How to Agentify Contact Center at Enterprise Scale
Overview
| Experience | In Person |
|---|---|
| Track | Artificial Intelligence & Agents |
| Industry | Energy & Utilities, Communications, Media & Entertainment, Retail & Consumer Goods |
| Technologies | Unity Catalog, Databricks Apps, Lakebase |
| Skill Level | Intermediate |
At Europe’s largest gas distributor, the customer contact center handled approximately 120,000 calls per month using a traditional IVR system. This resulted in 22,000 calls dropped during navigation and 10,000 calls resolved through automation. Thus, the IVR was replaced with a conversational AI agent built on Databricks. The AI Gateway provides the backbone for real-time agent serving, while MLFlow tracks performance. Data is securely accessed from Unity Catalog, while the tools are served via MCP through a Databricks app. The agent interacts with callers in natural language, gathers contextual information, and accurately routes users to the correct resolution path. The solution reduced call drops during navigation by 50% and increased automated resolution by 44%. This session shares how we designed and operationalized a production-grade conversational AI on Databricks, used Unity Catalog as a governed knowledge layer for GenAI and defined measurable KPIs
Session Speakers
Delli, Serena
/Head of Data&AI Architectures
Bludigit, Italgas
Nicola Giorcelli
/Lead Data&AI
Cluster Reply
Full Summary
From rigid IVR to agentic voicebot: how Italgas rebuilt its contact center on the Databricks Platform
Automated phone menus trained customers to expect long, linear trees and little flexibility. At Italgas, a major European gas distributor, that model could not keep up with 120,000 monthly calls. The legacy IVR forced callers down brittle branches, produced a high abandonment rate, and routed too many trivial issues to human operators.
FAQ
MCP servers expose tools and data sources the agents can call. In Italgas's setup they wrap retrieval, internal data APIs, and external systems like the CRM, so the voicebot can carry out tasks such as creating or changing appointments and producing quotations, not just answer questions.
A lightweight intent classifier sits in front of the conversation. Routine flows like authentication and FAQs stay on the fast real-time model, while specialized intents, for example quotations or appointment changes, are routed to dedicated vertical agents built to execute those workflows.
Callers can still tell it is synthetic, according to the speakers, yet the conversational flow is natural enough to reduce frustration and drop-offs. The ability to interrupt, recontextualize, and be rerouted keeps users engaged during beta.
Operators are no longer buried in trivial requests. The system absorbs routine resolutions, freeing the human team to handle complex issues that require judgment and improving the contact center's ability to keep up with volume.
Yes. With agents, data, and governance centralized on the Databricks Platform, expansion to channels like WhatsApp or the public website mainly requires new integrations at the experience layer. The underlying logic, security, and permissions remain consistent.