evidence after the fact
Supervisors assess whether evidence is reproducible.
Frequently asked questions
Do we need a conformity assessment?
What is the difference between provider and deployer?
What if we are unsure whether a system is high risk?
Where do you stand?
Knowing where you stand is the first step. A structured analysis of your supervisory exposure is the second.