How U Mobile built agentic telco operations with Amazon Bedrock AgentCore
B · Useful, limited A detailed architecture and rollout account from U Mobile's IT division, co-written with AWS and a partner on AWS's blog; outcome figures are internal measurements with no control group.
Summary based on the full text·Original: about 12 minutes to read
In short
An AWS Machine Learning blog post on AiDaH, the internal agentic helpdesk that U Mobile, a Malaysian 5G operator, built on Amazon Bedrock AgentCore with AWS and partner Axrail. It serves L2 call-centre and branch staff through Microsoft Teams. The most useful detail is the routing: most requests never reach a model. The agent proposes a remediation and a staff member confirms it before anything changes on a live subscriber account.
Key numbers
- 70 / 20 / 10 percentrequests resolved by keyword match with no model, by a lightweight model, and by the full orchestrator
- 84.5 percentfall in monthly provisioning and network tickets, January to September 2026
- 48.7 percentfall in billing and charging tickets over the same period
- more than 150documented operational scenarios covered, with 8 specialist agents
- 4.6 out of 5internal survey rating for ease of use
As reported by the source.
English·Web page·opens aws.amazon.com
What it reports
- Timeline: two pilot agents in late 2025; a Q1 2026 pilot with 4 agents and 20 scenarios; 6 agents by mid-2026; live to 50 L2 call-centre and branch staff in Q3 2026 with 8 agents and more than 150 scenarios across six business divisions.
- Routing: about 70% of requests resolve through deterministic keyword matching with no model call, 20% are classified by a lightweight model, 10% need the full orchestrator.
- Architecture: six layers (Teams and web interface, orchestration with pre-flight access checks, 8 specialist agents talking over Agent2Agent, MCP gateways to backends, knowledge bases, on-premises systems over VPN). Sensitive fields are masked before reaching the model.
- Security: a Well-Architected Security Review of the landing zone before scale-up; a six-tier role model per agent, tool and data source; full traces to CloudWatch.
- Adoption: open "ask anything" did not work; structured prompt templates per scenario did. AiDaH is now written into the L2 support standard operating procedures.
- Results (U Mobile internal measurements, monthly ticket volumes January to September 2026): provisioning and network -84.5%, general inquiries -81.6%, billing and charging -48.7%.
- Built with spec-driven development in Kiro on a 12-week sprint cadence.
How they got it
A customer case written with the vendor. Ticket figures are internal before/after counts; the survey score is from an internal survey. No method, sample or control is described.
Limits
- Vendor co-authored and promotional for AgentCore and Kiro.
- Ticket declines have no control group and may reflect other changes over nine months.
- No error rate, cost or staff-time figure is given.
Why it matters for companies
The 70/20/10 split is the transferable lesson: a cheap deterministic layer in front of the model handles most traffic and keeps cost and risk down. Adoption needed scripted prompts for named tasks, which is a reminder that frontline staff want a button, not a blank box. Writing the tool into the formal procedures, with an audit trail, is what moved it from pilot to operations.
Summary based on the full text. Read the original for the full argument and data.Read the original on aws.amazon.com