BuzzWhy.
ai · SG

UOB’s MERaLiON AI for Singapore bank calls: what could change for customers?

UOB is exploring AI prompts and post-call checks for relationship managers. Here is what the MERaLiON announcement confirms and what Singapore customers should still wait to learn.

UOB wants to help the person on the call

On 18 November 2026 in Singapore, UOB announced that it was exploring a multilingual assistant powered by A*STAR's MERaLiON and signed a memorandum of understanding with A*STAR's Institute of Advanced Intelligence and Computing. The bank described a system for its relationship managers. It could surface relevant product or risk information during a conversation and help review what was said afterwards. UOB says managers would remain accountable for the advice they give.

The practical change is in the manager's workflow. A Singapore customer who switches between English and Mandarin during a call might eventually get a clearer follow-up because the system could help the manager keep track of the question. That is an illustrative scenario. The announcement does not establish a customer launch date, coverage for any particular product, or that the tool is already running on all UOB calls.

Why switching languages changes the job

A standard transcript records words. A live assistant has to work out whether a sentence is a question, a correction or a request to confirm a prior point. A short exchange may switch languages while leaving product names or figures in English. IMDA's technical note on MERaLiON describes support for code-switching and regional forms of speech, including Singlish. That helps explain the model choice. Actual performance on bank calls still needs evidence from the bank's deployment.

Consider a fictional caller who starts in English, asks in Mandarin whether a charge applies only after the first year, then repeats a date in English. A useful prompt would identify the unresolved question and point the manager to the applicable terms. An inaccurate summary could turn that question into a statement that the caller accepted a charge. Both transcripts might look polished. Their downstream consequences differ considerably.

What a useful post-call review would check

UOB described a post-call review that could use evidence from the conversation to assess disclosures and support staff coaching. A sensible evaluation would compare the new workflow with human review of the same calls. A reviewer should be able to trace a flagged issue back to a specific stretch of audio, distinguish who said what, and correct a summary before it affects any follow-up. These are proposed evaluation criteria. UOB did not publish results for them in its announcement.

Average transcription accuracy alone would miss the costly errors. A pilot could separately measure missed disclosure flags, false flags, mistakes on amounts and dates, and results on calls where the speaker changes languages. Review time matters too. More alerts offer little value if managers or reviewers spend longer sorting through them than they did with the original process.

What to look for in UOB’s next update

The next useful details for Singapore customers are which teams and call types will use the tool, how customers will be notified, and how a customer can correct an inaccurate follow-up. The November release also describes regional ambitions. Language support in the underlying model does not establish availability in each market. Until UOB publishes customer-facing terms and deployment details, the evidence supports describing this as a proposed tool for relationship managers with further evaluation ahead.

Explore the original sources

  1. UOB announcement on the MERaLiON collaboration, 18 November 2026 ↗
  2. IMDA technical annex on Singapore’s national multimodal language models ↗