What banks in the UAE need to know about AI model risk

UAE banks already run model-risk files: IFRS 9, credit scoring, pricing, fraud. An "AI model" is not a new universe. It is a model whose claim is harder to reconstruct, and whose training data may not match the book in the Emirates.

Start from the file you already have

If the system approves credit, prices a policy, or ranks alerts, it belongs in the same inventory as the older scorecards. Give it an owner, a claim, a monitoring window, and an exit condition. A vendor slide deck is not a model card.

The local book is the test

A model trained on US or European bureau data will look accurate on a vendor demo and still misfire on thin-file residents, salary-transfer customers, or Arabic-language input. The assurance question is whether it works on the population you serve, not whether it won a Kaggle benchmark.

Language is a model-risk issue

Chat interfaces, document extraction and collections workflows that "also work in Arabic" need an Arabic test, including dialect and mixed Arabic-English. That is why the lab's second tool is an Arabic and African-language eval suite, after the credit toolkit.

What to ask a vendor

• What decision does this take, and for whom?
• What data was it trained on, and does that match our book?
• What are the approval or error gaps across nationality, gender and language?
• What happens when a feature is missing or a prompt is hostile?
• Can a second team rerun the test on our data?

Where SAFE Labs sits

We run the tests and write the finding. We do not sell the model. Editorial control stays with the lab; no client may exceed 40% of revenue. AI assurance · Request a review