Cutting support response times 14.4x for Angkas with multi-agent AI on AWS
Angkas handles ~250,000 bookings a day. ChatGenie built a multi-agent AI support platform on Amazon Bedrock and Amazon EKS that cut support headcount 77% and made responses 14.4x faster.
Lower support OPEX, from 39 agents to 9 at the same volume
Faster responses, from ~360 seconds to ~25
Response accuracy, up from ~96%
About the client
Angkas operates one of the largest motorcycle ride-hailing networks in the Philippines, with around 250,000 average daily bookings and more than 50,000 active riders.
The challenge
Support volume is high and highly contextual: passengers and riders share the same social messaging channels, raising concerns that range from booking and payment issues to rider onboarding and account verification, in English, Filipino, code-switched Tagalog-English, and regional dialects.
Support depended entirely on 39 human agents across three shifts, so costs grew in line with user growth and response times varied widely. Sensitive or high-priority cases were not consistently routed to the right person quickly, and support leaders lacked real-time visibility into sentiment and service quality.
The solution
Swarm engaged ChatGenie as the specialist Hive team to design and deploy a multi-agent AI platform on AWS.
A Guard Agent filters off-topic and harmful messages before any model call, acting as both a safety and cost gate. An Orchestrator Agent generates responses grounded in the customer’s knowledge base through a retrieval pipeline powered by Amazon Bedrock (Cohere Embed v4 Multilingual), which handles mixed-language and dialect inputs without a translation layer.
Supporting agents classify each user as a passenger or rider in real time so the right workflow applies, and track sentiment across conversation turns to escalate distressed users to human agents automatically, creating a ticket in the customer’s system and syncing its resolution back.
All workloads run on Amazon EKS with Amazon DocumentDB, Amazon RDS, Amazon ElastiCache, and AWS Secrets Manager inside a secured VPC.
Impact
Support headcount fell from 39 to 9 agents while handling the same volume, a 77% reduction in support OPEX.
Average response time dropped from around 360 seconds to about 25 seconds, a 14.4x improvement that holds steady during peak demand because the platform scales automatically on Amazon EKS. Response accuracy rose from roughly 96% to 99%.
AI agents now handle routine inquiries and triage end to end, human agents focus on complex and sensitive cases, and support leaders have real-time visibility into sentiment and service quality.
