Quick Answer: UOB is exploring whether A*STAR's MERaLiON multilingual AI can help its relationship managers understand customers, handle languages across five ASEAN markets, and review calls for fair dealing. The project is at the exploration stage under a memorandum of understanding, not a finished rollout. Relationship managers stay accountable for the advice they give.

Most bank AI announcements promise faster service. UOB's latest targets a narrower problem: a customer who starts a sentence in Bahasa Indonesia, finishes it in English, and drops in a local expression. On 18 November 2026, UOB said it is the first bank to explore A*STAR's MERaLiON multilingual AI for this kind of conversation across Singapore, Malaysia, Thailand, Indonesia and Vietnam. Here's what the announcement covers, what it doesn't, and where the real risk sits.

What UOB Is Actually Piloting

As UOB's press release describes it, the first phase runs through Project AURA, the bank's programme to deploy advanced AI at scale. Relationship manager enablement is one of its priority domains.

The collaboration is still early. UOB and the ASTAR Institute of Advanced Intelligence and Computing (ASTAR IAIC) signed a memorandum of understanding to explore how MERaLiON's capabilities could be enhanced for safe, assured and scalable deployment. The bank's own wording is "looking at" and "explore." Nothing in the release says the tool is live with customers.

The solution is titled "Compliance Intelligence Engine and Evidence-Grounded Post-Call Compliance Review System." It has two main functions:

  1. During the call: prompts on considerations such as key product information and risk disclosures, based on interpreting questions across languages, accents and communication styles.
  2. After the call: an evidence-grounded review to check whether key information was communicated clearly and whether fair dealing standards were met.

Here's where most readers stop, and where the more interesting questions start.

How MERaLiON Handles Code-Switching

MERaLiON stands for Multimodal Empathetic Reasoning and Learning in One Network. The release says it supports multilingual processing and code-switching (switching languages within one exchange) across 11 ASEAN languages, including Malay, Bahasa Indonesia, Thai and Vietnamese. It also covers English, Mandarin, Cantonese and Singlish.

UOB's claim is that MERaLiON has the lowest average word-error rate among open-source models across languages. The footnote lists Mandarin Chinese, Malay, Indonesian, Vietnamese, Thai, Tamil and Burmese as the languages behind that figure. Two things to note.

First, the language list mixes ASEAN and non-ASEAN languages. Mandarin and Tamil aren't among the region's official languages, though both matter in the markets UOB serves. If you're citing the "11 ASEAN languages" figure, check it against the release before you repeat it.

Second, word-error rate measures transcription accuracy. It doesn't tell you whether the system understood a customer's intent, or whether it handled a loan discussion with a small-business owner in Chiang Mai correctly. A low WER on general speech can coexist with poor performance on domain jargon and regional idiom. When you evaluate any speech or language model, ask for WER per language on audio from your own domain, not just a headline benchmark.

That said, the code-switching problem is real. When an RM misses a term because the customer switched languages mid-sentence, the risk disclosure may never get said clearly. That gap is what UOB is trying to close.

Why Post-Call Review May Matter More Than Live Prompts

Most coverage frames this as a real-time translation tool. The common assumption is that putting AI in front of the customer is the valuable part. The release points somewhere else. The post-call review, which creates a record tied to the conversation's evidence, is described as "an additional layer of governance." That's arguably the more important function.

Here's the reasoning. A live prompt competes for an RM's attention during a conversation. An RM under time pressure may treat a prompt as authoritative, even when the model misread a code-switched phrase. Cause and effect: when the prompt is wrong and the RM follows it, a customer gets the wrong framing, and nothing in the moment catches it. A post-call review works differently. It runs after the call, checks the transcript against the disclosures that should have been made, and produces a record that compliance teams can audit and coach from.

This is where I'd push back on the lazy version of the story. Real-time AI assistance sounds more impressive, but it also adds a second voice in a live conversation that the RM has to judge. The review layer is less glamorous and probably more defensible.

The table below breaks down the four functions the release describes. The "main risk" column is my analysis, not UOB's claim.

FunctionWhen it runsWhat it relies onMain risk to watch
Live RM promptsDuring the callReal-time interpretation of questions across languages and accentsRM over-trusts a prompt that misread code-switched speech
Post-call compliance reviewAfter the callEvidence from the conversation itselfReview quality depends on transcript accuracy
Manual translation supportDuring or after the callMultilingual processing across 11 ASEAN-linked languagesLoss of local expressions and cultural nuance
RM coaching insightsPeriodically, from aggregated reviewsPatterns across many reviewed callsCoaching built on the model's own errors

Where Fair Dealing and Accountability Sit

The release is clear on one point: "RMs would remain accountable for the conversations and advice they provide, with AI complementing their professional judgement." That's the right design principle. It's also the easiest one to erode in practice.

The phrase "fair dealing" carries specific weight in Singapore. The Monetary Authority of Singapore (MAS) sets fair dealing expectations for financial institutions, and this MoU builds on A*STAR's collaboration with the MAS-established Future of Finance Institute to test AI against real-world financial-sector needs. So the framing isn't marketing language. Regulators are watching how these tools are used.

Here's what that means for you if you run a compliance or front-line function. The value of a post-call review depends on whether it can point to specific evidence in a conversation. A review that says "the risk disclosure was unclear" without showing where and why is hard to act on. Insist that any system you evaluate ties its findings to the transcript.

The real-world trade-off is speed against certainty. Live prompts are fast but can be wrong in the moment. Post-call reviews are slower but give you something to audit. Most organisations will want both, but they should be deployed with different expectations about who checks what.

What to Ask Before Copying This Approach

Plenty of banks will want to replicate this. Before you do, get clear answers on a few points that the release doesn't address:

  • Where audio and transcripts are processed, and whether they leave the bank's environment.
  • How long recordings and model outputs are retained, and who can access them.
  • How accuracy is measured per language and per product line, not just as a single headline figure.
  • How RMs are trained to treat prompts as input to their judgement rather than instructions.
  • How false positives in compliance review are handled, so that good conversations don't generate unnecessary coaching.

A.STAR's own framing names the priorities: accuracy, explainability and cultural context. UOB's Group CFO, Leong Yung Chee, described the goal as applying AI "with purpose" to help RMs "navigate language and cultural nuances." Those are the right criteria. The release doesn't yet show how they'll be measured, so treat the claims as a starting point for your own evaluation. You can read more about A*STAR's research programmes on its official site.

Frequently Asked Questions

What is MERaLiON multilingual AI?

MERaLiON is a multilingual AI model developed by Singapore's Agency for Science, Technology and Research (A*STAR). Its name stands for Multimodal Empathetic Reasoning and Learning in One Network. UOB says it supports multilingual processing and code-switching across ASEAN languages, alongside English, Mandarin, Cantonese and Singlish.

How will UOB use MERaLiON with its relationship managers?

UOB is exploring a solution that prompts relationship managers on product information and risk disclosures during calls. It also plans an evidence-grounded review after each conversation. The first phase runs through Project AURA, and the project is still in exploration under a memorandum of understanding.

Why does code-switching matter in bank customer calls?

Customers often switch languages or add local expressions within one exchange. If a system or an RM misses a term, a key risk disclosure may not land clearly. That gap can affect fair dealing outcomes, which is why UOB is testing multilingual processing for these conversations.

Can AI replace relationship managers in customer conversations?

No. UOB states that RMs remain accountable for the conversations and advice they give. The AI is positioned to complement their professional judgement, not replace it. Any deployment should keep that accountability clear in both process and training.

Relationship managers don't need to wait for a finished product to act on this. The core lesson is that the post-call record is where accountability lives, so it deserves as much attention as live assistance. Pass this to a compliance lead who's wrestling with multilingual call reviews, and ask them one question: can your review tool point to the exact transcript evidence behind every finding? Then compare their answer with what UOB has described for its MERaLiON multilingual AI pilot.