Skip to main content

Copilot Causal Toolkit

In short The Copilot Causal Toolkit is a set of Python / Jupyter notebooks that estimate the causal effect of Copilot usage on a workplace outcome using double machine learning (DML). Pick a scenario below, then follow the five-step path to a result.

It helps you move beyond correlation and answer questions such as:

Choose your scenario

Each scenario maps to one outcome variable and one or two notebooks. The notebook file names describe the outcome (e.g. AftCollabHours), not the scenario label — use these cards to pick the right one.

🤝

Seller Productivity

Outcome: External collaboration hours

CI-DML_ExtCollabHours_PQ.ipynb (Person Query)
CI-DML_ExtCollabHours_SUR.ipynb (Super Users Report)

🌙

Burnout Prevention

Outcome: After-hours collaboration hours

CI-DML_AftCollabHours_PQ.ipynb (Person Query)
CI-DML_AftCollabHours_SUR.ipynb (Super Users Report)

💬

Employee Engagement

Outcome: Ordinal survey metric (e.g. eSat)

CI-DML_Engagement_PQ.ipynb (Person Query only)

All three use Copilot usage (Total_Copilot_actions_taken) as the treatment variable. The Employee Engagement notebook is a template — because Glint survey metrics vary by organization, you update the outcome name, scale, and confounders to match your survey before running. See Preparing your data for the columns each scenario needs.

The path to a result

  1. Set up your environment Download the toolkit, install Python + the required packages, and open the folder in your editor. → Setup & installation
  2. Prepare your data Export a Person Query (recommended) or a Super Users Report, and include the columns your scenario needs. → Preparing your data
  3. Configure the notebook Edit a handful of parameters in the notebook cells — file paths, attributes, and date range. → Configuring the notebook
  4. Run the analysis Run cell-by-cell the first time, then all at once. Troubleshoot common errors. → Running & troubleshooting
  5. Interpret the outputs Read every plot, table, and sensitivity metric the toolkit produces. → Interpretation Guide

Continue reading

🧰 1 · Setup & installation Download the toolkit, install Python and the required packages, and open the project. Start → 📤 2 · Preparing your data Export from Viva Insights or a Super Users Report, and the columns each scenario needs. Explore → ⚙️ 3 · Configuring the notebook The handful of parameters to edit before running, plus a pre-flight checklist. Explore → ▶️ 4 · Running & troubleshooting Run cell-by-cell or all at once, fix common errors, and an FAQ. Explore → 📈 5 · Interpreting the outputs A file-by-file walkthrough of every plot, table, and sensitivity metric. Explore → 🔬 How it works The DML and causal-forest methodology, assumptions, and limitations. Explore →

What this analysis can and cannot prove

Read before you act on results This toolkit estimates the causal effect of Copilot usage under the assumptions of double machine learning — chiefly unconfoundedness (all relevant confounders are measured and included), overlap (both Copilot users and comparable non/low-users exist across covariate values), and a correctly handled treatment definition.

When those hold, it can support statements like “for comparable employees, higher Copilot usage is associated with a change of X in the outcome that is plausibly causal.” It cannot prove causation if important confounders are unobserved (e.g. unmeasured motivation or role changes), if there is no overlap between treated and untreated groups, or if Copilot usage is itself driven by the outcome (reverse causality).

Treat the estimates as decision-support evidence to be triangulated with experiments and domain knowledge, not as definitive proof. The Interpretation Guide discusses these caveats per output.

New to causal inference more broadly? Start with the conceptual Causal Inference in Copilot Analytics guide.

Last updated: Jun 22, 2026 Edit this page on GitHub