AI-risk framework
The four risk levels, how they show up in finance, and the borderline cases where a system shifts level.
The Global Risk Hierarchy
Classified by potential impact on financial stability, consumer rights, and individual opportunity. Based on the EU AI Act framework.
Minimal Risk
Systems with negligible impact on human lives, financial decisions, or data security.
Internal spam filters, email triage, or simple analytics dashboards.
Deep dive+
What the rules require
No AI-specific legal obligations. Follow general rules such as GDPR and internal model governance.
Typical finance scenarios
- Spam filter on the bank's internal email
- Forecasting branch cash demand
- Summarising internal meeting notes
- Detecting duplicate invoices in accounting
Checklist
- ✓Keep an inventory of all AI tools in use
- ✓Monitor quality so errors are caught early
- ✕Don't let an internal tool drift into customer decisions unnoticed
Limited Risk
Systems that require transparency so users know they are interacting with AI.
Customer-facing chatbots and generative AI tools that draft reports or emails.
Deep dive+
What the rules require
Transparency duties: people must know they are talking to AI, and AI-generated content must be labeled.
Typical finance scenarios
- Customer service chatbot answering questions about fees
- AI-drafted market commentary sent to clients
- Fund or insurance recommendations where a human decides
- Synthetic voice in a phone bank
Checklist
- ✓Tell customers clearly that they are talking to AI
- ✓Offer an easy way to reach a human
- ✕Don't present AI-written analysis as if a human wrote it
High Risk
High-stakes systems that affect access to essential services, employment, or legal standing.
Automated credit scoring, loan approval algorithms, and hiring tools.
Deep dive+
What the rules require
Strict requirements: risk management, high-quality data, logging, technical documentation, human oversight, accuracy testing and registration before launch.
Typical finance scenarios
- Automated credit scoring for consumer loans
- Risk-based pricing of life and health insurance
- AI screening of job applicants at the bank
- Fraud models that block accounts automatically
Checklist
- ✓Test the model for bias across groups
- ✓Let a human review and overturn decisions
- ✓Give customers a clear reason for a rejection
- ✕Don't launch without documentation and an audit trail
Unacceptable Risk
Systems that pose a clear threat to safety, fundamental rights, or livelihoods.
Real-time social scoring that denies banking services, or manipulative AI.
Deep dive+
What the rules require
Banned in the EU. These systems may not be developed, sold or used.
Typical finance scenarios
- Social scoring that denies banking based on unrelated behaviour
- AI exploiting financial distress or age to push expensive loans
- Hidden manipulation that pressures customers into products
- Emotion recognition of employees at work
Checklist
- ✓Stop the project and escalate to compliance
- ✓Redesign the use case around legitimate data
- ✕Don't assume a vendor's tool is legal just because it is sold
Industry Spotlight: Banking & Fintech
Algorithmic Bias
Models must not discriminate based on zip code, gender, or ethnicity when scoring creditworthiness.
Explainability
Regulators increasingly require clear reasons for automated loan rejections.
Human Oversight
High-risk decisions must be reviewable by humans, not left to algorithms alone.
Where AI risk shows up in finance
Automated loan approval affects economic opportunity and requires fairness audits.
Internal transaction monitoring is usually minimal risk; if it blocks customer accounts, it becomes high risk.
AI assistants must disclose they are not human and escalate sensitive requests.
Using behavioral scores to deny essential banking services is prohibited.
When Does the Risk Level Shift?
The same AI system can land in different risk levels depending on how it is used. Small changes in scope, autonomy, or audience can move a system up the hierarchy. These borderline cases show what to watch for.
Fraud detection
A model that flags suspicious transactions for an analyst to review is internal tooling — minimal risk. The moment it can freeze a customer's account or block payments on its own, it restricts access to an essential service and becomes high risk.
The shift happens when the model gains authority to act on customers without a human in the loop.
Customer chatbot
A chatbot that answers questions about fees and opening hours is limited risk — it just needs to disclose that it is AI. If the same chatbot starts giving binding advice on mortgages or rejecting applications in the conversation, it influences access to credit and moves into high-risk territory.
The shift happens when the conversation starts producing decisions, not just information.
Marketing personalization
Recommending a savings account based on a customer's stated interests is ordinary, permitted personalization. But targeting people in financial distress with high-interest loan offers — exploiting a known vulnerability — is prohibited manipulation.
The shift happens when personalization exploits vulnerability instead of serving a stated need.
Employee productivity tools
An AI assistant that drafts internal reports is minimal risk. If the bank instead uses AI to infer employees' emotions from their voices or webcams during work, it crosses into a banned practice under the AI Act.
The shift happens when a tool stops assisting employees and starts monitoring their emotions.
Rule of thumb
Ask three questions whenever a system changes: Does it now affect customers directly? Can it act without human review? Does it touch credit, employment, or essential services? A yes to any of these usually means the risk level has moved up — and the compliance work needs to move with it.