From a flooded support inbox to an operation that runs itself.

Context

Onoff Telecom is a European telecom operator running a high-volume B2C support channel: SIM and eSIM activation, number management, billing questions, app issues. Telecom support has a specific shape: a large share of inbound is repetitive level-1, the same dozen questions asked thousands of times, while a smaller tail is genuinely complex and needs a human.

The brief was not to add a chatbot. It was to make the whole support operation measurable and partly self-running: agents stop drowning in repetitive tickets, managers stop assembling reports by hand. The starting point:

• High inbound volume with a heavy repetitive level-1 share
• First-response times stretched by queue depth, simple questions slowing down the hard ones
• Reporting via a manual export pulled into a spreadsheet once a week, stale by the time anyone read it
• No single real-time view of what support was actually doing

Challenge

Three problems, one system that had to solve all three at once. A point solution would have stranded the value: an agent that cannot see its own data is a toy, a dashboard nobody automates is a gadget. The win required the conversation, the automation, the data warehouse and the report to be one loop.

Solution: one closed loop

The system is a single closed loop across five layers. The agent handles support and logs everything, ingestion and BigQuery turn that into reliable data, the dashboard makes it visible live, and the weekly report closes the loop in the stakeholder's inbox. No single tool does that: the integration is the product.

1. 1

   Request triage agent

   A conversational agent sits on the front line of the repetitive level-1 flow, grounded in Onoff's knowledge base, not in generic model memory. It optimizes for genuine resolution, not vanity containment: a request counts only when the customer never needed a human.

2. 2

   Response automation

   High-frequency questions (activation, eSIM, number management, billing basics) are answered automatically. Anything ambiguous, sensitive or out of scope is handed to a human agent with full context attached, and every interaction is written back into the pipeline.

3. 3

   BigQuery ingestion

   Support events are captured at the source, then cleaned, deduplicated and modeled into a single source of truth: deflection rate, first-response time, volume by category, escalation rate. The unglamorous layer that makes everything downstream trustworthy.

4. 4

   Real-time dashboard

   A live view on top of BigQuery for the support manager: volumes, response and resolution times, escalations, trends. No more waiting for the weekly file.

5. 5

   Automated weekly report

   A scheduled job reads the modeled data, computes the week's deltas, writes a short narrative summary and sends it by email or Slack. The manual export disappears entirely: the report writes and sends itself.

Results

Every figure below is a market-calibrated range derived from industry benchmarks for AI-assisted telecom support, never a contractual Onoff metric. Each range is replaced by the real number the moment actual volumes are shared.

Stack