Practice 04 / 07
Decide whether an AI feature is ready to release.
We review what the system can access, test how it handles sensitive information and document limitations for the people approving its use.
01Challenges addressed
Teams adopt models, copilots and LLM features faster than they are inventoried, risk-assessed or approved.
AI features need tests that cover both the application and the model’s behaviour, including how sensitive information is handled.
The EU AI Act introduces risk-based obligations, and regulators in other regions are issuing AI guidance.
Is the model accurate? Is it biased? Can we explain its decisions? Committees need documented answers.
02Services & capabilities
Open a service to see the scope it covers.
03Methodology
A conceptual sequence. Each engagement is scoped to its objectives, and not every engagement includes every step.
Step 01
Identify AI systems in use and in development, their owners, data and purposes.
Step 02
Assess risk and regulatory classification — including EU AI Act risk levels where relevant.
Step 03
Establish policies, frameworks and risk registers aligned to ISO/IEC 42001, ISO/IEC 23894 and the NIST AI RMF.
Step 04
Test AI and LLM applications against the OWASP Top 10 for LLM Applications and MITRE ATLAS.
Step 05
Evaluate accuracy, bias, explainability, robustness and data protection, documented for your governance committee.
04Engagement outputs
Typical outputs — the exact set depends on the services in scope.
05Standards & frameworks
References indicate the standards we work with. Which apply to you is confirmed during scoping.
06Related practices
Practice 03
Know where personal data lives, moves and leaves.
Practice 01
Find and prioritise weaknesses in the systems you depend on.
Practice 05
Give leadership a clear view of security risks, who owns them and which decisions need funding.
Contact
Tell us what you need to assess or prepare for, and any deadline you have in mind.