Context
PLUS Malaysia, the country’s largest highway operator, handles a large volume of repetitive customer contact: toll rates between two exits, why a barrier did not lift, how to dispute a charge, whether a stretch is closed. Much of it arrives outside the hours when a contact centre is cheap to staff, and much of it is answerable from information the operator already publishes.
The brief was not “add a chatbot”. It was to take the predictable half of the contact volume off the queue without making the unpredictable half worse.
What I did
- Started from the contact logs, not from the technology. We clustered real enquiries into intents and found that a few hundred of them covered the overwhelming majority of traffic.
- Designed the agent bilingually from day one. Malaysian customers code-switch mid-sentence, so treating Bahasa Malaysia as a translation layer over an English bot would have failed. Both languages were first-class in intent design and in testing.
- Built the escalation path first. Every conversation the agent cannot close hands to a human with the customer’s context attached, so the handover does not restart the conversation.
- Set up the content operation that keeps it accurate: named owners for each answer domain, a review cadence, and a feedback loop from unresolved conversations back into intent design.
- Integrated with the systems of record, so the agent could answer account-specific questions instead of only reciting published FAQ text.
Outcome
The toll, traffic and account questions that make up the bulk of the volume get answered at any hour, in either language, without a queue. Human agents spend their time on the exceptions, which is what they are good at. The win was less “cost saved” than “the contact centre stopped drowning at 11pm”.
What I’d do differently
I would have invested in the content operation earlier. We built a capable agent and then discovered that the limiting factor was whether someone owned each answer and kept it current. An agent is a publishing product wearing a machine-learning costume: if nobody owns the content, accuracy decays within a quarter and trust never recovers.
I would also have resisted the pull toward breadth. Every stakeholder wants their department’s questions added. Depth on the top intents beats shallow coverage of everything, and the second is much harder to walk back.