AI Analyst Builder

The Problem

A business question shouldn't take a week to answer

The standard cycle for an IC analyst with no dedicated data engineering team, told in six days.

Monday Business question surfaces in a meeting. Day 1

Tue–Wed Data exploration: find the right tables, understand the schema, write exploratory queries. Day 2–3

Thursday Build charts, validate numbers, spot-check aggregations. Day 4

Friday Write the narrative, build the deck. Day 5

Following Monday Present findings — often after the decision has already been made on gut feel. Day 6

→ With AI Analyst Builder, the same cycle runs in under 2 hours, target across the first 10 live analyses.

The Pipeline

Frame. Analyze. Story. Deck.

Independent agents resolve dependencies and run in parallel, up to three at once. Quality gates checkpoint the work before it moves to the next phase.

Frame

Frame the business question
Generate testable hypotheses
Checkpoint before touching data

Analyze

Explore data, verify integrity
Drill down to root cause
Validate findings, size opportunity

Story

Design Context-Tension-Resolution beats
Generate SWD-styled charts
Review visuals against the checklist

Deck

Write the narrative
Build the branded slide deck
Export as PDF or HTML

Specialized agents

Validation layers

606

Tests passing

Slash commands

Every finding passes structural, logical, business-rules, and Simpson's-paradox checks before it reaches a slide. Grade C and below stays internal — only grade B and above is safe to share with a stakeholder.

The Moat

After 50 analyses, the system knows your schema, your metric definitions, your common errors, and your business context. A generic SaaS tool starts from scratch every session.

Corrections, query archaeology, and the business glossary compound — every logged fix and proven SQL pattern makes the next analysis faster and more accurate.

The Roster

Twelve of the eighteen agents on shift

Each agent is a markdown prompt template with a defined contract — read, substitute variables, execute. Editable, not a black box.

Hypothesis

Generates testable theories across product, technical, and external causes.

Data Explorer

Profiles schema, quality, and gaps before analysis begins.

Source Tie-Out

Cross-checks pandas against SQL before analysis proceeds. Halts on mismatch.

Descriptive Analytics

Segments, funnels, and decomposes the metric in question.

Root Cause Investigator

Drills down through dimensions to a specific, actionable cause.

Validation

Runs the four-layer check and assigns a confidence grade.

Opportunity Sizer

Quantifies business impact with sensitivity analysis.

Story Architect

Builds Context-Tension-Resolution narrative beats from findings.

Chart Maker

Builds each chart to Storytelling-with-Data standards.

Visual Design Critic

Reviews every chart and deck against the SWD checklist.

Deck Creator

Assembles the branded Marp deck with speaker notes.

Comms Drafter

Drafts the Slack summary, email brief, and exec version.

In The Terminal

One question in. One slide out.

Terminal

$ /run-pipeline data_path=data/checkout/ question="Why is conversion dropping?"

▸ FRAME    4 hypotheses generated
▸ ANALYZE  root cause isolated — mobile checkout latency
▸ STORY    6 narrative beats mapped
▸ DECK     12 slides, speaker notes attached

✓ Confidence grade: B+
✓ 606 tests passing · 4-layer validation clear

Output — Slide 4 of 12

Checkout · Mobile

Conversion dropped 8pts after the mobile redesign shipped

Source: checkout_events, Mar 1–Apr 30 · n = 412,880 sessions

Validated Root cause: latency Grade B+

Pipeline Modes

Run the whole thing, or just the part you need

full_presentation Complete analysis to slide deck, start to finish. 18 agents

deep_dive Analysis only — no presentation layer. 12 agents

quick_chart Just make one chart, straight to Chart Maker. 1 agent

refresh_deck Re-do the presentation using existing analysis. 4 agents

validate_only Check existing work against the four-layer gate. 3 agents

Why This, Not…

Question → validated root cause → branded deck. Nothing else goes end to end.

Tool	What it does	What's missing
ChatGPT / Claude (chat)	Generates SQL and charts on request.	No memory across sessions. No validation. No business context. Restarts from zero every time.
Claude Code alone	Writes the SQL you describe.	A coding assistant, not an analyst. You drive every step manually. No framing, no validation pipeline, no deck output.
BI tools (Looker, Tableau)	Persistent dashboards.	Answers "what" happened, not "why." No narrative, no root cause, no recommendation.
AI SQL tools (Julius, Noteable)	Natural language to SQL.	No validation layer, no framing, no storyboard, no deck. Closed SaaS — no customization.
Jupyter + Copilot	AI-assisted notebook.	Still requires the analyst to drive every step. No orchestration. No deck output.
AI Analyst Builder	Question → validated root cause → branded slide deck.	Automated, resumable, and self-learning — corrections logged today sharpen every analysis after.

Success Metrics

What "working" means, before it ships

Targets for the first 10–60 days post-publish — not results claimed yet.

<2h

Time-to-deck · from ~5 days

Target · first 10 runs

80%

Pipeline success, no halts

Target · 60 days

70%

Chart pass rate, first try

Target · 60 days

200

GitHub stars

Target · 90 days

Get Started

Three steps to your first deck

Bring your own data — CSV, DuckDB, Postgres, BigQuery, or Snowflake. No bundled datasets.

Install Claude Code

npm install -g @anthropic-ai/claude-code

Clone and set up

git clone ai-analyst-lab/ai-analyst
cd ai-analyst && pip install -e ".[dev]"

Connect data, ask a question

/run-pipeline question="…"

Before You Start

You are the expert. This handles the other 80%.

It catches data mistakes, validates findings, and builds the deck. You validate the output — correct it once, and it remembers.

Read the Setup Guide