TL;DR: Prohibited practices are banned (for example social scoring, manipulation causing significant harm, untargeted facial-image scraping). High-risk systems are safety components of regulated products (Annex I) or listed uses (Annex III: biometrics, critical infrastructure, education, employment, essential services and credit, law enforcement, migration, justice). Transparency duties apply to chatbots and synthetic content. Everything else is minimal risk. A high-risk provider must run a risk management system, ensure data governance, produce technical documentation, enable logging, give instructions for use, design for human oversight, ensure accuracy, robustness and cybersecurity, operate a quality management system, pass conformity assessment, affix CE marking, register, and do post-market monitoring and serious-incident reporting.
How to approach it
Explain classification by intended use, list the tiers with examples, then go through high-risk provider obligations translated into engineering work, and close with dates after the 2026 Digital Omnibus.
A strong answer
- 1
1 · Prohibited
Banned practices
- Social scoring, harmful manipulation, untargeted face scraping
- 2
2 · High-risk
Annex I products, Annex III uses
- Risk management, data governance, documentation, logging
- Human oversight, accuracy, robustness, cybersecurity, conformity assessment
- 3
3 · Transparency
Chatbots and synthetic content
- Tell users; mark generated outputs
- 4
4 · Minimal
Everything else
- Voluntary codes
Classification is by intended purpose. The same model can be minimal risk in a writing assistant and high-risk in a hiring tool. Annex III uses can escape high-risk classification if the system performs only narrow procedural or preparatory tasks and does not materially influence decisions, but systems that profile individuals remain high-risk; providers must document that assessment.
High-risk provider obligations.
| Obligation | Engineering form |
|---|---|
| Risk management system (Art. 9) | Lifecycle risk process: threat models, evaluations, mitigations, residual risk |
| Data and data governance (Art. 10) | Documented, relevant, representative datasets; bias examination and mitigation |
| Technical documentation (Art. 11, Annex IV) | Design, data, training, evaluation, monitoring documentation |
| Record-keeping (Art. 12) | Automatic event logs over the lifetime |
| Transparency to deployers (Art. 13) | Instructions for use: capabilities, limitations, accuracy, oversight measures |
| Human oversight (Art. 14) | Ability to understand, monitor, override and stop |
| Accuracy, robustness, cybersecurity (Art. 15) | Declared metrics; resilience; protection against data poisoning, model poisoning, adversarial examples, confidentiality attacks |
| Quality management system (Art. 17) | Processes for compliance, design control, testing, data management, incident handling |
| Conformity assessment, CE marking, EU declaration, registration | Pre-market procedures and database registration |
| Post-market monitoring and serious incident reporting | Monitoring plan; incident reporting to authorities |
Deployers of high-risk systems have their own duties: use per instructions, human oversight by competent people, monitoring, keeping logs, informing affected workers, and for some deployers a fundamental rights impact assessment.
Timing (checked November 2026). The Digital Omnibus (Regulation (EU) 2026/1744) moved Annex III high-risk obligations to 2 December 2027 and Annex I obligations to 2 August 2028. Prohibitions have applied since 2 February 2025 and GPAI model obligations since 2 August 2025.
What interviewers probe next
- "Is a customer-service chatbot high-risk?" Usually not; it has transparency duties. It becomes high-risk if it decides access to essential services, credit or similar.
- "How does Article 15 change your security programme?" It turns AI-specific attack resilience into a documented legal requirement for high-risk systems: threat models, testing and controls for poisoning, adversarial inputs and confidentiality attacks.
- "What makes a deployer become a provider?" Putting its name on the system, making a substantial modification, or changing its intended purpose so it becomes high-risk.
What if?
What if · interview follow-ups
Hold your answer as the constraints change
A startup offers two products in the EU: an AI system that screens job applicants and a customer service chatbot.
Start here
What are the risk tiers?
Which tier has bans, which has heavy obligations?
Prohibited practices are banned (for example social scoring, manipulation causing significant harm, untargeted facial-image scraping). High-risk systems are safety components of regulated products (Annex I) or listed uses (Annex III). Transparency duties apply to chatbots and synthetic content. Everything else is minimal risk.
Carry it forwardFour tiers: prohibited, high-risk, transparency, minimal.
Weak answerAssuming every AI system is high-risk.
Constraint changes
Which tier is the applicant screening system?
Is employment listed in Annex III?
High-risk, because employment, including recruitment and selection, is an Annex III use. As provider, the startup must run a risk management system, ensure data governance, produce technical documentation, enable logging, write instructions for use, design for human oversight, ensure accuracy and cybersecurity, operate a quality management system, pass conformity assessment, register and monitor after release.
Carry it forwardEmployment uses are high-risk under Annex III.
Weak answerTreating hiring tools as minimal risk because a human decides.
Constraint changes
Which tier is the customer service chatbot?
Does it make decisions in a listed area?
Usually the transparency tier: users must be told they are interacting with AI unless that is obvious. If the chatbot generates synthetic text or media, marking duties may apply. It becomes high-risk only if used for a listed purpose, such as deciding access to essential services.
Carry it forwardChatbots mainly carry transparency duties.
Weak answerIgnoring the chatbot because it is low risk.
Constraint changes
What if a customer configures the chatbot to screen loan applicants?
Does classification follow the product or the use?
Then the use changes the classification: creditworthiness assessment is a listed high-risk use. Whoever puts the system into service for that purpose may take on provider obligations. Define intended use in contracts and instructions, and monitor for uses outside it.
Carry it forwardClassification follows the intended use.
Weak answerAssuming a product's tier is fixed regardless of use.
Constraint changes
When do the high-risk obligations apply?
How long does building a conformity file take?
As checked in November 2026, the Digital Omnibus moved Annex III high-risk obligations to 2 December 2027. That is time to build, not time to wait: conformity assessment needs documentation, testing and a quality system that take months to put in place. Confirm current dates with official sources.
Carry it forwardUse the time before application to build evidence.
Weak answerDeferring all work until the deadline.
Now defend it
What would you tell the startup to do?
Classify by intended use: the screening system is high-risk under Annex III and needs the full provider programme (risk management, data governance, documentation, logging, oversight, accuracy and cybersecurity, quality management, conformity assessment, registration, monitoring); the chatbot needs AI disclosure and any marking duties. Control intended use in contracts, monitor for reclassifying uses, and track dates through official sources.
How it's scored
Weak: Lists tiers vaguely, or classifies by model type.
Hire: Correct tiers with examples, classification by intended purpose, and the main provider obligations.
Strong Hire: Translates obligations into engineering artefacts, highlights Article 15, covers deployer duties, the Annex III exception, role changes and current dates.
Common mistakes
- Calling every LLM application high-risk.
- Ignoring deployer obligations.
- Quoting the original 2026 high-risk date as current.