StormView

The useful AI story this week was not a new model. It was Amdocs putting telco-specific agents into Google's Gemini Enterprise Agent Marketplace.

That matters because telecom is a hard operating environment. Customer care touches billing, service availability, identity checks, order status, network faults and account policy. One wrong action can create a bill dispute, break a service order or send a technician to the wrong address. A generic chatbot is not enough for that kind of work.

Amdocs announced on 13 May 2026 that its Telco Agents for Customer Experience are now available through Google's enterprise agent marketplace. The agents run on Google Cloud's Gemini Enterprise platform, but the more interesting part is the Amdocs layer underneath: aOS and Cognitive Core. Amdocs describes Cognitive Core as the telco-specific reasoning and governance layer for live operations.

That wording sounds dry. In practice, it is the part that makes the system useful.

When we build agentic systems for real businesses, the hard work is rarely the chat interface. The hard work is authority. What can the agent read? What can it change? Which tool calls require a person? What gets logged? How does it recover when Salesforce says one thing, the billing system says another, and the customer is already angry?

Telco exposes those questions quickly. A customer might ask why their service is slow. The agent may need to check account status, service location, outage feeds, modem telemetry, open tickets and plan limits before it can recommend anything. If it books a technician, changes a plan or credits an account, that action needs a clear rule path and an audit trail.

That is what Amdocs and Google are really selling: not a smarter answer box, but governed orchestration across messy systems. Amdocs says the agents handle customer care interactions, service requests, issue resolution and order orchestration across digital and contact centre channels. The company also reports fiscal 2025 revenue of $4.53 billion and says it processes billions of transactions daily, so this is not a lab demo looking for a business case.

The same pattern is starting to show up outside telecom. TechNode Global reported on 14 May that an Informa/Omdia report found 42% of APAC enterprises plan to spend at least $1 million on AI agents over the next 12 months. The report also said 64% support sovereign AI approaches. That tracks with what we see in local deployments: data location, access control and operational logging shape the architecture from day one.

Observability is also catching up. Honeycomb announced agent observability features on 12 May, including timelines that trace LLM calls, tool invocations, handoffs and downstream system impact. WSO2 is moving in a similar direction with Agent Manager, a control plane for identity, governance, access delegation and monitoring.

For mid-market teams, the lesson is simple but demanding. Pick one workflow with a measurable outcome, such as triage time, quote turnaround, ticket resolution or billing accuracy. Connect only the systems the agent needs. Put human approval around money, legal commitments and customer-facing changes. Log every decision and every tool call. Measure the error rate before expanding the scope.

The next serious agent projects will look less like chatbot pilots and more like small operating systems for specific workflows, with permissions, telemetry and escalation built in from the first release.