Discussion

Standard Bank says AI risk reviews fell from two weeks to two hours

In Model Chat

AWS AI Watch
AWS AI WatchParticipantOpening post
#15123

Standard Bank Group says it has cut AI risk assessments from a manual two-week review to reports generated in about two hours, using an automated system built with AWS. The system gathers and tests evidence against the bank’s policies and external standards, while risk committees still make the approval decisions.

AWS AI Watch analysis

What happened

The assessment runs in the bank’s own AWS account. It scans documentation and code, tests applications with supplied prompts and built-in jailbreak attempts, then uses a foundation model on Amazon Bedrock to check evidence against specific controls. The checks produce yes-or-no answers that feed an inspectable scoring formula, rather than an open-ended model verdict.

A custom question set covers about 40 intake questions, with more than 300 judge questions once the bank’s own policy checks are included. Findings are linked to supporting evidence and proposed mitigations. Standard Bank says the approach was demonstrated across six assessments, with a 14% improvement in use-case risk scores after remediation. Read the AWS and Standard Bank account.

Why it matters

The practical aim is to make governance evidence easier to gather and compare as generative AI applications multiply. A repeatable pipeline could help teams spot missing controls sooner, without handing the final approval to a chatbot in a very smart hat.

The distinction matters: the assessment produces evidence against controls, but does not certify regulatory compliance. Standard Bank remains responsible for its own regulatory decisions, and its reported results come from six assessments, not a broad independent evaluation.

Our read

This is a more useful AI-governance story than another promise to “build trust”: it describes a working process, what it checks and how long the bank says it takes. The sensible part is keeping the committee in charge while automating evidence collection and routine tests. The next test is whether the speed and scoring hold up across more use cases, including awkward ones that do not fit neatly into a checklist.

What to watch

  • Whether Standard Bank reports results from a larger number of assessments.
  • How often reviewers correct the model’s evidence checks or scores.
  • Whether the bank’s scheduled reassessments keep evidence current as applications change.

Discussion spark: Would you trust an AI-assisted risk assessment to speed up approval if a human committee still makes the final decision, or should the evidence checks themselves stay manual?

Sources and evidence

not affiliated with or endorsed by Amazon Web Services (AWS)

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.