Telco: AI agents for service and churn.

Subscriber service is where telco and broadband providers win or lose customers. In Southeast Asia, a large share of that service now happens on messaging channels, and much of it is repetitive: when will my technician arrive, why is my connection slow, what does my plan include, why is this bill higher than last month.

Each of those conversations is also a churn signal. A subscriber who waits days for an installation, contacts support twice about the same fault and then sees an unexpected charge is a customer who is already comparing offers. Most providers hold that history in separate systems and only see the risk when the cancellation request arrives.

What does the subscriber journey look like?

We map six stages. Each produces service demand, and each leaves a signal of how likely the subscriber is to stay.

Which conversations should an agent handle first?

We score candidate intents on volume, simplicity and risk. In subscriber service, the first wave is usually informational and data-backed:

• Installation and activation status, read from the order or field service system.
• Plan details, coverage and add-ons, answered from current product content.
• Known outages and planned maintenance in a subscriber's area.
• Simple account questions such as due dates and payment methods.

HubSpot describes Customer Agent as able to “resolve inquiries automatically across channels”. HubSpot reports that teams using it close 77% more customer tickets per month on average, with 39% faster ticket resolution. These are vendor averages; a provider's result will depend on its content and data.

The deployment controls matter for a high-volume operation. HubSpot's documentation lists WhatsApp, Facebook, live chat, forms, email and custom channels, with calling in beta. You can set working hours and the percentage of conversations the agent handles, and HubSpot says credits “are only consumed when a conversation is resolved” (HubSpot Knowledge Base). That makes a staged rollout, one channel and one intent at a time, straightforward.

What should stay with people?

Some subscriber conversations should go to a person by design:

• Cancellations and disconnection requests, which are retention conversations.
• Billing disputes and refunds.
• Complaints, and any subscriber who has contacted support about the same issue before.
• Business and enterprise accounts with service level commitments.

HubSpot's handoff settings support this. By default the agent hands over when it cannot answer, when the visitor asks for a human or when it is paused, and you can add custom triggers; HubSpot's own example uses “Cancellation, Refund, or Can't log in”. You can choose a live handoff to the next available person or an asynchronous one for later follow-up (HubSpot Knowledge Base).

On WhatsApp, Meta's Business Messaging Policy requires businesses running automated experiences to provide “prompt, clear, and direct escalation paths” to a human (WhatsApp Business Messaging Policy). Design the handoff before launch.

How do service conversations become churn signals?

The agent is only half the value. The other half is what the conversations reveal. When every installation delay, repeat fault and billing query is on the subscriber's record, churn risk becomes visible months before the contract ends.

HubSpot's customer success workspace lets teams “build unique customer health scores” and prioritise their activities (HubSpot). We build health scores from the signals that matter in each business: repeat contacts, unresolved tickets, missed installation windows, payment issues and how close the contract is to its end. The retention team then works the list before the subscriber calls to leave.

Where do the billing and network systems fit?

The billing system and the operational support systems remain the record for plans, usage, charges and network status. The CRM and the agent should read from them, not copy them by hand. The design work is deciding which fields the agent may read, which it may update, and how quickly outage and order status reach the subscriber's record. See our integration approach.

What should a telco measure?

• Resolution rate on the intents the agent owns, and handoff rate with reasons.
• Repeat contact rate for the same issue.
• Installations completed on time.
• Health score distribution across the base, and how it moves.
• Churn and save rate among subscribers the retention team contacted first.

How do you roll the agent out without risking subscribers?

Telco volumes are high enough that a poorly designed agent can damage satisfaction across a whole market quickly. We use a staged rollout.

1. Shadow. Test the agent against a large sample of real historical questions. Review every answer with the best people on the service floor.
2. Partial. Assign a small share of live conversations on one channel during staffed hours, with live handoff switched on.
3. Expand. Raise the share as resolution holds and feedback stays healthy. Add intents one at a time, starting with those backed by live data.
4. Next channel. Only then move to the next channel, language or market.

Each step has an owner and an exit rule. During major outages, consider lowering the agent's share or limiting it to outage updates, so subscribers with urgent problems reach people quickly. Language matters too: test in every language your subscribers use, not only English.

Where to start

Pick one channel and three intents with stable, data-backed answers. Connect order and outage status, write the handoff rules, and run the agent on a small share of conversations. In parallel, bring service history onto the subscriber record and build a first health score. Our Customer Agent page sets out how we deploy and govern it.