.png)
Two-Thirds of Malaysia's Chatbots Don't Understand Their Own Customers
By Shawn Tan, Interactive Rewards Asia
Published 20th August 2026
"100% of my interactions have been escalated to a human agent. And the worst part is, I have to re-explain everything." That's a Malaysian shopper describing one of the country's most popular e-commerce chatbots, quoted in Entermind's June 2026 whitepaper on AI customer service. It isn't an edge case. It's the median experience — and it's happening in the exact channel banks, telcos, and retailers across the region bet on to cut cost and lift satisfaction at the same time.
Why This Matters Now
Chat is no longer a side channel in APAC — it's the default one. WhatsApp penetration sits above 80% of adults in Malaysia, Singapore, and Hong Kong, and Indonesia is now the platform's third-largest market globally. Every major bank, telco, and e-wallet has responded by pushing a chatbot into that channel. Entermind's Enterprise Chatbot Quality Index put 24 of them through 26 standardised tests across comprehension, access, experience, functional capability, and safeguards. The finding that should worry every CX leader in the region: two-thirds of Malaysia's customer service chatbots fail to understand how Malaysians actually talk. Financial services — the sector spending the most on this — scored the worst of any vertical tested, averaging just 38.3%.
Where Most Teams Get It Wrong
Most teams are confusing being reachable with being useful. The two failure modes below show up in the index over and over.
Access is mistaken for capability
Maybank and UOB's chatbots pass every access test — always on, easy to find — but function purely as triage layers. Maybank's bot offers only preset FAQ menus with no free-text input at all, which makes it structurally impossible to even score on comprehension. It's not a chatbot; it's a phone tree wearing a chat bubble.
Deflection is treated as the finish line
Every one of the 24 chatbots tested, including the strongest performers, fails to carry memory across sessions. Each conversation starts from zero, so a customer mid-dispute today has to re-explain everything tomorrow — precisely what triggered the complaint quoted above.
"It's not a chatbot; it's a phone tree wearing a chat bubble."
What To Do Instead
Three moves separate the chatbots that work from the ones that don't.
1. Fund comprehension and localisation before adding channels
The eight chatbots that passed both language and slang tests — Maxis, CelcomDigi, Ryt Bank, AirAsia, Batik Air, Boost, Touch 'n Go, and Shopee — share one thing: investment in that layer specifically. Touch 'n Go topped the e-wallet category at 82.3%, the only player to pass all seven comprehension tests, at genuine scale: 26 million verified users, 13.5 million opening the app monthly. That investment is paying for itself directly — non-payment services now generate more than half of TNG's revenue, because comprehension earned enough trust for customers to do more inside the app.
2. Design for safe failure, not just full automation
Boost passed every safeguards test in the index — manipulation resistance, profanity handling, graceful error recovery — even with a lower comprehension score than Touch 'n Go. A bot that fails predictably and hands off cleanly beats one that fails silently and loops the customer in circles.
3. Build cross-session memory deliberately
Nobody in the index has cross-session memory yet, which means the first bank or telco to fix it turns a market-wide weakness into real differentiation — and it's an architecture problem, not a frontier-AI one. Ryt Bank, Malaysia's newest digital bank, is the clearest proof this works inside financial services, the sector everyone else is failing: four of six comprehension tests passed, all seven safeguards tests, natural-language handling of 80,000+ transactions a month, and a hallucination rate under 1.5%. It reached 1.2 million customers within seven months of launch.
The Outcome
The commercial logic is unusually clean
Comprehension is the single biggest predictor of overall chatbot performance in the index — 78% of chatbots that pass at least four of the seven comprehension tests also score above 60% overall — making it the highest-leverage investment on the table. Get it right and you land near Ryt Bank: sub-2% hallucination, six-figure monthly transaction volume through the bot itself, a million-plus customers trusting the channel enough to actually use it. Get it wrong and you get the opening quote on repeat — every escalation a re-explained complaint, every interaction quietly building the case against automation instead of for it.
This is an article published by Interactive Rewards.
Comments