How Cursor transformed our analytics & insights operating model

Most data teams are not getting the most out of AI yet

The standard workflow today looks something like this: open a browser chatbot, type a prompt, copy the generated SQL, paste it into a warehouse console, fix the syntax error, run it, paste the output into Excel, and manually build a slide deck.

This process isn’t just slow. It’s risky. Without warehouse context, AI models frequently invent logic, hit deprecated tables, or violate subtle business constraints (like using raw ARR > 0 instead of a governed measure like current_arr >= 300).

Over the past year, we set out to solve this by evolving Cursor from an AI code helper into a full-stack Analytics Operating System. By unifying code, context, live tools, and executive delivery inside a single environment, we have compressed multi-day cross-system analyses into hours while raising data reliability.

Here is a breakdown of how the system works.

1. Warehouse convergence: behavioral telemetry meets commercial reality

The foundation of any reliable analytics workflow is unified context. In our architecture, Fullstory Anywhere: Warehouse data lives in BigQuery alongside Salesforce CRM, ARR History, Contract SKU entitlements, internal ops data, and more. Because Cursor operates as a multi-root workspace with access to both our terminal and production repositories, the agent doesn't guess schemas:

• Direct Codebase Inspection: It reads production LookML views directly to trace how metrics, dimensions, and joins are modeled upstream.

• Live Query Execution: It uses the bq CLI to execute SQL directly, joining granular behavioral telemetry (session replays, event logs, tool calls) to paying account tiers and CSM assignments.

• Multi-System Orchestration: Cursor queries data across systems via Model Context Protocol (MCP) servers, including Fullstory MCP for in-app funnels, Looker MCP for dashboard parity checks, and Clozd MCP for win/loss context.

2. Eliminating metric drift through durable governance

The biggest risk with AI in data engineering is "metric drift": getting a fast, plausible answer that silently breaks governance rules. To prevent this, we maintain a dedicated fullstory-insights workspace repository containing three pillars of governance:

A. The living glossary (glossary.md)

A markdown-based repository of truth that codifies canonical business logic: paying customer filters, SKU history attach rules, adoption thresholds, and partition horizons.

B. Always-on system rules

Custom Cursor rules enforce strict pre-query protocols:

1. Confirm account grain and snapshot windows before writing SQL.

2. Forbid hardcoded headcounts or unverified assumptions.

3. Require explicit SQL output and sanity checks before drawing narrative conclusions.

C. Compounding memory

Whenever an analyst resolves an edge case or data anomaly, a custom capture-durable-context skill proposes an update to our glossary and runbooks. When we accept it, the next analysis inherits it.

3. Reusable skills, subagents, and last-mile executive delivery

Instead of prompting from scratch, complex analyses are packaged into 16 Runbooks, 9 Custom Skills, and 4 Specialized Subagents:

• bigquery-analyst: Handles SQL optimization and partition-aware query writing.

• looker-engineer: Validates LookML joins and dashboard consistency.

• metrics-strategist: Frames business questions and hypothesizes drivers.

• executive-analyst: Synthesizes data into executive-ready narratives.

Automated executive drumbeats

Weekly Business Review (WBR) slide updates, monthly metric bridges, and SKU one-pagers are generated via verified scripts with built-in benchmark validation.

Causal sensemaking

When asked “Why did metric X move?”, the system uses a custom sensemaking skill grounded in Klein’s Data-Frame theory. It formulates and tests competing hypotheses across BigQuery, Fullstory, Looker, Salesforce, and Clozd before committing to a final conclusion.

Board-ready deliverables

The workflow extends all the way to the final presentation layer. Cursor generates custom-branded PDFs (complete with multi-pass visual page checks by the agent) and interactive Cursor Canvases, with single-command publishing to Notion and Slack.

The bottom line

This operating model transforms a single analyst into a full-stack insights engine, handling everything from raw data validation and LookML modeling to strategic framing and executive communication. This is uniquely possible with Anywhere Warehouse + MCP.

More importantly, because all rules, glossaries, and skills are stored as version-controlled code in a shared repository, the entire system is completely portable, repeatable, and scalable across the team.