U Mobile cuts support tickets by 84.5% using agentic AI

Solving the coordination problem

U Mobile, a 5G network provider in Malaysia, faced a common hurdle in telecommunications: fragmented data. When a customer has a problem, like being unable to buy a data add-on, the solution usually requires checking multiple systems, including logs, subscriber eligibility, and internal policies.

Previously, human agents had to manually switch between these different portals to find the root cause. This manual process caused delays and often led to tickets being escalated to backend IT teams, creating a bottleneck for customer service.

To solve this, U Mobile built AiDaH, an AI-Driven Agentic Helpdesk. Unlike basic AI that just drafts text, an agentic system can navigate these various backend systems independently to analyze results and propose specific corrective actions.

From pilot to production

The company began conceptualizing the platform in late 2025, starting with a small pilot. This initial phase used Amazon Bedrock Knowledge Bases, which is a managed capability for Retrieval Augmented Generation (RAG) to ground AI responses in specific data.

By early 2026, the project expanded to include four agents covering 20 operational scenarios for the internal IT helpdesk. The platform continued to harden throughout the year, eventually reaching 50 users in call centers and branches by the third quarter of 2026.

Today, AiDaH covers more than 150 documented operational scenarios across six business divisions. The system is designed to scale by adding new 'gateways' for different domains without needing to rebuild the core architecture.

How agentic AI works in practice

The difference between traditional AI and agentic AI is best seen in how they handle complex tasks. A standard AI might suggest troubleshooting steps for a roaming issue, but the human still has to manually verify the status and eligibility.

With AiDaH, the system independently checks the subscriber's roaming configuration and identifies the exact point of failure. It then assembles a remediation plan, such as which services to reset or what configuration to push, for the human agent to review and confirm.

In one instance, the system identified that a customer had tried to purchase an add-on over 30 times in three days. It traced the failure to a specific 'pending order' error code and provided the fix in a single response, a task that previously required manual investigation across several systems.

The six-layer architecture

AiDaH is built on a layered architecture using Amazon Bedrock AgentCore. Staff interact with the system through Microsoft Teams, which allows them to use the tool within their existing workflow rather than learning a new application.

The system uses a routing brain to decide how to handle requests. About 70 percent of requests are resolved through simple keyword matching, 20 percent use a lightweight model, and only 10 percent require the full orchestrator. This ensures the most cost-effective path is taken for every query.

Specialist agents communicate via a protocol that allows them to delegate subtasks to one another. For example, if a port-in query reveals a billing block, the agent can hand off the task to a business support agent to merge the results into a single narrative for the user.

Security and data privacy

Because the AI can access sensitive subscriber data and billing systems, U Mobile implemented a multi-layered security foundation. They used a six-tier role-based access control model to ensure that agents can only reach the specific systems their human users are authorized to see.

Sensitive information is masked at the gateway before it ever reaches the AI model. Additionally, the system uses a secure VPN tunnel to connect to on-premises systems, ensuring that no data is exposed to the public internet during the process.

Every action taken by an agent is recorded in an audit trail. This includes prompts, tool calls, and latencies, which are written to monitoring tools to ensure the system remains compliant with operational standards.