Ready reckoner
Governance stack
| Layer | Instrument |
|---|---|
| Law | EU AI Act (risk tiers + roles), GDPR, sector rules |
| Management system | ISO/IEC 42001 |
| Risk framework | NIST AI RMF (Govern, Map, Measure, Manage), ISO/IEC 23894 |
| Impact assessment | ISO/IEC 42005, DPIA, FRIA |
- EU AI Act: prohibited practices · high-risk (e.g. recruitment, credit scoring) · transparency duties (chatbots, synthetic content) · minimal. Duties depend on role: provider vs deployer.
- Digital Omnibus moved Annex III high-risk deadlines to 2 Dec 2027; prohibitions, AI literacy (Art. 4, since 2 Feb 2025) and the GDPR already apply.
- Policy (intent) → standard (measurable) → baseline (configuration) → procedure (steps) → guideline (advice).
- Keep AI visible: inventory, intake workflow with a fast lane, procurement and discovery checks.
Responsible AI
- Principle → requirement → control → evidence. Vendor claims are not evidence.
- Removing a protected attribute does not remove bias: proxies. Feedback loops amplify.
- Impact ratio = less-favored rate ÷ most-favored rate; < 0.8 → investigate.
- Equal opportunity = equal TPR; equalized odds = equal TPR and FPR. Not all fairness metrics can hold at once → documented policy choice.
- Transparency (system) · explainability (decision) · interpretability (internals). Detailed scores help attackers.
- ISC2 Code of Ethics canons apply in order; raise harm through proper channels.
Risk
- Risk statement: threat source → vector → asset → event → impact, with owner and review date.
- ALE = SLE × ARO. Control justified if ALE reduction > annual cost.
- Tiering by blast radius and autonomy; regulation sets the floor.
- Treatments: mitigate, avoid, transfer (cost moves, accountability does not), accept (formal, owned by business, time-limited).
- Reassessment triggers defined in advance.
Privacy and IP
- Personal data lives in training data, indexes, prompts, outputs, logs, memory, weights and vendor copies. An embedding index is a copy.
- Reuse for training = new purpose → compatibility, lawful basis, DPIA. Public ≠ free to use.
- IP: can we train on it · what does the model license allow · who owns the output.
Third parties and supply chain
- Read assurance for scope, exceptions, complementary user entity controls.
- Contract: data use, change notice and pinning, incident notification, audit, sub-processors, IP indemnity, regulatory support, exit.
- Notification windows must fit yours: CERT-In 6 h · NIS2 24 h · GDPR 72 h.
- AIBOM turns "are we affected?" into a query. Confidential computing protects from host/operator, not app-layer attacks.
Accountability and awareness
- Three lines (own, oversee, assure) need independence.
- Named authority to intervene, override, suspend; emergency stop works outside the model and is tested.
- Shadow AI → approved alternatives + intake, not blanket bans.
- Training by role; measure with Kirkpatrick (completion is not a level). Zero reported incidents is a question.