# 90-day AI coding measurement plan

## Decision to make

- Team / workflow / adoption date:
- Comparable work categories and excluded work:
- Baseline and observation windows:
- Decision owner and review dates:
- Hypothesis and expected product outcome:
- Changes in staffing, work mix, architecture or process:

## Collect consistently

| Category | Unit / denominator | Source | Owner | Baseline | Observation |
| --- | --- | --- | --- | --- | --- |
| Tool and model spend | Actual currency / period | Billing | | | |
| Authoring | Active hours / comparable change | Work sample | | | |
| Review | Active hours and elapsed wait, separately | Review sample / repository | | | |
| QA | Active hours and elapsed wait, separately | QA / CI | | | |
| Rework | Non-overlapping active hours / period | Reopened work | | | |
| Incidents and support | Non-overlapping hours / period; incident counts and severity | Incident / support tools | | | |
| Delivery | Lead time and deployments / period | CI/CD | | | |
| Change failure | Failed changes / deployed changes | Incident-to-change mapping | | | |
| Customer outcome | Defined outcome / eligible exposed unit | Product analytics / experiment | | | |

Days 1–14: collect or reconstruct a defensible baseline; if no historical baseline exists, label the observation-only design. Days 15–45: compare consistent categories and inspect burden transfer. Days 46–75: evaluate an explicit workflow intervention. Days 76–90: decide and document uncertainty, ongoing maintenance and follow-up.

Separate active labor, waiting time, tool spend and realized revenue. Categorize an hour once. Do not count both loaded wages and an incident labor estimate for the same work. Released capacity is not automatically a cash saving. A before/after comparison alone does not identify AI's causal effect.