AI-risk framework
The four AI risk levels, with examples from everyday work and cases that show how a change in use can change the risk level.
The Four AI Risk Levels
The EU AI Act looks at what the system is intended to do and where it is used. The same technology can have different obligations in different settings.
Minimal Risk
Everyday uses that do not fall under the AI Act’s prohibitions, high-risk categories or specific transparency duties.
A spam filter or an assistant that drafts internal meeting summaries.
Limited Risk
Systems that require transparency so users know they are interacting with AI.
A retailer’s AI chatbot that answers questions about returns and delivery.
High Risk
Uses listed in the Act, such as screening job applicants, assessing students, or serving as a safety component in certain regulated products.
A recruitment tool that ranks job applicants and influences who is interviewed.
Unacceptable Risk
Specific AI practices prohibited by Article 5 because of the harm they can cause.
Using AI to infer employees’ emotions from their faces or voices during routine performance reviews.
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.
Recruitment assistant
An assistant drafts a generic job advert without targeting people or evaluating applicants. If it is then used to rank applicants and shape the interview shortlist, it enters a listed high-risk use.
It now helps decide who progresses, rather than just drafting text. A human making the final decision does not automatically make it low risk.
Student support chatbot
A chatbot that answers questions about course schedules is a transparency case. If it starts scoring admissions interviews to help decide who gets a place, the admissions use is high risk.
It evaluates applicants for access to education. Chatbot transparency duties can still apply.
Personalised offers
A shop recommends products from a shopper’s stated preferences. It crosses a line if it exploits an older person’s vulnerability to materially distort their choices and pressure them into a purchase likely to cause significant financial harm.
The prohibition turns on exploitation, distortion of behaviour and actual or likely significant harm, not personalisation alone.
Employee productivity tools
An assistant drafts internal reports. Using AI instead to infer employees’ emotions from biometric data such as facial expressions for routine performance reviews is prohibited.
The system moves from helping staff draft text to inferring emotions in the workplace for a non-medical, non-safety purpose.
Rule of thumb
If the purpose changes, classify the new use. Check whether it matches a prohibited practice, a listed high-risk use or a transparency duty. Greater autonomy is a reason to reassess, not a risk category on its own.
Sector notes
The general examples apply across sectors. DORA and NIS2 can add duties for organisations in scope, but neither determines an AI system’s risk level under the AI Act.
Finance & DORA
Credit scoring is different from fraud detection. Scoring a person’s creditworthiness is listed as high risk. Detecting financial fraud is expressly excluded from that entry. Automatically blocking a payment is a reason to review the impact, not enough by itself to classify the system as high risk.
DORA still applies to the ICT service. Before contracting for a vendor’s fraud monitoring service, an in-scope firm assesses the ICT risks and checks whether the service supports a critical or important function.
NIS2
Working in a critical sector does not make every AI tool high risk. At a water utility, using AI to draft shift notes is different from using it as a safety component in operating the water supply. The latter is a listed high-risk use unless an Article 6(3) exception applies.
NIS2 can apply even when the AI is minimal risk. If a compromised shift-note tool causes significant disruption to the utility’s service, the utility must assess it under the applicable national incident-reporting rules.