U Mobile's AI helpdesk: 70% of requests skip the model, tickets fall 84.5%

On 4 November 2026 the AWS Machine Learning Blog described AiDaH, an AI-driven helpdesk built by the Information Systems Division of U Mobile, a Malaysian 5G network operator, with AWS and partner Axrail. According to the post, the system now covers more than 150 documented operational scenarios and U Mobile's internal measurements show ticket volumes in key categories down by as much as 84.5%.

What it does

The post says support staff reach AiDaH through Microsoft Teams. Instead of drafting a reply, the agents check systems themselves: logs, subscriber eligibility, business-policy documents. In one production case it traced a failed add-on purchase to a "Subscriber has pending order" error and found the customer had retried more than 30 times over three days. Eight specialist agents, from customer records to number portability and finance, hand work to each other over the Agent2Agent protocol, and tools reach backends through Model Context Protocol (MCP) gateways.

The cost control worth copying

Every request first passes a security check covering role-based access, identity and guardrails. It then goes to the cheapest tier that can handle it. Per the post, roughly 70% of requests match a phrase such as "port-in status" and run with no model invocation, 20% are classified by a lightweight model, and only 10% need the full orchestrator.

The post adds that a human stays in the loop where it matters: the agent proposes a remediation plan, and the support worker confirms before anything changes on a live subscriber account.

Controls and rollout

• Identity: staff sign in through Microsoft Entra ID, short-lived tokens authorise agent calls, and the post says an agent can reach only systems its invoking user may see, enforced by a six-tier role model.
• Data: sensitive subscriber fields are masked at the gateway before they reach the model.
• Audit: prompts, tool calls, token counts and latencies are logged to Amazon CloudWatch; AiDaH is now written into U Mobile's level-2 support procedures and is subject to operational audit.
• Adoption: the team gave staff structured prompt templates for common cases, and an internal survey rated ease of use 4.6 out of 5.

The reported timeline and results

Design began in late 2025 with two pilot agents. A Q1 2026 pilot ran 4 agents and 20 scenarios, and the system reached 8 agents by Q3. For January to September 2026, U Mobile reports provisioning and network tickets down 84.5%, general inquiries down 81.6% and billing and charging down 48.7%.

What the post leaves out

This is a vendor-published customer story. It gives no starting ticket counts, so a percentage drop cannot be turned into hours saved. It does not say how many fewer tickets come from AiDaH rather than from other changes in the period, what the system costs to run, how often its diagnoses were wrong, or how often staff overrode a proposed fix. The "back to humans" rate is the figure a finance lead would ask for first.

The routing idea transfers without AWS: send routine, recognisable requests down a path that needs no model, and keep the expensive model for the rest.