it_support · logistics · workflow
GrabGPT: How Grab built an internal ChatGPT-like tool after a failed support chatbot experiment
Grab's ML Platform team support channels were overwhelmed with repetitive user inquiries, with on-call engineers spending more time answering questions than building innovative solutions.
How it works
Common implementation structure
How this type of workflow is generally built, generalized across documented cases — not tied to any one vendor's stack. Click any stage to read what happens there. Specific products that implement these stages appear in “Tools commonly seen” below.
Stage 1 · Support channels overwhelmed
Slack channels were flooded with questions and on-call engineers were spending more time answering repetitive queries than building innovative solutions.
Tools used
chatbot-uiGPT-3.5-turbocatwalkGrabGPT
Outcome
GrabGPT became one of the most widely used internal tools at Grab, with 300 users registering on day one, over 3,000 users and 600 daily active users by month three, and eventually almost all Grabbers using it.
What failed first
An initial documentation-fed chatbot built on chatbot-ui and GPT-3.5-turbo failed because the model's token context window could not accommodate the full platform documentation, requiring heavy summarization that only covered a handful of FAQs; embedding search also proved ineffective.
Results
Time saved300
Volume600
Running sinceMarch 2023
Grounding & classification
Source type: technical build writeup
30 fields verified against source quotes.
chatbotconversational aiknowledge searchragknowledge basebuilder submittedfailure mode describedmetric backednamed customerproduction runtime claimedtools describedworkflow describedfinancial serviceslogisticsemployee productivitythroughput increasetechnical build writeupback office opsit supportrag answering