Standard Bank Hands AI Safety Checks To A Machine, Humans Still Decide
Standard Bank Group has found a way around the slowest part of rolling out AI at a big bank: instead of people reading documents and testing every new tool by hand, an AI system now gathers the evidence and grades each one for risk, while humans keep the final say.
Key Takeaways
- The assessment runs entirely in Standard Bank Group’s own AWS account
- It scores each generative AI use case against international AI standards and the bank’s own policies
- The workflow uses about 40 intake questions and more than 300 judge questions
- Final review and approval stay with the risk committee, with findings traced back to evidence and controls
AWS explained that Standard Bank Group worked with AWS Professional Services to automate evidence gathering and testing for generative AI approval gates, after a review process that depended on people reading documents and testing systems by hand.
The bank’s goal was not to hand decisions to a model. Review and approval still sit with the risk committee, but each decision now starts with evidence collected the same way across use cases before new systems reach customers or employees.
How The Assessment Works
The setup stays inside the bank’s environment. Documents, code, and test results remain in an Amazon Simple Storage Service (Amazon S3) bucket in Standard Bank Group’s own AWS account. The AWS Well-Architected Tool handles intake, AWS Step Functions orchestrates the pipeline, and a foundation model on Amazon Bedrock evaluates the evidence. Standard Bank controls access through AWS Identity and Access Management (IAM), and Amazon Bedrock does not share the bank’s prompts or outputs with model providers or use them to train models.
The workflow has four steps: a pre-assessment agent collects evidence, the pipeline examines that evidence and tests the application directly, the model evaluates the results against controls, and the system summarizes the outcome in a standard format for the people who act on it. The custom lens in the Well-Architected Tool holds about 40 questions, and Standard Bank added judge questions built from its own standards to bring the total to more than 300.
For each question, the model returns yes or no on whether specific evidence appears in the artifacts and in the application’s responses. Each risk gets a likelihood score based on the share of required evidence the model could not find, plus a fixed severity level that the bank’s risk team can adjust. Together, they produce a rating from Very low to Critical. Each missing piece maps to a mitigation and to the clause of the standard or policy it came from, so a risk officer can trace findings back to their source when explaining decisions to a regulator.
Also Read: AWS Uses Okta Logins To Stop AI Agents Going Rogue At Work
Why This Matters
Teams can now build a generative AI app in days, but each one still needs a risk committee’s sign-off before it reaches customers or staff. When that sign-off depends on people reading documents and testing by hand, governance becomes the slowest step, and banks either slow down or cut corners. Automating the evidence gathering, while keeping humans in charge of approval, is how a heavily regulated bank can move at AI speed without losing control.
Standard Bank did not need a new AI policy. It already had an established model risk practice for concerns such as explainability and fairness, and a change risk process that assesses each new use case according to its level of risk before it goes live.
What changed was the pace. Generative AI adds complexity because models and tooling change every 3 to 6 months, yet each use case still has to meet the same high security and compliance bar. Jörg Fischer, Group Chief Information and AI Officer, said responsible AI should “build and scale with confidence.”
What To Watch Next
The assessment is built so its output can be checked. Instead of asking the model for an open-ended opinion, it asks narrow yes-or-no questions worded to match specific controls, and a fixed formula that the risk team can inspect turns those answers into scores.
Each evaluation runs multiple times to measure how consistent the answers are, and any inconsistent findings are flagged in the review interface. Every judgment is stored with the evidence behind it so reviewers and auditors can trace how it was reached.
Every model interaction also passes through Amazon Bedrock Guardrails, which screen for harmful content and attempts to manipulate the model. AWS noted one limit up front: supported foundation models vary by AWS Region. At the same time, Amazon Bedrock’s single interface means the team can move to a newer judge model by changing a configuration setting instead of rebuilding the pipeline.
The bigger test is whether other banks copy the approach. Regulators increasingly want banks to show how they checked an AI system before using it, and a setup that ties every finding to the exact rule and evidence behind it gives risk officers a ready-made answer. If it works at Standard Bank, AI risk checks could become a standard product that cloud providers sell to every bank.
Read Next: Sumsub Targets AI Agent Risk With 3 Controls For Banks, Insurers And Marketplaces
