How to Write an AI Risk Appetite Statement
An AI risk appetite statement is a practice, not a statute: Article 5 prohibited is the hard floor, and Annex III high-risk needs executive or board sign-off.
An "AI risk appetite statement" is a board-level practice, not a requirement in the EU AI Act. The Act sets a legal floor; a risk appetite statement is how a company decides, above that floor, which AI it will and will not pursue, and who has to say yes. It turns a vague instinct ("we should be careful with AI") into a written rule that a product team can act on without escalating every decision. Done well, it speeds good projects through and stops bad ones early.
The discipline is to anchor it in the law where the law is hard, and use management judgment where the law is silent. Under Regulation (EU) 2024/1689 the hard floor is Article 5; everything above it is your call.
The hard floor: Article 5 prohibited uses are off-limits, full stop
Start with the one part that is not a matter of appetite. The Article 5 prohibited practices — social scoring, manipulative or vulnerability-exploiting techniques, emotion recognition in the workplace and education (Article 5(1)(f)), untargeted facial-image scraping, certain real-time public biometric identification — are banned, and have been since 2 February 2025. From 2 December 2026 the Digital Omnibus adds two more: AI that generates non-consensual intimate deepfakes of identifiable people ("nudifier" apps) and AI that generates child sexual abuse material (Article 5(1)(ba) and (bb)). No risk appetite, however bold, can authorise them; a breach is the top penalty tier, €35,000,000 or 7% of worldwide annual turnover (Article 99(3)). So the first line of any honest statement is simple: we do not build, buy or deploy any Article 5 prohibited practice. That is a legal floor stated as policy, not a choice.
The board-sign-off tier: Annex III high-risk
Above the floor sits the territory where appetite actually operates. Annex III high-risk systems (recruitment, credit scoring, biometric ID, essential-service eligibility, and the rest) are lawful but carry the full requirement stack and real liability. A sensible statement makes these a named board-or-executive decision: no high-risk system goes live without explicit sign-off from the accountable executive, on the back of a completed classification, the Article 9 risk assessment, and — where it applies — an Article 27 FRIA. You are not banning high-risk AI; you are refusing to let it ship by default.
| Tier | Appetite stance | Decision rights |
|---|---|---|
| Prohibited (Art 5) | Zero — hard floor, not a choice | Automatic no |
| High-risk (Annex III) | Case by case, with controls | Accountable executive sign-off |
| Limited (Art 50) | Acceptable with transparency duties met | Product owner |
| Minimal | Acceptable, recorded | Product owner |
Thresholds and escalation
Appetite is only useful if it sets thresholds that trigger escalation. Write them in concrete terms a team can apply at intake:
- Any use that could fall in Annex III → mandatory classification and central review before build.
- Any system that profiles natural persons → treat as high-risk (it always is in Annex III) and escalate.
- Any use touching vulnerable groups, biometric data, or automated decisions with legal effect → escalate regardless of tier.
- A change of intended purpose or a substantial modification → re-run classification before shipping.
Pair each threshold with a clear route: who is notified, who decides, and what evidence (classification record, Article 9 assessment, FRIA) must accompany the decision.
Tie it to the Article 9 risk-management system
A statement that lives in a slide deck is theatre. Connect it to the live machinery: the Article 9 risk-management system for high-risk systems is the continuous process where appetite is actually tested — the risks you said you would not accept are the ones the Article 9 process must drive to mitigation or to a no-go. Feed the outcomes into Article 72 post-market monitoring, so a system that drifts outside appetite once live is caught and re-decided. Appetite, in other words, is the policy; Article 9 and Article 72 are where it bites.
Worked example: Lindqvist Finans AB, ~95 employees, ~€19M turnover
Lindqvist, a Gothenburg lender, writes a one-page statement. Floor: no Article 5 practice — explicitly no workplace emotion recognition and no social scoring. Board sign-off: its credit-scoring model is Annex III point 5(b) high-risk, so it ships only on the CRO's signature, backed by the Article 9 assessment and an Article 27 FRIA (creditworthiness deployers owe one). Thresholds: any new use touching applicant data or automated lending decisions escalates to the CRO; marketing and internal-productivity AI (minimal/limited) stays with product owners. Escalation: a proposed "behavioural trust score" pulling unrelated social data is flagged at intake as straying toward prohibited social scoring and is killed before any build. The statement gives Lindqvist a fast yes for low-risk AI and a hard, documented process for the rest.
How Confir helps
Confir operationalises the statement: it screens every use against Article 5 at intake (the hard floor), classifies under Article 6 and surfaces the Annex III high-risk systems that need executive sign-off, and records the Article 9 and Article 27 evidence the appetite requires before go-live. The engine is deterministic and rule-based, so the line between "auto-approve" and "escalate" is applied the same way every time, with the rule that fired shown in plain English — appetite enforced by a consistent process, not by memory.
Related guides
- AI governance, end to end
- Building an AI governance operating model
- Board reporting on AI risk
- The three lines of defence for AI
- Prohibited vs high-risk AI
- How to classify your AI system
Manage your EU AI Act compliance in one place
Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.
Start free trial →