U Mobile’s AI Helpdesk Cuts Tickets by Up to 84.5%—With Approval at the Point of Change

A customer cannot buy a mobile-data add-on. The answer may be split across an error log, a billing rule, subscriber eligibility and a provisioning system. U Mobile’s new case study describes an AI helpdesk built to investigate that chain before a support employee opens each system manually. Its reported ticket reductions are striking; its human approval boundary is the part other operators can assess for themselves.

A support ticket becomes an investigation

In an AWS case study published November 4, U Mobile describes AiDaH as an operations platform built on Amazon Bedrock AgentCore. It went live to 50 level-two call-center and branch users in Q3 2026 and now covers more than 150 documented scenarios across six divisions, according to the case report. For an add-on purchase failure, the agent can inspect an error, check customer eligibility and consult operational knowledge, then present a proposed remediation to the support worker.

The example is more revealing than the headline percentage. In one case reported by AWS, AiDaH traced a blocked purchase to a pending order and found that the subscriber had retried more than 30 times across channels over three days. That is a concrete demonstration of cross-system diagnosis in a customer-specific scenario, but it is still one case described by the company and its cloud provider—not a controlled comparison with a baseline helpdesk.

The metric needs its denominator

AWS reports U Mobile’s internal measurements show provisioning and network tickets down 84.5%, general inquiries down 81.6%, and billing and charging tickets down 48.7% between January and September 2026. The different reductions matter: “up to 84.5%” does not describe every ticket or all support work. The post does not publish the underlying raw counts, a control group, category definitions over time, or an independent audit of the attribution. The numbers should therefore be treated as U Mobile’s reported operational outcome, not proof that AgentCore alone caused the change or that another carrier can reproduce it.

A useful evaluation would publish the ticket volume and resolution rate for each category, changes in staffing and network conditions, repeat-contact rates, customer satisfaction, and the share of cases the agent could not resolve. Operators should also measure incorrect remediations and the time taken to reverse them. The NIST AI Risk Management Framework Playbook recommends documenting oversight, overrides, reported errors and escalation outcomes; those measures make an automation claim more useful than a single maximum percentage.

Keep authority close to the action

The case study says AiDaH investigates, assembles a remediation plan and lets a support employee review and confirm actions that alter a live subscriber account. It also describes identity checks, role-based access, masking sensitive fields, and audit traces for tool calls. The approval point is crucial because a diagnosis and an account change are different actions. A likely fix can be helpful; permission to execute it must still come from the right person and policy.

AWS’s AgentCore policy documentation describes controls that can check identity, tool inputs and session history before a call is allowed. These are platform capabilities, not evidence that every deployment configures them correctly. A telco adopting a similar design should verify that each tool enforces least privilege, sensitive subscriber data is minimized, approvals are bound to the exact proposed operation, and the resulting change is auditable.

U Mobile’s report is worth reading as a workflow case study, not as an industry benchmark. The most transferable result is not “AI reduces tickets by 84.5%.” It is the pattern: gather context across fragmented systems, avoid model calls when deterministic rules are enough, make a constrained proposal, and preserve human confirmation when the action changes a customer’s service. See our guides to agent identity and permissions and fresh operational state for the controls around that pattern.