Telecom Generative AI Customer Support
Intent-based chatbots and manual agent lookups replaced with LLM-driven conversations and real-time agent assistance, deployed inside the existing contact centre infrastructure.
The Challenge
A Canadian telecommunications leader serving millions of customers across the country. Its customers expect support that resolves problems on the first contact, in the channel they choose. The systems behind that support were built for a pre-LLM era, and it showed.
The virtual assistant ran on intent-based natural language understanding: every conversational path hand-built, every change a development cycle
Rigid intent trees could not keep up with how customers actually phrase problems, so the digital experience felt scripted and dead-ended often
Maintenance consumed the team. Keeping the intent model current was continuous manual work that never got ahead of demand
Live agents had no real-time guidance during calls or chats. They worked from manual lookups and static knowledge resources while the customer waited
The result compounded: slower resolutions, inconsistent answers across agents, and rising workload on the support floor
Approach
Reengineered the virtual assistant on Vertex AI Agent Builder and Dialogflow Messenger, replacing hard-coded intents with LLM-driven conversational flows
Shifted flow design to natural language, so product owners define how a conversation should go instead of specifying intent trees for engineers to build
Implemented Generative Knowledge Assist, surfacing context-aware answer suggestions to agents in real time during live interactions
Implemented Smart Reply, generating conversation-aligned response options agents can use directly
Implemented Conversation Summarizer, producing concise summaries of long support interactions for handoffs and follow-ups
Deployed all agent-assist capabilities within Google Contact Center AI, integrated with the client's existing contact centre infrastructure rather than replacing it
What Was Delivered
Virtual assistant live in production on LLM-driven flows, replacing the intent-based system
Conversation flows defined in natural language by product owners, removing the per-change development cycle
3x faster information retrieval for agents in live interactions, replacing manual knowledge lookups
Three agent-assist capabilities in production: generative knowledge suggestions, smart replies, and automatic conversation summaries
Case handoffs and follow-ups now carry an auto-generated summary instead of manual notes
Full deployment inside existing contact centre infrastructure via Contact Center AI, with no rip-and-replace
Business Impact
Maintaining the virtual assistant stopped being an engineering backlog. The people who own the customer conversation now shape it directly.
Agents answer with context in front of them instead of searching while the customer waits. Retrieval that took manual lookups now happens 3x faster, in the flow of the conversation.
Handoffs no longer start from zero. The next agent reads a summary, not a transcript.
The client identified improvements in first-contact resolution and customer satisfaction as the new capabilities took hold, alongside reduced cognitive load on the support floor.
The support platform is now AI-native and extensible. New capabilities land on the same foundation instead of requiring another rebuild.
Vertex AI Agent Builder - Dialogflow Messenger - Google Contact Center AI - Generative AI
Frequently asked questions
- Why replace an intent-based chatbot instead of expanding it?
- Intent-based systems fail structurally, not incrementally. Every customer phrasing the model has not seen is a dead end, and every fix is manual work that adds to the maintenance load. LLM-driven flows handle variation in how people actually talk, so the system improves with the model rather than with headcount. Expanding the old system would have bought more of the same maintenance burden.
- What does agent assist actually change during a live call?
- It removes the search step. Instead of putting a customer on hold to check static resources, the agent sees context-aware answer suggestions, ready-to-use replies, and a running summary of the interaction in real time. The agent stays in the conversation. In this deployment, information retrieval during live interactions became 3x faster than the manual lookup process it replaced.
- Does modernizing the contact centre mean replacing its infrastructure?
- No. The generative AI capabilities here were deployed within Google Contact Center AI and integrated with the contact centre platform already in place. The economics of the project stay tied to the workflow being improved, not to a platform migration. That is also what makes the foundation extensible: new capabilities deploy against existing rails.
Related industry
AI for Telecommunications