01
Overview
The Problem
A third-party IVR vendor unexpectedly exited late in the year, removing a critical support deflection layer for a servicing client. With no runway and limited engineering capacity, we needed to quickly replace a system that handled high-volume technical and access-related issues.
Why It Mattered
Without a replacement, support volume would spike at the same time the platform was preparing for a major client launch. Any increase in manual support risked overwhelming teams, delaying launch work, and degrading the borrower experience.
Who This Affected
Borrowers encountering login and technical issues, support teams responsible for managing increased inbound volume, and internal product teams balancing incident response alongside critical delivery commitments.
Success Criteria
- Launch a functional chatbot and knowledge base within days
- Use AI to accelerate research and content creation without sacrificing clarity
- Create a repeatable system that could scale across clients
- Reduce reliance on manual support for common technical issues
02
Context & Constraints
Constraints
The IVR vendor exit left no buffer time. The chatbot and knowledge base needed to be designed, built, and launched within days, with limited engineering capacity and an active client launch already underway.
Known Risks
If the replacement failed to cover common technical issues, support volume would spike immediately. Poorly structured or unclear content risked frustrating users and undermining trust in automation. There was also risk in relying on AI-generated content without a process to validate accuracy quickly.
Dependencies
The solution depended on Zendesk's chatbot and knowledge base capabilities, fast turnaround from client partners for content approval, and close coordination with support teams to identify high-volume issues worth addressing first.
Tradeoffs
We prioritized speed, coverage, and repeatability over completeness. Initial content focused on the highest-impact technical and access issues rather than exhaustive documentation. AI-assisted generation accelerated output, with human review ensuring accuracy and tone before launch.
03
Approach
Start from the borrower's questions
With no runway for iteration, we framed the work around the questions a borrower would ask when something broke. Login failures and technical issues were treated as moments of stress, and content was designed to answer the most critical questions clearly and quickly.
Treat the knowledge base as the system
The chatbot was only as effective as the content behind it. We designed a clear information architecture and consistent article structure optimized for both human readers and automated retrieval. Each article focused on a single question, surfaced resolution steps early, and avoided ambiguity.
Use AI to accelerate thinking, not replace it
To move quickly, we designed a two-part AI-assisted workflow. First, a research "kernel" was generated for each borrower question, using available context to assemble the core information needed to answer it. That kernel was then passed into a second AI-driven article generator that followed a strict, predefined template.
The template itself was designed to ensure clarity, consistency, and suitability for chatbot use. All outputs were reviewed and refined before publishing.
Build for repeatability from day one
The goal was never a one-off launch. The same workflow for defining questions, generating kernels, producing structured articles, and expanding coverage is still in use today. What began as an urgent response became a scalable support layer and a reliable input into ongoing product research.
04
Outcomes
The chatbot and supporting knowledge base launched within days and quickly became the primary deflection layer for technical and access-related support issues.
Since launch, over 23,000 borrowers have interacted with the chatbot, with an engagement rate of 71%. While transfer rates remain higher and automated resolution rates lower than the long-term target, this was expected given the initial scope. At launch, the bot covered roughly 25% of total feature and issue scenarios, prioritizing the highest-impact technical and login-related questions.
Beyond deflection, the chatbot became a valuable research signal. The questions borrowers asked helped validate assumptions about where confusion and friction existed, and directly informed a prioritized content expansion schedule. Patterns in usage also reinforced that borrowers were seeking more than one-off transactional answers, helping validate broader, AI-powered support initiatives focused on deeper guidance and self-service.
The AI-assisted workflow enabled rapid iteration as coverage expanded. That same system is still in use today, supporting ongoing article creation, faster updates, and continuous improvement as new insights emerge.
05
Reflection
This project reinforced that AI is most effective when designed as part of a broader system, not treated as a standalone feature. The chatbot worked because it supported a larger goal: reducing support cost and friction in an industry where call centers are major cost centers and margins are thin.
It also shifted how we think about support. Rather than viewing it as a reactive function, we began treating it as strategic infrastructure. The chatbot became a surface for learning. The questions borrowers asked revealed patterns in confusion and friction, helping prioritize content expansion, inform product decisions, and support more grounded conversations with servicing clients.
The solution was incomplete at launch. What mattered though was building something durable under pressure that could expand and improve over time. Designing a structured, AI-assisted workflow made speed possible without sacrificing clarity, and reinforced that thoughtful systems design is what enables progress at pace.