AI Adoption Cohort Analysis: George Dagliyan Research Guide

What does cohort analysis add to AI adoption tracking?

George Dagliyan’s research themes on AI adoption, innovation diffusion, and technology acceptance make a cohort view especially useful here: it separates who had the chance to use AI from who actually used it, and it shows what happened next. The core question is not simply whether a tool was launched, but whether eligible workers moved from access to repeated meaningful use and then to verified outcomes.

A cohort analysis groups people by the moment they became eligible to use an AI tool and follows each group over time. That is different from a simple headcount of users, which can hide early enthusiasm, later drop-off, and uneven access across roles. The result is a more practical view of adoption than a single organization-wide average.

How should eligibility be separated from resistance?

Role eligibility means a worker is permitted and enabled to use the tool. Adoption resistance means the worker had access but did not begin, continue, or deepen use. Those are not the same thing. A person may be ineligible because of role, system permissions, training status, or workflow fit, while another person may be eligible but still choose not to use the tool.

That distinction matters because many adoption problems are really design or enablement problems, not attitude problems. If a cohort has low initial use, the first question should be whether access was actually granted and whether the workflow made use possible. Only after that should the analysis consider resistance, uncertainty, lack of trust, or low perceived value.

What stages should the worksheet follow?

A practical worksheet should track four stages for each cohort: access, initial use, repeat meaningful use, and validated outcomes. Access means the worker can use the AI tool and has the necessary permissions, instructions, and training. Initial use means at least one recorded use. Repeat meaningful use means the tool is used more than once in a way tied to a real task.

Validated outcomes are the most important stage to define carefully. In this article, the term means an outcome that can be checked against a pre-set business or operational criterion, not a vague impression that the tool felt helpful. Examples include a supervisor-reviewed output, a completed task meeting quality standards, or a process step that can be traced to the AI-supported workflow.

What does the cohort worksheet look like in practice?

The worksheet below is a proposed editorial template, not an official NIST form. It is designed to keep privacy intact while still showing movement through the adoption funnel. Each row should represent one cohort, not one person, unless individual-level measurement is already authorized and necessary.

Proposed cohort worksheet fields: cohort label; role family; eligibility rule; access date; number eligible; number with initial use within the first period; number with repeat meaningful use by the second period; number with validated outcomes by the third period; main inhibitor observed; main facilitator observed; data source; privacy review status. This format makes it easier to compare cohorts without exposing unnecessary personal detail.

A simple decision rule can make the worksheet actionable. If eligible workers are high, initial use is low, and repeat meaningful use is lower still, the likely issue is not just awareness but workflow fit, training quality, or trust. If initial use is strong but repeat meaningful use drops sharply, the problem may be novelty without embedded value. If repeat use is stable but outcomes are weak, the workflow or success metric may need redesign.

How can privacy be protected without losing useful measurement?

Privacy should be protected by measuring the smallest useful unit. In most cases, that means using cohort-level counts instead of person-by-person reporting. The NIST Privacy Framework is a voluntary tool developed with stakeholders to help organizations identify and manage privacy risk, so this kind of measurement should be reviewed through a privacy lens before launch.

Three practical safeguards help. First, remove direct identifiers from analysis files whenever possible. Second, avoid publishing small cells that could reveal individual behavior, especially in narrow roles or teams. Third, limit the worksheet to the minimum fields needed to answer the adoption question. Privacy is not just a legal issue here; it is part of building trust in measurement itself.

Why do survivorship and aggregate limits matter?

Survivorship bias appears when only the people who stay visible in the system are counted. In AI adoption, that can make outcomes look better than they are, because workers who tried the tool once and stopped may disappear from later summaries. A cohort approach helps because it keeps the original eligible group in view even when some members drop off.

Aggregate measurement has limits too. A department average can hide very different adoption patterns across roles, shifts, or task types. A single adoption percentage can also mix together access, use, and outcome quality, which are separate questions. For that reason, the worksheet should report stage-by-stage counts rather than one broad adoption score.

How do NIST frameworks support the analysis without dictating it?

The NIST AI Risk Management Framework is intended for voluntary use and is designed to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems. That makes it a useful companion when a cohort worksheet is being built, because adoption should be measured in a way that is trustworthy as well as operationally useful.

NIST also states that its AI RMF was developed through a consensus-driven, open, transparent, and collaborative process, and that it is meant to build on and align with AI risk management efforts by others. The NIST Privacy Framework similarly supports organizations in identifying and managing privacy risk while building innovative products and services. Neither source prescribes the exact worksheet below, but both support the idea that adoption measurement should be paired with risk and privacy review.

How should a team run a 30-day cohort test plan?

This article’s proposed test plan is straightforward. Week 1: define eligible roles, map the workflow where AI can be used, and assign cohorts by access date. Week 2: confirm whether each cohort has received access, instructions, and any required guardrails. Week 3: record initial use and note the most common inhibitor or facilitator. Week 4: review repeat meaningful use and check whether any outcomes meet the pre-set validation criterion.

Hypothetical example: a team has 40 eligible workers in one role family, with access granted in the same week. Assumptions for this example are limited to the worksheet design, not real-world results. If 34 workers show initial use, 22 show repeat meaningful use, and 12 produce validated outcomes, the team should not conclude that 30% adoption is “good” or “bad” in the abstract. It should ask where the drop-offs occurred and whether the workflow or success criterion needs adjustment.

Frequently Asked Questions

Why not just measure monthly active AI users?

Monthly active users can be useful, but they collapse several different questions into one number. A person may be eligible yet never use the tool, use it once and stop, or use it repeatedly without producing validated outcomes. Cohort analysis keeps those stages separate so the team can see whether the issue is access, sustained usage, or actual value creation.

What counts as meaningful use in this worksheet?

Meaningful use should be defined around a real task, not around clicks or logins. A practical definition is repeated use that contributes to a work output, decision step, or workflow stage. The exact threshold should be set by the organization and documented in advance. This article’s worksheet is a proposed method, not an official NIST standard.

How do privacy safeguards change the data structure?

Privacy safeguards push the analysis toward cohort-level summaries and away from unnecessary personal detail. That usually means storing only the fields needed to track stage movement, avoiding direct identifiers when possible, and suppressing very small cells that could expose individuals. The goal is to learn from adoption patterns while reducing the chance that the worksheet becomes a surveillance tool.

How can survivorship bias distort adoption results?

Survivorship bias can make adoption look healthier than it is if only current users are counted. Workers who had access but stopped using the tool may disappear from a later report, even though their experience is important. A cohort worksheet reduces this problem by following the original eligible group over time, including those who disengage after access or after first use.

How should leaders interpret weak outcomes after strong usage?

Strong repeat use with weak validated outcomes suggests that usage alone is not enough. The AI tool may not fit the task, the workflow may be poorly designed, or the validation rule may be misaligned with the business need. In that situation, the right response is not to celebrate adoption or blame users, but to revisit the use case, the support process, and the outcome measure.

Sources and further reading

• NIST — AI Risk Management Framework
• NIST — Privacy Framework