Designing AI into a bank

AI in retail banking has had three lives. The first was rules-and-classifiers, used silently for fraud, AML, and credit scoring. The second was conversational, mostly chatbots that promised more than they delivered. The third, ongoing now, is generative and assisted, where models can summarise, explain, propose, and complete drafts inside the workflow. The third life is the one redrawing job descriptions in the bank, not just inside design but across operations, risk, and product.

For a senior designer at ENBD, the working assumption should be that AI is already in the stack and the question is how to surface it well. The temptation to put a glowing button in a corner labelled "Ask AI" is real and almost always wrong. The right placements are quieter: an explanation that appears beside a transaction the customer does not recognise, a summary at the top of a long statement, a draft of a dispute that the customer can edit, an automatic categorisation of spend that can be corrected, a search that understands "the rent transfer last month" without forcing the customer to learn a query language.

Where AI earns its place

Three families of use cases tend to deliver real value in retail banking. The first is comprehension assistance: summarising statements, explaining transactions, surfacing answers buried in the fine print. The second is journey acceleration: pre-filling forms, drafting messages, suggesting recipients, flagging duplicates. The third is risk and safety: detecting unusual transfers, scoring incoming requests for fraud, monitoring sessions for compromise. Each has a UX shape. Each fails differently.

The states that matter

AI features have a wider set of UX states than most others. A button is pressed or not, but a model can be confident, partially confident, working in the background, missing data, or simply wrong. Designing those states is most of the job.

The trust ladder

Customers do not extend AI a uniform amount of trust on day one. They extend it incrementally as the system proves itself. A useful mental model is a ladder of permissions. At the bottom, AI summarises information the customer can see anyway. One step up, AI categorises spend or detects duplicates, with the customer correcting as needed. Higher still, AI drafts messages and pre-fills forms, with the customer reviewing before submission. Higher again, AI takes small actions on the customer's behalf within tight rails (move spare change to savings, decline a known scam transaction). Only at the top does AI act with significant agency and money at stake, and that step should be reached deliberately, with explicit consent and a clear off-switch.

Bilingual AI is harder than monolingual AI

The UAE customer base is bilingual at minimum, often trilingual. AI features that work in English and not in Arabic are unfinished by definition. Beyond translation quality, there is the harder problem of cultural and tonal calibration. Arabic registers vary across dialects and contexts; a model that treats them as a single language loses its voice. The right approach is to test AI features in Arabic and English from the first sprint, with native speakers, in real customer scenarios, on real devices.

Compliance is your design partner

An AI feature in the bank touches data, model, and decision risk. The right partners are not lawyers blocking your work; they are model risk colleagues, fraud, legal, and Sharia (where applicable) co-designing the safe envelope. Bring drafts early. Document where the data goes, what the model can and cannot say, what is logged, and what the customer can see. Done well, this becomes a public-facing trust statement. Done poorly, it becomes a leak waiting to happen.