An agent platform can do useful work while sending most requests around its full orchestrator. U Mobile’s helpdesk offers a concrete example of that choice.

In a November 4 AWS case study, the Malaysian operator describes AiDaH, a helpdesk built with AWS and partner Axrail on Amazon Bedrock AgentCore. The account says roughly 70% of requests resolve through deterministic keyword matching, 20% use a lightweight classifier and 10% require full orchestration. Those proportions describe this deployment’s routing pattern, rather than a target every enterprise should copy.

The distinction matters because familiar requests and unfamiliar investigations create different work. A recognized status query can follow a known route. An ambiguous request may need classification. A problem spanning several backend systems may warrant a model coordinating the investigation.

AiDaH reaches more than 150 documented operational scenarios across six divisions, according to the case study. Its proposed remediation still goes to a support employee for confirmation before live subscriber-account changes execute. That places the approval at the point where a diagnosis becomes an action affecting a customer.

The reported business result needs its denominator. U Mobile’s internal measurements show monthly provisioning and network tickets falling 84.5% between January and September 2026. This is a category-specific observation reported by the operator and its technology partners. It does not establish that AI alone caused the decline or that the same reduction applies across the helpdesk.

For a team evaluating this pattern, start with its own request history. Separate repeatable lookups from investigations that need judgment, and measure how often each route resolves the problem. Count escalations, mistaken routing and staff review time alongside model charges. The routing decision is worthwhile when it reduces the work required to reach a correct, authorized outcome.