Enterprise AI Chatbot: Indonesian Telecom

Customer experience · Deployed December 2025 · Major Indonesian telecom operator

A major Indonesian telecom operator needed to improve customer service without continuing to add agents in proportion to demand. The resulting GenAI conversational platform handled approximately 300,000 conversations per month, combining automated resolution with a deliberate route to human help.

Across a reported four-month operating period, first-contact resolution increased from roughly 50% to 90%, weekly NPS moved from −39 to +25–+35, and agent-related costs fell by approximately 44%. These outcomes describe the service operation as a whole, including its human escalation path.

The problem: rising demand and a deteriorating service experience

Prepaid customers contacted the operator about billing, troubleshooting and service activation. They needed help across channels and languages, including outside normal working hours. A customer could begin with an apparently simple question and then require account-specific information or a change in a backend system.

The existing operation was under pressure. NPS had fallen to −39, agent handover exceeded 32%, and weekly unassigned ticket backlogs reached 178 cases. Adding more agents was becoming an expensive response to a problem that also involved routing, context and workflow design.

The objective was therefore broader than answering frequently asked questions. The platform needed to understand what the customer was trying to accomplish, retrieve the relevant information, execute supported actions and recognize when a person should take over.

The approach: resolve the routine, escalate the uncertain

The platform used language models for intent understanding and conversational context, with an NLP pipeline for entity extraction and multi-turn state. A structured knowledge base connected services, policies and workflows, while secure integrations connected the conversation to customer and operational records.

The defining product decision was intelligent escalation over forced containment. Keeping a customer inside a bot conversation is useful only when the issue is progressing toward resolution. Complex or uncertain cases were routed to agents with their conversation context intact.

This is a logical view of the reported platform. It shows the responsibilities and integration boundaries without exposing the customer's deployment details.

Delivery approach and the work involved

Connect customer intents to operational workflows

Billing, troubleshooting and activation required different responses. A policy question could be answered from approved knowledge. An account-specific question required customer context. A supported transaction required an API interaction and a confirmed result.

This separation made the knowledge base and the integration layer complementary. The model could explain a service, while the relevant operational system remained responsible for the customer's actual account state.

Preserve context across the conversation

Intent detection alone was insufficient. The platform needed to retain entities and context across follow-up questions, corrections and clarification. Asking for information already supplied adds friction, especially when the customer has moved from an automated exchange to an agent.

Multilingual handling covered Indonesian, English and regional languages, with automatic detection and transliteration support. Language handling belonged throughout the journey: interpreting the request, finding the right workflow and presenting the next step coherently.

Make human handoff part of the product

Confidence-based routing directed complex cases to agents while preserving the exchange. The handoff was an intended completion path for the automation layer, rather than an exception to hide in the metrics.

The agent needed enough context to continue the work: what the customer wanted, what had already been discussed and where the automated path had stopped. This reduced the need for the customer to begin the interaction again.

Connect channels to a common service layer

The platform's channel coverage included web, mobile apps, WhatsApp, Facebook Messenger, IVR-to-text and portal integrations. The broader channel stack also supported government portals. Channel availability and the reported total conversation volume are separate facts; the operating figures do not provide a channel-by-channel breakdown.

A modular, API-first architecture supported these entry points alongside billing, CRM, provisioning and customer-database integrations. High availability, autoscaling, PII masking and encryption were part of the reported platform design.

Improve through operational review

Human review connected unresolved cases and agent feedback to knowledge and workflow improvement. This is particularly important for a telecom service, where policies, products and the distribution of customer problems can change.

The relevant feedback was not simply whether an answer sounded natural. It was whether the customer's issue was resolved, whether the next owner could act, and whether repeat contacts or unresolved tickets exposed a gap in the workflow.

Results across four months of operations

The NPS range is important: the improvement reaches 74 points at +35 and 64 points at +25. Describing the entire range as “70+ points” would overstate its lower end.

The cost comparison is a reduction of 61M in the reported units, or 43.9%. The public case study does not assign a currency or annualize the amount because the supplied cost figures do not identify those details.

What the metrics mean together

First-contact resolution and autonomous resolution answer different questions. A case can be resolved during the first contact after an agent takes over. The 90% service-level FCR figure should therefore not be read as a 90% bot-only resolution rate.

Similarly, the 55–63% autonomous resolution rate is explicitly measured among answered conversations. It should not be multiplied by the 300,000 monthly total without confirming that the populations match. Handover, autonomous resolution and unresolved backlog also need consistent definitions before they can be reconciled into a single funnel.

The combination is nevertheless commercially meaningful: better reported satisfaction and first-contact resolution, fewer unassigned tickets, lower agent handover and lower agent-related cost. The goal was a more effective customer-service operation, with automation and people contributing to the result.

The delivery scope

The delivery work spanned the conversational experience and the operation around it: telecom intent and context handling, multilingual access, knowledge-to-workflow mapping, enterprise integrations, human escalation and ongoing service review.

The client is described by sector and geography, and the operating figures are retained without exposing customer records or confidential integration details.

The delivery lesson

The strongest result was the combination of scale and service quality over a sustained operating period. A platform handling hundreds of thousands of conversations still needed to recognize the point where a human could resolve the issue more effectively.

The engineering principle is to optimize the completed customer journey: a reliable answer, a confirmed action or a useful handoff. Containment becomes valuable when it supports that outcome.

Related reading: governed enterprise search and the FIELD delivery methodology.