Adoption Metrics for Claude, Cursor, Copilot & Codex
What HR and execs should track to measure adoption of Claude, Cursor, Copilot, and Codex. Tool-specific metrics, ROI signals, and how to attribute usage.
Most tech-forward companies in 2026 are paying for at least three AI tools at once. Most can’t tell you which ones are actually being used. License costs add up fast, and adoption-by-vibes isn’t a defensible answer when finance asks for ROI.
This article breaks down what to track for each of the four major enterprise AI tools: Claude, Cursor, GitHub Copilot, and Codex. The metrics differ by tool category. Coding assistants need different signals than chat assistants. For the broader framework behind these metrics, see our guide to measuring AI tool adoption.
Quick Comparison
| Tool | Tool Type | Key Adoption Metrics | Enterprise Analytics API |
|---|---|---|---|
| Claude | Chat, code, office | Messages, artifacts, Code sessions, edit acceptance, office sessions | ✔ |
| Cursor | Coding assistant | Agent edits, tab completions, acceptance rate, chat mode | ✔ |
| GitHub Copilot | Coding assistant | Active users, completions accepted, chat usage, PR assist rate | ✔ |
| Codex | Autonomous coding | Threads, turns, tokens, client surface | ✔ |
Claude: Chat, Code, and Office Agent Metrics
Claude is the trickiest tool to measure because it spans three surfaces: chat, Claude Code, and office agents for Excel and PowerPoint. Each surface needs its own metric set. Aggregating them into a single “Claude adoption” number loses too much signal.
Chat (claude.ai): conversations started per user, messages sent per active user, artifacts created (the strongest depth signal), and skill/connector adoption.
Claude Code: sessions per engineer per week, pull requests assisted, lines added and removed, edit acceptance rate (target above 60% for a healthy pattern).
Office agents: sessions, skill invocations, connector usage across Excel and PowerPoint.
Claude exposes all of this through its Enterprise Analytics API. Windmill’s Claude integration reads it directly without touching conversation content or output.
Cursor: Edits, Completions, and Agent Mode
Cursor’s analytics are organized around four interaction modes: agent edits, tab completions, chat, and plan mode. The metric that matters most is acceptance rate, which tells you whether the AI’s suggestions actually fit the engineer’s workflow. Acceptance rate below 30% usually means the engineer hasn’t found a use case where Cursor reliably helps.
The metrics to track:
- Agent edit acceptance rate (primary signal for stickiness)
- Lines added and deleted per accepted edit (depth)
- Tab completion usage (inline autocomplete frequency)
- Chat and plan mode usage, broken down by model
- MCP tool and extension adoption
Cursor’s annualized revenue hit roughly $2 billion by early 2026, with more than half of Fortune 500 companies running it. The cost-per-seat pressure makes adoption tracking a finance issue, not just a People Ops one. Windmill’s Cursor integration pulls these metrics directly via the org chart.
GitHub Copilot: Completions, Chat, and PR Assist
GitHub Copilot has the largest installed base of any AI coding tool: 4.7 million paid subscribers as of January 2026, with 29% of developers using it at work per JetBrains’ 2026 developer survey. GitHub’s enterprise admin API exposes granular metrics by language.
Core metrics to track:
- Daily and weekly active users against provisioned seats
- Completions suggested versus accepted (acceptance rate)
- Acceptance rate by programming language — useful for stack fit
- Chat usage — sessions and messages per engineer
- Pull request assist rate — PRs where Copilot generated descriptions or commits
Teams on Copilot typically pull metrics through GitHub’s admin dashboard or a data warehouse. GitHub’s data is good, but stitching it together with your other AI tool data requires extra work.
Codex: Threads, Turns, and Client Surface
Codex is OpenAI’s autonomous coding agent, used through ChatGPT or IDE integrations. Its adoption pattern looks different from Cursor or Copilot because it runs tasks end-to-end rather than offering inline suggestions. Instead of per-keystroke acceptance, you track per-task throughput.
The metrics to track:
- Threads per active user per week (how often engineers hand work to Codex)
- Turns per thread (depth of interaction)
- Token and credit consumption
- Client surface breakdown (ChatGPT vs IDE vs other clients)
Codex revenue hit $2.5 billion annualized by February 2026. Codex’s Enterprise Analytics API exposes user-level usage, and Windmill’s Codex integration reads it without touching prompts or generated code.
Cross-Tool Metrics That Matter Most
When you’ve stitched all four tools together, three cross-tool metrics matter more than any single-tool number. These are the metrics finance and leadership actually want, and they’re impossible to compute without unified data.
- Total AI active users as a percentage of the org — your real adoption rate, not the per-tool one
- Tool overlap — how many people use two or more tools (an indicator that your AI stack has gotten too sprawling)
- Adoption by team or function — engineering at 95% while sales sits at 20% is a real finding that surveys often miss
For roles where AI use is part of the job, these metrics increasingly belong in performance reviews — not as a target to hit, but as context for understanding how someone gets their work done.
Pulling It All Together
Pulling data from one AI tool is straightforward. Pulling from four, attributing each user through your org chart, and producing managerial-grade reports is harder. The hard parts: identity matching across tools, org chart attribution, and trend visualization over time.
Windmill’s AI Adoption report pulls Claude, Cursor, and Codex natively and attributes through your org chart. For Copilot, teams typically supplement with GitHub’s admin dashboard or a data warehouse. Either way, the goal is the same: one view of who at your company actually uses each tool, broken down by team and trended over time. JetBrains’ 2026 survey found 90% of AI-using developers save at least an hour a week, with the top 20% saving 8 hours or more. The gap between average and top adopters is the lever managers can actually pull.
The Bottom Line
Single-number “AI adoption rate” metrics hide what’s actually happening. Each tool needs its own metric set, attribution has to flow through your org chart, and the data has to land in front of managers who can act on it. Without all three, you’ll renew licenses next year on the same set of vibes you renewed them on last year.
Ready to pull adoption data from Claude, Cursor, and Codex automatically? Book a demo to see Windmill’s AI Adoption report.
Frequently Asked Questions
What metrics should HR track for Claude adoption?
Track conversations started, messages sent, and artifacts created on the chat side. For Claude Code, track sessions, pull requests assisted, lines added or removed, and edit acceptance rate. For office agents, track sessions, skill invocations, and connector usage. Each Claude product needs its own metric set.
How do you measure Cursor adoption across an engineering team?
The four metrics that matter for Cursor are agent edit acceptance rate, tab completion usage, chat and plan mode sessions, and MCP tool adoption. Acceptance rate is the single best signal for whether the tool fits an engineer's workflow. Anything under 30% suggests the engineer hasn't found a use case that works for them.
What are the right metrics for GitHub Copilot adoption?
Track active users (daily and weekly), completions accepted versus suggested, chat usage per engineer, and pull request assist rate. GitHub's enterprise analytics also expose acceptance rate per programming language, which helps identify whether the tool fits your specific tech stack.
How do you measure Codex adoption?
Codex exposes thread and turn metrics that show how engineers use the tool for autonomous coding tasks. Track threads per active user, turns per thread, token consumption, and client surface breakdown — whether engineers use Codex through ChatGPT, IDE integrations, or other clients.
Can you measure adoption across multiple AI tools at once?
Yes, but you have to pull data from each tool's enterprise analytics API and attribute usage to employees through a unified org chart. Most teams either build this themselves with a data warehouse and BI tool, or use a people analytics platform that pulls the integrations natively.