Your dashboards tell you who touched the tool. They cannot tell you whose work changed. Adoption telemetry closes that gap — computing change-management stage progression directly from the usage signals your AI deployment already produces, and mapping every stall to a class of intervention.
of generative-AI pilots reported no measurable P&L impact, in one widely cited and widely debated study.
of companies abandoning most of their AI initiatives — a jump of one year.
of AI users reach the stage where AI is consistently embedded in their workflows.
The reported causes of failure are organizational — workflow, sponsorship, behavior — not model capability. The diagnosis is consistent. What’s been missing is an instrument.
NANTE classifies a deployment population — never individuals — into five stages of behavior change, computed deterministically from telemetry an enterprise already generates. Thresholds are public, falsifiable, and open to dispute.
nante (Twi): “walk” — as in nante yiye, walk well. The model measures a population’s walk through a change.
In the paper’s synthetic simulation, two cohorts a usage dashboard calls identical — over 98% of both reach recurring use. One is integrating the tool into how work gets done; one has stalled at the surface. NANTE separates them.
Healthy: 25.1% reach depth (Transform + Embed) · composite 60.5 · no stall flag. Stalled: 0% past Navigate · composite 49.9 · shallow-plateau flag → workflow redesign, not more licenses. Figures are the reference and shallow-plateau cohorts from the paper’s evaluation (synthetic populations; reproducible via make results).
Adoption telemetry: continuous measurement of behavior change from production signals, interpreted through an explicit model of change — distinct from evals, usage analytics, product analytics, and survey-based change management.
NANTE: five stages, six diagnosable patterns (one healthy, five failure modes), defined thresholds, and a stall-to-intervention map — so measurement ends in action, not a score.
agent-adoption-kit (Apache-2.0): a local pipeline over your own exports. Every table and figure in the paper regenerates from the open-source code. Raw events never leave your environment.
The evaluation demonstrates computability and discrimination on synthetic populations. Every threshold is a proposed default, not a calibrated value; validating stages against real outcomes is the paper’s stated research agenda — and the reason the research collaboration (design partners) below exists.
The framework’s next step requires what no synthetic population can supply: production deployments with observed outcomes. Polywise is convening a small group of organizations running enterprise AI at cohort scale (50+ seats) to validate and calibrate NANTE against reality.
Participation is scoped, confidential, and cohort-level only — NANTE never scores individuals.
Young, D. A. (2026). Adoption Telemetry: Measuring Enterprise AI Adoption from Production Signals. Zenodo. https://doi.org/10.5281/zenodo.21943954. arXiv:2608.23617
@article{young2026adoption,
author = {Young, Damon A.},
title = {Adoption Telemetry: Measuring Enterprise AI Adoption
from Production Signals},
year = {2026},
eprint = {2608.23617},
archivePrefix = {arXiv},
primaryClass = {cs.HC},
publisher = {Zenodo},
doi = {10.5281/zenodo.21943954},
url = {https://doi.org/10.5281/zenodo.21943954}
}