AI-native Analytics Agents

The problem

A FinOps analyst's day is a loop: someone asks a cost question, the analyst finds the right data, runs the analysis, spots what matters, builds the dashboard, repeats. The loop doesn't scale — every new question costs analyst hours, and most customers don't have analysts to spare.

What the agents do

From a single natural-language request, the agents interpret the question, route it to the right analysis and data, analyze cost and usage, surface anomalies and optimization opportunities, and generate dashboards. Watch the shape of a run:

What I owned

• Intent & analysis routing — how the agent decides what to analyze and which data to reach for, grounded in FinOps domain knowledge.
• Human-in-the-loop review workflows — an analyst stays in control of what reaches the customer; the agent drafts, the human approves.
• The prototype itself — I built the working prototype hands-on in Claude Code and Figma Make. Engineering shipped to production from that prototype. This is the clearest proof of how I like to work: prototype first, spec second.

Why this matters

Most “AI agent” stories are demos. This one is a production system with guardrails — and the design questions that mattered were product questions, not model questions: when should the agent act autonomously vs. draft for review? How does domain knowledge get encoded into routing? What does a customer need to see to trust an agent's answer?

Scope, honestly

The run above is a simulation built for this site — the shape and pacing are faithful, the numbers are invented. The production system is real, and it ships nothing without an analyst's approval. “Human in the loop” here is a design decision baked into the workflow, not a disclaimer bolted on after.