Skip to content
context Tools and apps

Standard Bank Group automated generative AI risk assessment using Amazon Bedrock

only one source so far

Standard Bank Group has automated generative AI risk assessment using LLM-as-a-judge on the Amazon Bedrock platform. The system collects and tests evidence against more than 300 controls from AI standards, but the decision still rests with the risk committee.

Standard Bank Group, in collaboration with AWS Professional Services, has deployed an automated risk assessment system for generative AI applications. The system evaluates each new use case against international AI standards and the bank's own internal policies before the case is forwarded to the risk committee for approval. According to the company, the solution is not intended to replace human decision-making but to speed up the preparation of supporting materials — approval authority remains with the risk committee, which now works with consistent, automatically collected evidence instead of manually reading documentation and testing each application separately.

Technically, the solution runs entirely within Standard Bank Group's AWS account. Documents, code, and test results remain in the bank's Amazon S3 bucket, orchestration is handled by AWS Step Functions, and evidence evaluation is performed by a foundation model on Amazon Bedrock using the LLM-as-a-judge technique, where the model compares the evidence against defined criteria and returns a yes/no answer. According to the company, Amazon Bedrock does not share prompts or model outputs with model providers and does not use them for training; the bank controls access to the model via AWS IAM, just as with its other AWS workloads.

The process has four steps: a pre-assessment agent on AWS Fargate first automatically scans the code repository and documentation for the given use case and pre-fills answers in the tool's custom lens for the Well-Architected Tool (roughly 40 questions on risk areas such as missing documentation or insufficient human oversight), after which delivery teams review and complete the answers. This is followed by live testing of the application via a Lambda function or API, including built-in jailbreak attempts and consistency checks on answers using an embedding model. Each check is linked to a specific control from an international AI standard or framework, and after adding the bank's own standards, the total number of checks exceeds 300.

What changed

Why it matters

For companies in regulated industries (banks, insurance companies), this demonstrates a concrete approach to speeding up governance of new AI applications without bypassing approval processes — automation collects and pre-fills evidence, while the human decision at the risk committee is preserved. The architectural solution is also relevant: data and prompts remain in the client's account and are not shared with model providers, which addresses concerns about sensitive data leakage when deploying LLMs in compliance processes.

Relevant practical impact

What this means

01

For a business

Standard Bank Group sped up the risk committee's approval process for AI use cases by automating the collection and testing of evidence of compliance with AI standards using LLM-as-a-judge on the Amazon Bedrock platform.

Risks and compliance
What to decide Consider introducing automated mapping of evidence to AI standard controls for internal generative AI projects before submission to the risk committee, especially in regulated industries.
More business impacts →
AI governance automation AWS enterprise generative AI Risk assessment

Check the original

Event sources

only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.

1
AWS Machine Learning Blog primary source · first detected How Standard Bank Group automates AI risk assessment for generative AI