# Enterprise AI adoption measurement

Connect successful tasks, answer quality and repeat use to a testable retention story.
Test activation first: a trial sized for its rollout rule needs about 360 accounts under hypothetical assumptions.

For data leaders deciding what to instrument and whether a trial is feasible. Independent portfolio work. No official affiliation, endorsement, employer data or new career claim.

## Read and reproduce

Open `index.html` directly. All proposed definitions, experiments, calculation outputs and local method snapshots work offline; JavaScript only runs the theme control. Expand the details panels for the instrumentation and analysis contract. For an HTTP preview, run `python3 -m http.server 8000 --bind 127.0.0.1 --directory .` from this directory and open localhost on port 8000. The header and footer links to the portfolio home resolve when the page is served inside the portfolio.

Pinned runtime: Python 3.12.3, standard library only. No install step, API key or downloaded dataset is needed. From any directory run the absolute path to `check.py`; from this directory:

```sh
python3 -B check.py
```

This runs 19 tests, reproduces all three static pages and the power JSON in memory, checks the README calculation paragraph below against the same templates, checks the vendored design-system files and the frozen snapshots, and resolves local links. It changes no files. `python3 -B power.py` prints the full calculation; it never makes an API call. The render module is read-only by default.

## Real versus hypothetical

The example assumes grounded retrieval, shared knowledge, reusable agent capabilities and workflow automation. These capabilities define a hypothetical workplace assistant, not measured outcomes. All enterprise rates, margins, cluster sizes and trial durations are explicit hypothetical planning assumptions. Proposed event names describe instrumentation to build, not telemetry already available.

The base E1 calculation assumes activation changes from 30% to 35%, 80% target power, two-sided alpha 0.05/3, equal account allocation, exactly 20 users per account, ICC 0.05 and no attrition. Its rollout rule needs the lower bound of the 98.33% interval above zero. It yields 180 accounts per arm, 360 total accounts and 7,200 users; the chance of a rollout decision is 80% if the true gain is 5 points. An earlier version also required the estimate itself to reach 5 points; with the same size that rule's chance was 50%, and keeping it would need 610 accounts at a true 6-point gain and 186 accounts at a true 7-point gain. Sensitivity rows use ICC 0.00, 0.05, 0.10, 0.20, 0.30 and one row with unequal account sizes (CV 1.0, 542 accounts), every other assumption fixed.

These counts are planning approximations, not observed sample supply or a guarantee of achieved power. Guardrail tests are not sized, and requiring them to pass can only lower the chance of a rollout. E2, E3 and guardrails are not sized here. The earlier assumptions and calculation output are kept unchanged in `results-history/`.

Activation has one definition: the first qualifying task completes within 7 days of assignment and is not reopened in the next 7 days, so every outcome is final on day 14.

The two `related/` pages are static projections of earlier reviewed portfolio results. The mobility study uses public Chicago data; its source notice is included. The marketplace study is entirely synthetic. Their JSON projections, source release hashes and selected fields are recorded in `dependencies.json`. They illustrate paired evaluation and interference; neither supplies an enterprise power input. The original interactive evaluation and executable simulation are outside this snapshot. No raw trip data or model-response cache is copied here.

## Method and limitations

The calculation uses a two-proportion normal approximation and multiplies variance by the design effect of a fixed-weight user average across accounts, 1 + ((1 + CV²) × m − 1) × ICC, which is the equal-cluster design effect when account sizes are equal. It rounds up to whole accounts, omits continuity and small-cluster corrections and ignores the opposite rejection tail in planning. The tests include an independently inverted normal-tail calculation, the R documentation's rounded reference, covariance enumeration for equal and unequal account sizes, the decision table against a direct search, a seeded simulation of both rollout rules (seed 20261010, 2,000 simulated trials, tolerance 0.04 fixed before the first run), arm symmetry, monotonicity, rounding, invalid inputs and a check that the page prose follows the assumptions. Full formula and variance assumptions appear on the page.

Account sizes, intraclass correlation, recruitment, outcome maturity, missingness and guardrail precision need measurement before a trial. Account connectivity can invalidate independent clusters. The final estimator and power must be validated together using the actual account distribution. Three primary tests use Bonferroni allocation; secondary metrics remain exploratory. No forecast of enterprise revenue or treatment benefit is made.

The tree is a causal story to test, not proof that adoption causes renewal. Month one is devoted to replacing assumptions. If recruitment or measurement is inadequate, the fallback is a descriptive pilot without a causal claim.

![Proposed metric relationships](assets/metrics-tree.svg)

## Design

The pages use the portfolio design system, release 2.0.4, vendored unedited under `vendor/viz-design-system/` with its fonts, licences and adoption record (`vendor/viz-design-system-adoption.json`).

## Provenance and timing

Started October 8, 2026. Actual elapsed wall time to the first complete static page: 18.24 minutes. Elapsed time through the browser and initial complete-copy checks: 31.77 minutes. On October 11, 2026 the sizing rule, sensitivity rows and page design were revised; that rework's elapsed time is recorded in `build-info.json`. These are observed build intervals, not delivery promises.