AI Workforce Analytics

The problem: Every company is pouring money into AI tools — Claude, Copilot, Cursor, ChatGPT — and nobody can answer the basic question: is it working?

The existing observability tools (LiteLLM, Helicone, Portkey, Langsmith) track API requests from applications. They tell you "API key xyz consumed 500K tokens." They can't tell you which person used those tokens, on which project, whether they picked the right model, or whether the $200/mo subscription is actually delivering value.

There's no system of record for how humans use AI. Not backend API calls — how people use AI as part of their daily work.

What We're Building

A workforce-level analytics layer for AI tool usage. The unit of analysis isn't the API request — it's the human AI session: who was working on what, in which tool, for how long, using which model, and was it a good use of resources.

The vision is tool-agnostic: Claude Code today, Claude Desktop and Cowork tomorrow, Cursor and Copilot after that. One dashboard that shows a manager — or a power user — exactly how AI is being used across their team.

First PoC: Chalk (Claude Code + Cowork + Desktop Chat)

The chalk/ directory contains a working proof of concept that reads session data from Claude Code, Cowork, and Desktop Chat — no proxy, no traffic interception — and produces detailed usage analytics with rich reporting.

Real numbers from the first user (Jan–Mar 2026):

The PoC immediately surfaced actionable insights: a single project consuming 67% of all cost, weekly trend analysis showing productivity patterns, 542 sessions spawning subagents for parallel work, and 48 Cowork sessions that were previously invisible.

Session Summarizer & Maturity Scoring

A two-phase batch summarizer produces per-session reports combining local analysis with AI-generated summaries:

• Local analysis — session health, tool leverage, workflow maturity scoring, key actions, files modified, and improvement tips — all computed without any API calls
• AI summaries — concise session descriptions generated via the Anthropic Messages API (Haiku by default), rate-limited and cached so subsequent runs skip unchanged sessions
• Maturity scoring — each session is rated Exploratory → Ad-hoc → Structured → Optimized based on 7 best-practice checks (planning, scope control, tool leverage, verification, etc.)
• Client-type detection — automatically classifies sessions as Claude Code, Cowork, or Desktop Chat based on tool signatures and working directory patterns
• Subagent tracking — Task tool spawns are counted per session, with parallel execution detection and per-project breakdowns

Rich Batch Report

The --report flag generates a comprehensive analytics report with:

• Weekly Activity Trend — last 12 weeks with quality and high-productivity percentages, trend comparison
• Cost Per Project — top 10 projects by API spend with percentage breakdown
• Session Complexity Scoring — weighted formula across duration, tools, compactions, MCP usage, and messages; 4 tiers from Lightweight to Heavy
• Subagent Analysis — per-project subagent usage, distribution histogram, parallel execution detection
• Client Type Breakdown — Claude Code vs. Cowork vs. Desktop Chat distribution

See chalk/README.md for setup and usage, or concept-doc.md for the full product thesis and technical architecture.

Team Layer (NEW)

The Team Analytics Layer enables managers to answer "How is my team using Claude Code?" while preserving developer privacy for personal projects.

Quick Start

# Install
pip install chalk
chalk install

# Local usage (no account needed)
chalk report

# Cloud sync (optional)
chalk login
chalk sync

# Join a team
chalk team join <invite-code>
chalk share-patterns import-suggested
chalk sync

Features

For Individual Developers:

• Local reports — chalk report works instantly, no account needed
• Optional cloud sync — View your data in a web dashboard at chalk.io
• Privacy control — Configure sharing patterns to separate work/personal projects

For Team Leads:

• Team dashboard — See aggregate stats: members, sessions, cost
• Member breakdown — View individual contributions (work projects only)
• Suggested patterns — Set default tracking patterns for your team
• Sync compliance — Track who's syncing and when

Privacy Model

All data syncs to the server, but views are filtered by role:

Work patterns use glob syntax: github.com/acme-corp/* matches all repos under the org.

Architecture

• CLI — Python package with auth, sync, team commands
• Server — FastAPI + Supabase (Postgres with Row Level Security)
• Dashboard — Next.js web app for visualization

See server/migrations/ for the database schema.

Where This Goes

Individual → Team → Org. The team layer is now implemented! Team leads see per-person breakdowns, adoption insights, and cost tracking across their engineering org while respecting developer privacy.

Claude Code → Everything. The PoC already captures Claude Code, Cowork, and Desktop Chat sessions from separate storage locations. The same model extends to API usage and eventually third-party tools like Cursor and Copilot. One system of record for all AI activity.

Built by Jeremy Greven and Ryan Meyers.