AAIR AI life cycle risk management
AI Life Cycle Risk Management is worth 21% of AAIR — around 19 of the 90 questions and the smallest of the three domains. It sounds like the most technical part of the exam, which is why candidates over-prepare it and under-prepare the two risk domains that together make up 79%.
You need to know what happens at each stage and what risk it introduces. You do not need to be able to perform any of it.
Design, development or procurement, and documentation
The first decision is build or buy, and the two carry different risk:
| Build | Buy or use an API | |
|---|---|---|
| Visibility | You see the data and process | Largely opaque |
| Control | You decide everything | Provider decides |
| Responsibility | Entirely yours | Shared, but accountability stays with you |
| Main risk | Doing it badly | Inheriting unknown properties, and silent change |
Documentation is named explicitly and is the artefact risk depends on: intended use, limitations, data sources, design decisions. Undocumented systems cannot be assessed, which makes missing documentation a risk finding in its own right.
Model training, testing and validation
What matters from a risk perspective:
- Training — who can initiate it, on what data, with what approval. An unauthorised retrain is a production behaviour change.
- Testing and validation — what each technique demonstrates. Validation against held-out data shows performance on unseen cases; adversarial testing shows behaviour under hostile input; fairness testing shows outcome disparity across groups. A system can pass one and fail another.
- Acceptance thresholds defined in advance. Testing without a pre-defined threshold produces a number, not a decision.
Implementation, maintenance and decommissioning
Implementation — deployment approval, staged rollout, fallback if behaviour is unacceptable.
Maintenance — the stage where AI diverges most from conventional systems. Behaviour changes three ways: retraining, a provider updating a hosted model, and drift. Only the first reliably appears in a change record.
Decommissioning — routinely overlooked and reliably examined. Retiring an AI system leaves three things behind:
- The model artefact, which encodes what it learned
- The training data, with its retention and privacy obligations
- The decisions already made, which may need to remain explainable and defensible long after the service is switched off
Data and asset management
Models and training data are assets on the register:
- Provenance — a defensible record of origin
- Integrity — detecting modification
- Classification — sensitivity of the data, and of the model that learned from it
- Retention and disposal — including that removing a record from a dataset does not remove its influence from a trained model
Sample questions
Question 1. An enterprise is deciding between fine-tuning an open-weights model in-house and using a hosted third-party model API for a regulated process. Which risk is MOST specific to the hosted option?
- A. The enterprise may lack the skills to manage compute capacity
- B. The provider may change the model, altering behaviour and invalidating prior validation
- C. Infrastructure costs may exceed the forecast
- D. Data labelling may require more effort than expected
Show answer
Answer: B
A hosted model can be changed by its provider, altering behaviour in production and invalidating validation evidence, which is the risk unique to not controlling the artefact. Compute skills and infrastructure cost apply more to the in-house option, and data labelling effort is common to both.
Question 2. A model passed validation against a held-out dataset before release and is used to screen job applicants. Which additional validation should the risk practitioner expect?
- A. Fairness testing showing outcome disparity across affected groups
- B. Load testing at peak application volume
- C. Usability testing with recruiters
- D. Penetration testing of the hosting environment
Show answer
Answer: A
Overall performance can conceal materially different outcomes between groups, so fairness testing across affected groups is the necessary additional validation for a system screening applicants. Load testing addresses capacity, usability addresses experience, and penetration testing addresses exploitability of the environment.
Question 3. An AI service supporting insurance pricing is being decommissioned. Which is MOST likely to be overlooked in the decommissioning plan?
- A. Cancelling the supporting software licences
- B. Releasing the hosting infrastructure
- C. Disposition of the model artefact and training data, and retaining the ability to explain past decisions
- D. Updating the asset register to mark the service retired
Show answer
Answer: C
Retiring the service leaves the model artefact, the training data and the record of past pricing decisions, all of which carry retention, privacy and defensibility obligations that outlive the system. Cancelling licences, releasing infrastructure and updating the asset register are routine steps rarely missed.
What to practise
Walk one AI system through all five stages and name the risk at each: design, training, validation, implementation, decommissioning. Five lines. Then answer the decommissioning question properly, because it is the stage nearly everyone skips and the one the exam likes.
Then stop, and put your remaining time into the governance and programme management domains — together they are 79% of the paper.