How to identify and assess AI risks: the six-phase method

Why identification comes first

Every AI risk framework asks for it. The EU AI Act's Article 9 starts the risk management system with the identification and analysis of known and reasonably foreseeable risks. ISO/IEC 42001 requires an AI risk assessment in clause 6.1.2, and ISO/IEC 23894 gives guidance on running one. None of them tells you how to find the risks in a real organisation, with its vendors, its spreadsheets and its pilots that quietly became production. A risk that is never identified is never scored, never controlled and never reported.

The method comes from bank risk identification, where regulators expect a complete, reconciled and documented risk inventory. It was rebuilt for AI with a new taxonomy, a new register and a fifth impact dimension, harm to individuals.

Phase 1: Foundation setting

Map the standards and laws that apply, set governance and roles, adopt the taxonomy, and set the context. Build the AI system inventory: every system, whether it is bought or built, its model and version, its purpose, its users, the persons it affects, its role and classification under the AI Act, and its owner. Set the risk criteria. Then take the register as a starting universe, so that the work does not begin from a blank page. Chapters 2 to 5.

Phase 2: Dual-track identification

Run two tracks, because each misses what the other sees. Top-down, senior people work through workshops, SWIFT (a structured what-if technique), scenarios, Delphi rounds and red-team review. One SWIFT prompt asks what happens if the vendor changes the model behind the service and does not tell you. Bottom-up, the people who operate each system complete the AI use-case template: a five-line system header and eleven fields, from the underlying drivers to the data quality rating. Ten specialist sub-processes cover what a general template cannot: data protection (DPIA), security (OWASP), fairness testing, model validation, third-party due diligence, content and intellectual property, safety for physical systems, accessibility, environmental impact, and human factors. Chapters 6 and 7.

Phase 3: Reconciliation and the portfolio view

Merge the two lists in five steps: match, gap, challenge, resolve and record. A risk that only one track found is a blind spot of the other, and the disagreement log keeps the arguments that scoring will settle. Then step back to the portfolio: the enterprise AI risk map and the concentrations behind it, such as one foundation-model provider under many use cases. Chapter 8.

Phase 4: Assessment

Score impact on five dimensions (financial, regulatory and legal, reputational, harm to individuals, and customer and operational), and let the highest set the score. Score the likelihood of the harm, not of the technical error, per stated unit of exposure: per output, per decision, per person affected or per year. Record vulnerability and speed of onset beside the score, and rate the evidence behind every score with a data quality rating. Score inherent and residual risk; a control that has not been tested reduces nothing. A line is material if its inherent score is 15 or more on the 5 × 5 grid, or if its impact is 5. Chapter 9.

Phase 5: Interaction

A heat map cannot show links. The interaction matrix records, for each pair of material and linked lines, whether one triggers the other, amplifies it or shares a driver with it. Bow-ties set causes, controls and consequences around each material risk, and the matrix shows cascades: a silent vendor update amplifies automation bias, which triggers decisions that cannot be explained. Look for hidden concentration: two suppliers may run on the same underlying model. Chapter 10.

Phase 6: Documentation

Keep the result as a living AI risk inventory: the fourteen fields of the bank method plus five AI columns (lifecycle stage, role, persons affected, regulatory hooks and ISO/IEC 42001 Annex A references). Each material risk gets a profile that a board member can read in five minutes, and every change leaves an audit trail. Chapter 11.

After the six phases

The register feeds the rest of the AI management system: fundamental rights impact assessments under Article 27 of the AI Act, data protection impact assessments, the ISO/IEC 42001 Statement of Applicability, post-market monitoring and the board report (Chapter 12). It is kept current by a cycle (Chapter 13):

Chapter 14 covers technology, including using AI to find AI risk, where the tool proposes and a person decides. Chapter 15 maps the method to the EU AI Act and to rules in the UK, the US, China, Korea and elsewhere. Chapter 16 tests the method against public AI failures.