Copilot for omnichannel call center agents

Context

TeleRed Mexico is a telecom operator (fixed internet, mobile, and pay-TV) with a contact center of ~600 agents across two sites (Mexico City and Guadalajara), handling voice, web chat, WhatsApp and email. Agents had to navigate 6 different systems (CRM, billing, network diagnostics, ticketing, knowledge base, omnichannel history) to resolve a single query.

The challenge

• High average handle time (AHT) from constant screen-switching.
• Low first-contact resolution (FCR) on technical queries (network diagnostics, billing crossed with plans).
• New-agent onboarding of 6-8 weeks before reaching full productivity.
• Inconsistent responses between agents for the same query.
• Customer data (contracts, partial national ID, payment history) subject to LFPDPPP — restricting use of AI tools that send that data to uncontrolled external services.

The solution: a private-model copilot integrated into the agent desktop

TeleRed deployed an AI copilot living inside the agent's omnichannel desktop, with the model hosted on private infrastructure (own VPC) to keep control over customer contract and billing data.

ARCHITECTURE

• Knowledge layer (RAG): the copilot queries the internal knowledge base (policies, procedures, plan catalog) in real time and answers with the cited source, avoiding invented responses.
• Unified context: when the interaction starts, the copilot already brings the customer's history (recent tickets, plan, payment status, recent network diagnostics) without the agent having to search manually.
• Live transcription and summary: on voice calls, it transcribes and suggests the next best action while the customer is still on the line; at close, it automatically generates the summary and case categorization.
• Real omnichannel: the same copilot layer assists the agent regardless of channel (voice, WhatsApp, chat, email), maintaining context continuity if the customer switches channels mid-interaction.
• Private model, not exposed to third parties: contract and billing data never leaves TeleRed's infrastructure; the copilot runs on a proprietary model tuned with the company's documentation and conversation history.
• Human always in charge: the copilot suggests — it doesn't decide or send anything to the customer without agent confirmation.

Results at 5 months

Metric	Before	After
Average handle time (AHT)	Base 100%	-32%
First-contact resolution (FCR)	Base 100%	+28 pts
Onboarding to full productivity	6-8 weeks	3 weeks
Response consistency (quality audit)	Base 100%	+40%
Different apps open per interaction	6	1 (unified)

Supervisors reported new-hire agents reaching senior-agent performance in half the usual time, with the copilot suggesting procedure and response at every step.

Key success factors

• Copilot integrated into the agent's real workflow, not a separate tool to check independently.
• RAG with internal source citation — reduces hallucinations and gives the agent confidence to use the suggestion.
• Private model over regulated data — avoids legal friction and protects contractual and billing information.
• The agent keeps control — the copilot assists but doesn't automate decisions toward the customer without human review.

Lesson to replicate

The biggest impact didn't come from "adding a chatbot" — it came from eliminating constant switching between systems and giving the agent all the context and suggested response in one place — with the model running on proprietary infrastructure to avoid exposing regulated customer data.