Chat Interface

CloudBolt asked: could a conversational AI interface — similar to how modern search tools work — enhance the FinOps platform? The hypothesis was that natural language queries might lower the barrier for users who found the filter-heavy reporting interface too slow to navigate.

The challenge was that enterprise FinOps users are not a forgiving audience. They conduct detailed investigations across large datasets. They need to compare, drill, and export. A chat interface that works for simple questions might fail precisely when the stakes are highest.

The design question was not just "can we build a chatbot?" — it was where does a conversational interface belong in a tool built around data density, and how do you design it so it earns trust rather than feeling like a consumer novelty bolted onto an enterprise product.

This project was unusual: I held the Design Lead and Engineering Manager roles simultaneously. That dual position shaped every decision in ways that a single-function role wouldn't have allowed.

On the design side, it meant being able to scope features to what was technically achievable in the time available — rather than designing optimistically and watching scope get cut later. On the engineering side, it meant the team built toward a clear design intent rather than interpreting ambiguous specs.

The tradeoff was depth. Holding both roles required constant triage: which design directions needed more exploration, and which technical constraints needed an answer before the research could move forward. The direction the research landed on reflects those tradeoffs — not the most expansive version of the concept, but the one most likely to be trustworthy in an enterprise context.

The chat-style interface suits straightforward queries well. "What are my top 5 idle EC2 instances?" works. The interface handles it cleanly.

The harder case is a FinOps engineer conducting a detailed investigation — comparing cost trends across multiple cloud providers over a 90-day window, drilling into specific accounts, and annotating findings for a stakeholder report. That workflow involves referencing multiple data points simultaneously, which chat history — scrolling off-screen — handles poorly.

The research acknowledged this explicitly rather than papering over it. The chat interface was positioned as a starting point for investigations, not the entire investigation. Its job would be to surface the right view in the existing product quickly — not to replace the reporting interface entirely. That framing made it possible to propose a direction without overpromising.