Layer 04 · AI Agents
20,000 hours recaptured. An AI agent that answers every question, instantly.
Enterprise SaaS · Internal Support Automation
The Client
1,700 employees. Decades of answers. No way to find them.
A publicly traded enterprise SaaS company had grown past 1,700 employees, and its institutional knowledge had grown with it: 33 Slack support channels, a deep Zendesk ticket history, and years of Confluence documentation. The knowledge existed. Finding it was the problem — and the people who could find it fastest were the same experts everyone kept asking.
The Challenge
The bottleneck was the people who knew the answers
The company was fielding 1,200+ support questions a month across those 33 channels. SMEs were the bottleneck: the same people being pulled out of high-value work to answer questions that already had answers, buried in Zendesk tickets, Confluence docs, and old Slack threads. In 2024, between 150 and 240 questions went unanswered every single month.
Our Approach
Start with a knowledge audit, not a technology decision
Before writing a line of code, we ran a knowledge source audit to map where resolution knowledge actually accumulated. The finding: the most valuable institutional knowledge wasn't in Confluence. It was in Zendesk, in the closed-ticket history of how real problems had actually been solved. That breakdown drove the entire architecture: Zendesk (47%), Confluence (34%), Slack history (19%).
The Architecture
Three sources in. One grounded answer out.
Zendesk, Confluence, and Slack feed a weekly-refreshed vector knowledge base; Ask Genie answers in Slack with citations, and monitoring covers every ingestion run.
Design Principles
Grounded answers, with citations
Every response is grounded in retrieved source material with citations. When the knowledge base doesn't have a confident answer, it says so.
Meet users where they work
No new tool to learn — Ask Genie lives in Slack, where the questions already happen.
Build for freshness from day one
A static knowledge base goes stale within weeks. Incremental ingestion was part of the architecture at launch, not Phase 2.
Monitor the pipeline, not just the agent
Automated monitoring recipes alert stakeholders to ingestion failures and knowledge-base capacity limits before users ever notice a stale answer.
The Results
Every question now gets a response
Live since mid-January 2025. A one-time full knowledge load at launch, then weekly delta ingestion ever since — the knowledge base stays current without anyone curating it by hand.
The 50% deflection target fell in month two. In February 2025 alone, Ask Genie deflected more than 2,000 questions that would otherwise have landed on subject-matter experts.
No more silence. In 2024, between 150 and 240 questions a month went completely unanswered. Since launch, every question has received a grounded, source-cited response — and when the knowledge base isn't confident, the agent says so rather than guessing.
20,000 hours, measured by the client. Twelve months in, the client's own estimate put the time recaptured at 20,000 hours — the number this page leads with. It is the same audit-first, grounded-agent standard we bring to every AI agent engagement.
"just wanted to say i love the genie!!"
"Really enjoying trying this out! Congrats team! Huge win."
End users · via Slack, first weeks after launch
What's Next
From one agent to a fleet
The same ingestion architecture is now being configured for additional team-specific knowledge agents, and the platform is being scaled to keep pace with a growing knowledge base. One agent proved the pattern; the pattern is what scales. See how we package it across our solutions.
Technology
Drowning in the same questions every week?
Bring us the Slack channel nobody can keep up with. We'll show you what a grounded, source-cited knowledge agent looks like in a 30-minute working session.