Two-Thirds of Malaysia’s AI Customer Service Chatbots Fail to Understand How We Talk

I rarely interact with AI customer service chatbots in Malaysia, but when I do, I notice that most of the time, I’m being given the runaround and fixed answers that don’t solve my problem. The repeated ordeal for every new enquiry I have peeves me at times, enough to drive me to barge through the chatbot flow in search of a human.

Turns out I’m not the only one.

“100% of my interactions have been escalated to a human agent. And the worst part is, I have to re-explain everything.”

Those were the words of another Malaysian user describing an eCommerce chatbot on one of the country’s most popular platforms, as shared by Entermind, a native data and AI consultancy here.

The firm’s recent whitepaper, “Confessions of an AI Chatbot,” found that most Malaysian chatbots amount to what it shares is a “glorified search bar”, according to one of its evaluators.

Yet what we, the consumers, actually need are chatbots with genuine conversational capabilities, robust security, and most importantly, the ability to solve the issues we’re having.

So who’s delivering in Malaysia’s AI chatbot space? Entermind sought for insights into customer service chatbots here with its Enterprise Chatbot Quality Index, testing 24 chatbots across E-Commerce and E-Wallet, Travel, Telecom and Financial Services.

Each chatbot was assessed against 26 standardised binary tests, grouped into five weighted categories, namely Comprehension, Access, Experience, Functional Capability, and Safeguards. The assessments were conducted by trained evaluators with sector-specific scripts.

Touch ‘n Go Tops the List in E-Commerce & E-Wallet Chatbots

Entermind’s analysis points to Touch ‘n Go as one of only two chatbots to pass all seven Comprehension tests, handling everything from slang and topic changes to multilingual support and negation.

It helps that the e-wallet has the scale, too. TNG eWallet serves 26+ million verified users, 13.5 million of whom open the app every month, engaging with the platform about twice a day on average. The company recently refreshed its homepage, too, as services beyond payments now contribute more than half of its revenue.

Touch ‘n Go topped the e-commerce and e-wallet category at 82.3%, but Boost was not far behind at 74%, and it earned its score in a different manner.

Boost passed all seven Safeguards tests, including manipulation resistance, profanity handling and error recovery.

The fintech also uses transaction data from its eWallet ecosystem to provide data-driven lending and embedded digital banking solutions for small businesses and individuals who lack conventional credit histories, according to The Star.

Shopee came next at 64.5%. One evaluator shared that Shopee’s chatbot is scoped to seller interactions and product queries rather than full customer service. Its focused design yielded 4/7 on Comprehension, and the chatbot passed Transaction, one of only three in the index to do so.

Notably, Shopee is said to offer the highest e-wallet limit in Malaysia as of January 2026 at RM30,000.

Lazada stood at 60.7%. Its Comprehension was strong at 5/7, including negation handling, which only 23% of chatbots pass. Lazada’s Experience was equally balanced at 4/5, but its Safeguards score was 2/7, with no fallback loop or error handling, according to the whitepaper.

Grab, whose chatbot operates more as a triage layer, is said to route users rather than resolve queries. Some of its approaches, though, still have merits, with one evaluator describing Grab’s refund system as “almost instantaneous.”

Ryt Bank Never Loses the User’s Context in Handoffs

The Financial Services segment had the lowest sector average in the Enterprise Chatbot Quality Index at 38.3%. Ryt Bank was the only entrant displaying genuine conversational capability. The other four chatbots remain in what the whitepaper described as developing implementations or triage layers.

Ryt Bank scored 7/7 on Safeguards, 4/6 on Comprehension, and is reported to process 80,000+ transactions per month through natural language, with a reported hallucination rate of below 1.5%. Ryt Bank recently hit 1.2 million customers in April 2026, seven months into its launch.

One evaluator noted that it “understands criticality and passed me on to a human right away,” while another praised its well-summarised bullet points, complete with links.

Allianz came in next at 38.5%, anchored with a solid safety score as it passed hallucination resistance, full accuracy and manipulation resistance. However, it scored 1/3 on Access and 1/5 on Comprehension.

Bank Islam held a score of 37.5%, passing all three Access tests, but it scored 1/7 on Comprehension and 0/3 on functional capability.

Both UOB and Maybank were reported to operate as triage layers, each passing Access at 3/3. UOB’s single additional pass came in personal data protection. Meanwhile, Maybank is said to offer only preset FAQ menus with no free-text input, rendering the entire Comprehension category structurally inaccessible.

The gap between Maybank and UOB’s sophisticated digital apps and their chatbot capabilities was the most notable disconnect. Separately, one evaluator shared, “I am more willing to trust a banking bot, because they already have my account details.”

Two-Thirds of Chatbots Don’t Understand How We Talk

Key takeaways include the fact that comprehension is the sharpest differentiator. 78% of chatbots that pass at least four of the seven Comprehension tests score above 60% overall.

The inverse holds too. All six triage bots pass Access but fail Comprehension, and Financial Services scored below 25% on the category, with only one in four chatbots in the sector reliably understanding natural language.

The consequences cascade. A chatbot that cannot recognise key details or handle topic changes forces users into rigid menu pathways or repetitive phrasing, and a related failure compounds the problem: localisation.

Only eight of the 24 chatbots, namely Maxis, CelcomDigi, Ryt Bank, AirAsia, Batik Air, Boost, Touch ‘n Go, and Shopee, pass both the language and slang tests.

This means that two-thirds of chatbots serving Malaysians cannot understand how we communicate. Prashant Kumar, CEO of Entermind, also recently shared in an interview with BFM 89.9,

“The thing is, in large corporations, between all the processes, approvals, and silos, it is very easy to forget what truly matters, which is how you create value at the customer interface. Customer value is the true North Star.”

The whitepaper traces a simple pattern across the top seven chatbots. All these businesses have zeroed in on language understanding, conversation handling and safety.

It also sets out the priorities for building better customer service chatbots. These include investing in language understanding, fixing fallback and error handling, and crucially, building cross-session memory, something none of the 24 chatbots evaluated could demonstrate.

AI chatbots in Malaysia seemingly greets its users as strangers every time. While leaders have shown that comprehension and safety are solvable problems, there is reason to believe, and hope, that memory will soon follow.

Featured image edited by Fintech News Malaysia based on an image by tete_escape on Magnific