Skip to main content
Lucy AI
CustomersSecurity
Book a Demo

Loading resources...

This may take a few moments

Lucy Logo

Creating digital experiences that inspire and innovate. We build tomorrow's solutions with today's cutting-edge technology.

Quick Links
  • Home
  • Products
  • Solutions
  • Customers
  • Security
  • Resources
Contact

Email Us

support@lucyhub.ai

Our Locations

Arlington, VA — USA
Addis Ababa, Ethiopia

© 2026 Lucy AI. All rights reserved.

INSA Security Assessed
Terms of ServicePrivacy Policy
Back to blog
Customer Stories2 min read

From pilot to production in two weeks: how one support team did it

A walkthrough of a real deployment, from the first exported call log to the day the agent started taking live traffic.

LA

The Lucy Team

June 26, 2026

Share
A support team reviewing call transcripts

On this page

  • Days 1 to 3: read the calls you already have
  • Days 4 to 7: write the knowledge base
  • Days 8 to 10: connect the number
  • Days 11 to 14: ramp on live traffic
  • What we would do differently

Key takeaways

  • The slow part is never the integration. It is agreeing on what the agent is allowed to say.
  • Start with read-only requests, add write actions after the first week of transcripts.
  • Route a percentage of live traffic rather than running a separate test number.

This is the sequence we ran with a support team handling a few thousand calls a week across four languages. Nothing here is unusual, which is the point. The parts that take time are predictable, and most of them are decisions rather than engineering.

Days 1 to 3: read the calls you already have

Before touching configuration, we exported three months of call reasons and sorted them by volume. Five intents covered just over 70 percent of inbound. That list became the scope of the pilot, and everything outside it was explicitly out of scope until the agent handled those five well.

Resist the urge to cover everything

A pilot that handles five intents correctly is a production system. One that half-handles thirty is a demo.

Days 4 to 7: write the knowledge base

The five intents needed about forty entries between them. The team wrote them in Amharic and English, and the longest arguments were about what the agent should say when a policy has exceptions. That conversation is worth having early, because it is the same conversation the agent will otherwise have with a customer.

  • Forty entries, authored by the support leads, reviewed by compliance.
  • Read-only actions connected first: balance, status, hours, branch lookup.
  • Escalation rules written as plain sentences, then encoded.

Days 8 to 10: connect the number

The telephony work took an afternoon. The agent was placed in front of the existing queue rather than beside it, so callers reached it on the number they already knew. Anything it could not resolve fell through to the same agents as before.

Live call monitoring on a phone
Supervisors watched live transcripts from day one.

Days 11 to 14: ramp on live traffic

Traffic went to the agent in steps: 10 percent, then 40, then everything. Running on real calls from the start is what made the ramp fast, because the failures that showed up were real failures rather than the ones a test script imagines.

5
Intents at launch
~40
Knowledge base entries
14
Days to full traffic

What we would do differently

Two things. Start the compliance review on day one instead of day four, because it was the only thing that ever sat idle waiting on someone. And add write actions a few days earlier: the team held them back out of caution, and the transcripts from week one showed the agent was ready before they were.

“Setup took days, not months. We connected our number and started routing the same week.”
Head of IT · Technology

Run the same two weeks

Bring a call log and we will tell you which intents to start with.

Book a demo
TagsDeploymentCase studyOperations
Share
LA

Written by

The Lucy Team

Lucy AI

Keep reading

More from the blog

A Lucy AI voice agent answering an inbound support call
Industry Insight

How AI voice agents cut call abandonment in a multilingual queue

Aug 18, 2026 · 4 min read

A dashboard showing voice agent performance metrics
Product

Six metrics worth tracking on a voice agent, and three that mislead

May 28, 2026 · 2 min read

A caller navigating a bank phone menu
Industry Insight

Your IVR menu is losing callers before they reach an agent

Aug 4, 2026 · 2 min read

See Lucy on your own calls

Tell us where your calls pile up and we will show you what Lucy does with them, in Amharic, Afaan Oromo, Tigrinya or English.