Think of one AI system you build, buy or use. Answer 8 short questions about it. 0 of 8 answered.
1. Impact if it goes wrong
If the AI gives a wrong, unsafe or manipulated answer, what is the worst realistic outcome?
Hardly any effect: a minor inconvenience Limited, recoverable: some rework or a small loss Significant: real financial, operational, safety or service impact Severe: widespread, safety-critical or possibly irreversible
2. How much it can do on its own
What can the AI do before a person steps in?
It only gives advice or text; people do everything It gives recommendations that people act on in a controlled process It starts consequential actions (payments, approvals, changes) with little human checking It plans and carries out multi-step consequential actions by itself
3. Sensitivity of the data
What kind of information does the AI handle?
Only public or non-sensitive information Internal business information or ordinary personal information Sensitive, confidential, financial or health information Highly sensitive: biometric, classified or critical-infrastructure information
4. Scale and reach
How widely and how often is the AI used?
Small internal use, low volume Controlled external use, or a moderate number of users Large-scale, frequent or customer-facing use Population-scale, national or critical-sector use
5. How reversible the harm is
If the AI gets something wrong, how easily can the effects be undone?
Easily reversed Recoverable with some effort Difficult, costly or time-sensitive to reverse Possibly irreversible physical, operational or systemic harm
6. How critical the function is
What does the AI support?
A non-critical supporting task An important business function A core or regulated operation Critical infrastructure, a safety-critical or national-security function
7. Exposure to attackers
Who can reach the AI, or feed it input?
Isolated or internal only Controlled external access: known users or partners Internet-facing or publicly accessible Actively targeted, or used in a high-threat environment
8. How hard it is to understand and audit
How easy is it to explain and check what the AI does?
Simple and transparent behaviour Some complexity, but bounded Complex, or depends on partly opaque components (for example a third-party model) A black box, or behaviour that emerges unpredictably
Anything that may need a higher tier? (optional)
The guidelines say these can justify a higher tier whatever the score is. Tick any that apply.
The AI performs safety-critical or consequential actions on its own It controls, or materially affects, critical infrastructure It can cause irreversible physical, operational or systemic harm It has privileged access to production systems, sensitive information or critical business processes It can run code, make transactions or send external communications on its own It operates in a high-threat or actively contested environment It cannot be reliably interrupted, rolled back or put into a safe state
What kind of AI is it? (optional)
Predictive or classic machine learning Generative AI (a chatbot or text generator) Retrieval-Augmented Generation (an AI answering from your documents) Agentic AI (the AI uses tools or takes actions) Several AI agents working together Images, audio or video as input or output Autonomous or physical systems
AI TātparyaSign in