Standard Bank: AI risk reviews cut from two weeks to a two-hour report
AWS and Standard Bank say automated checks turn a two-week AI risk review into a two-hour report; the risk committee still approves. No cost is given.
At a glance
- Company
- Standard Bank Group · Financial services · Size not stated
- Team and use case
- Legal and compliance · AI governance
- Tools
- AWS Professional Services, Amazon Bedrock, Amazon Bedrock Guardrails, AWS Well-Architected Tool, AWS Step Functions
- Stage
- In production, one team
- Results
- Risk review of one generative AI use case: 2 weeks, manual → Report in about 2 hours (6 assessments)
- Use-case risk scores after remediation: 14% improvement (base not given)
- Yes-or-no checks per assessment: More than 300
- Cost
- Not disclosed: no figure for AWS Professional Services, AWS usage or the bank's build effort.
- People
- The risk committee still reviews and approves; reviewers confirm or amend findings before each gate. Reviewer numbers before and after are not disclosed.
- Evidence
- Vendor-reported
Published by the company selling the product, usually with the customer. Wins are selected; cost is usually missing.
What this leaves out: AWS published the post, co-written with the bank's CTO. Six assessments; the two-week baseline is described, not measured, and two hours covers producing the report, not reviewing it. No cost, error rate or overturned-finding count.
Standard Bank Group has automated the evidence-gathering part of its risk review for generative AI use cases. A review that was "traditionally manual" and took two weeks now starts from a risk report produced in about two hours, according to a post on the AWS blog dated 20 November 2026 and co-written by Duncan Macdonald, the bank's chief technology officer, and Tala Qraitem of AWS. The figure was "demonstrated across six assessments".
Approval has not moved. "Review and approval stay with the risk committee", the post says; the committee now decides on evidence collected the same way for every use case. The post also reports "a 14% improvement in use-case risk scores after remediation", without saying how many use cases that covers or what the scores were.
What the check does
- Intake. A questionnaire of about 40 questions, each covering a risk such as incomplete documentation or insufficient human oversight. Software drafts the answers from the use case's code and documents with links to the evidence; the delivery team corrects them.
- Testing. The pipeline calls the application with questions a typical user would ask, a built-in set of attempts to talk it out of its rules (jailbreaks) and off-topic questions it should decline. It repeats and rewords prompts to check the answers stay consistent.
- Scoring. An AI model on Amazon Bedrock answers more than 300 yes-or-no questions on whether specific evidence exists. Each maps to a clause in an international AI standard or the bank's own policy. A fixed formula the risk team can inspect turns the answers into a rating from Very low to Critical.
- Review. Each evaluation runs several times and inconsistent findings are flagged. Reviewers confirm findings before a gate decision and can amend an answer with a business justification.
- Policy changes. A tool generates check questions from internal policy documents, so the post says a policy change becomes a test change on every later assessment.
The assessment is optional at proof of concept and mandatory at the build stage, where the approval gate takes its results as formal evidence. It runs in the bank's own AWS account. Re-running it on a schedule, or when code or configuration changes, is planned and not yet live.
What the post leaves out
- Cost. No figure for AWS Professional Services, usage or the bank's build effort.
- The other side of the two hours. Two hours is the time to produce the report. The post does not say how long the delivery team spends correcting the drafted answers, how long reviewers take to confirm findings, or how long the committee now takes to decide.
- Accuracy. No count of findings reviewers overturned or risks the checks missed. The judge model is not named.
- People. It does not say how many reviewers did the two-week review or what they do now, beyond "judgment calls and policy design".
- Regulatory weight. The post says the assessment "doesn't certify compliance with any regulation"; the bank remains responsible for its own determinations.
Reading this at work? Get one story like it explained every morning.
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.
What to do if you own AI governance or risk sign-off
- Legal and compliance: go through your AI policy and mark each clause that can be tested as a yes-or-no question against a document, a code file or a live answer. Those are the clauses a tool can check; the rest stay with reviewers.
- Whoever runs the approval gate: log the elapsed days and reviewer hours for your next two manual reviews. Standard Bank's two weeks is a description, not a measured baseline, and you will need your own to judge any vendor's claim.
- IT: ask any vendor proposing this what share of automated findings human reviewers changed in its last deployments, and require every judgment to be stored with its evidence so an auditor can trace it.
Take this to work
Forward this
AWS and Standard Bank Group say an automated assessment cut the bank's risk review of a generative AI use case from a manual two weeks to a report produced in about two hours, across six assessments, with the risk committee still approving. The post gives no cost and no count of findings reviewers overturned, so time our own reviews before treating that as a benchmark.
Questions to ask
- After the two-hour report, how many hours do the delivery team, the reviewers and the committee spend before a decision?
- What share of the automated findings did human reviewers change or overturn in the six assessments?
- What did the consulting, the build and the usage cost, and what does one assessment cost to run?
- Which of our own policy clauses cannot be written as a yes-or-no check, and who reviews those?
- Will a regulator or our auditor accept model-scored evidence, and can every judgment be traced to its source?
The arithmetic
AI use cases reviewed a year × (reviewer hours per manual review − reviewer hours to correct and confirm an automated report) − cost of building and running the assessment
Standard Bank: six assessments, each a two-week manual review replaced by a report produced in about two hours. The post does not give reviewer hours before or after, the use cases reviewed a year, or any cost, so the net saving cannot be worked out from it.
Questions this article answers
How does Standard Bank assess the risk of generative AI use cases?
An AWS blog post co-written with the bank's chief technology officer says an automated assessment gathers evidence from each use case's documents, code and live responses, then an AI model answers more than 300 yes-or-no checks tied to international AI standards and the bank's own policies. The risk committee still reviews and approves.
How much time did Standard Bank save on AI risk reviews?
The post says a traditionally manual two-week review became a risk report generated in approximately two hours, demonstrated across six assessments. It does not give the time reviewers and the committee spend on the report afterwards.
What did Standard Bank's automated AI risk assessment cost?
Not disclosed. The post gives no figure for AWS Professional Services, AWS usage or the bank's own build effort.
Does the automated assessment certify regulatory compliance?
No. The post says the assessment produces evidence against controls, does not certify compliance with any regulation, and that Standard Bank Group remains responsible for its own regulatory determinations.
Amazon Connect Resolutions: $0.40 a case, with human approvals billed as steps
AWS has put its AI agents for refunds, disputes and claims on general sale at $0.40 a case. Steps past the fourth cost extra, a person's sign-off included.
Get the briefing by email
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.