Are Employees Actually Using Generative AI? Measure Access, Behavior, Trust, and Avoidance Separately

When a major AI release dominates the news, many organizations rush to run an "employee AI usage and trust survey." If that survey only asks two vague questions—"Have you used AI at work?" and "Do you trust AI?"—the results will overstate adoption and will not say what actually needs to change. What looks like "employees distrust AI" is often simply that the tool was never enabled, the rules were never explained, or people did not know they were allowed to use it.

The fix is to split this compound question into four evidence layers measured separately: whether the organization authorizes and provides tools (availability), what employees actually do in concrete tasks (usage behavior), which object their trust refers to (trust), and where they actively avoid using AI (boundary concerns). Each layer needs its own question types, options, and denominators. Before writing items, write down the decision the study must serve and choose only the layers that support it, rather than packing all four into one questionnaire.

Decide First: Baseline, Adoption Drivers, or Governance Readiness

"Are people using AI?" sounds like one question, but for management decisions it is three. Match the study to the decision before writing items:

This framing also determines sampling. A pure usage baseline should measure behavior among all employees and use the whole population as the denominator. If the goal is to explain why usage differs across job functions, each function needs enough respondents for subgroup comparison instead of an even split of a few hundred invitations.

Four Evidence Layers, Each with Its Own Denominator

Treat "employee AI adoption" as a measurement object organized in four layers, and state the question and denominator for each:

• Availability and authorization (can they use it): whether the company provides or permits relevant tools, whether employees know they can, and whether access actually works on their work devices. This layer is an organizational fact; whether answers cluster on "yes" or on "don't know" changes how later results must be read.
• Actual usage behavior (do they use it): a task-scenario checklist paired with frequency—for example, which of research, editing, drafting, data analysis, meeting notes, coding, or translation they used in the past 7 or 30 days. The denominator is all target employees.
• Object-specific trust (trust in what): split trust into "the tool output itself," "how the company manages and uses AI," and "AI's effect on my role and advancement," and rate the three objects on the same scale.
• Boundary concerns and avoidance (where they hold back): list concerns first (data leakage, accountability for errors, being replaced, looking lazy, lack of supervisor endorsement), then confirm with a behavior item whether employees actually avoided tasks because of those concerns.

The common failure is mixing layers: using a single "do you trust AI" item as the explanatory variable for usage. Only when trust is split by object and usage is recorded by task and frequency is the relationship interpretable.

Illustrative Worked Example: One Team, Three Adoption Rates

The following is illustrative data that shows how the measure changes the conclusion; it does not represent a survey by WeJot or any real organization. Suppose a 120-person content and operations team runs an internal study after the early-September GPT-6 release:

In the same illustrative team, 39 people (33%) said they were unsure whether the company allowed or had enabled the tool, 31 (26%) chose "worried about data leakage or non-compliance" among concerns, and 24 (20%) said they feared being held responsible for mistakes. Reading 61% as "employees already actively use AI" would hide that one-third live in a gray zone of wanting to use it without knowing the rules; reading the 26% compliance concern as simple conservatism would also fail to explain the separate 20% whose issue is accountability.

Readers can verify a recurring pattern: most errors come from mismatched numerators and denominators—computing a high-frequency rate among users only, or moving "heard of it" into the numerator of "used it." Fix the denominator of every percentage in the analysis plan before quoting proportions.

Split Trust by Object: Output, Company, and Job Are Three Things

A Pew survey fielded in October 2024 among employed U.S. adults drawn from a national probability panel found that 52% of workers were worried about the future impact of AI use in the workplace, and 32% believed it would reduce their job opportunities in the long run; only about one in ten used AI chatbots at work every day or a few times a week, while 55% said they rarely or never did. Workers are not simply anti-tool; worry and usage coexist and must be measured separately. Pew also found that AI users were more likely to call the tools very or extremely helpful for speed (40%) than for improving quality (29%), a reminder that trust in efficiency and trust in quality are not the same thing.

Trust items should therefore name their object explicitly. Three parallel agreement items work well:

• "I can use AI-generated results directly in tasks that require factual accuracy" (trust in tool output);
• "The rules on which data may be fed to AI and who verifies outputs are clear and trustworthy" (trust in governance);
• "Using AI will not put me at a disadvantage in performance reviews or advancement" (trust in job impact).

The profile across the three objects often differs within the same respondents. When governance rules are unclear, people may personally trust output quality yet still avoid use for fear of accountability. When the concern is job impact, more training will not help; what is needed is communication about performance and role rules. Compressing the three into one "AI trust index" deletes exactly the differentiation a decision-maker needs.

Concern Items Should Land on Actual Avoidance Behavior

Asking only "what worries you" collects vague anxiety. A more reliable design pairs two kinds of items: first let employees select concerns (multi-select), then ask "In the past month, did you ever decide not to use AI for a task because of the concerns above?" and have those who answer yes describe the specific task in an open-ended item. The result is a trackable behavioral gap that training or rule changes can be mapped to by task type.

When analyzing concerns, separate organizational causes from personal ones. Organizational causes include the tool not being enabled, data rules never stated, and no clear owner for output quality; personal causes include not knowing how, finding it unreliable, and fearing job loss. Only after organizational causes are ruled out can the remaining variation plausibly be attributed to individual attitudes. NIST's Generative AI Profile under the AI Risk Management Framework likewise emphasizes that generative AI can confidently produce incorrect content, and that organizations need arrangements for human–AI interaction, output verification, and human oversight—governance actions that belong in the survey as checkable facts about "what the company has done," not only as attitude items.

When an AI Interview Layer Adds Value

Questionnaires answer how many people, in which tasks, at what frequency, and with what concerns. When the need is to understand how people actually decide to use or not use AI, follow up with a small number of AI interviews among typical segments. Signals that justify interviews include concerns concentrated on hard-to-name accountability, one function whose usage diverges sharply from the company, or open-ended answers that repeatedly mention socially awkward reasons such as "afraid of looking lazy." An AI interview can reconstruct one real work scenario—the task, the moment they stopped, and the alternative they switched to—which a questionnaire cannot fully enumerate in advance. For method trade-offs, see combining surveys and AI interviews.

Whichever mode is chosen, define the population first: all employees, one business line, or only those with tools already enabled? The population definition bounds every conclusion and drives screening and quotas; see defining the target survey population.

Example Items You Can Reuse

1. Availability (single): "Can you currently use generative AI tools that the company provides or permits in your role?" Options: yes and I know the rules / yes but I am unsure of the rules / not sure whether it is allowed / clearly not allowed.
2. Behavior (multi plus frequency): "In the past 30 days, in which of the following work scenarios did you use generative AI? Select all, then mark the one you used most." Scenarios: research and information synthesis / editing or rewriting / drafting content / data analysis or spreadsheet work / writing code / meeting notes or summaries / translation / other.
3. Object-specific trust (scale 1–5): "Please rate the following three statements, where 1 means strongly disagree and 5 means strongly agree": the three statements about tool output, company rules, and job impact above.
4. Avoidance behavior (single): "In the past 30 days, did you ever choose not to use AI for a task because of concerns about data security, accountability for errors, or other reasons?" Options: often / occasionally / almost never / I have not used AI. Those answering often or occasionally receive an open item: "Please give one example of the task and your concern."

Question order shapes answers. Pew's questionnaire guidance shows that earlier questions provide context for later ones; so ask about behavior and experience first and put attitudes and worries later, so that employees are not primed to think "the company really values AI" before reporting their own behavior. Pretest the items on a small group to confirm that options such as "not sure whether it is allowed" are read as real choices rather than checked off as an easy default.

Running This Study in WeJot

WeJot can turn this design into an executable study: scenario checklists with branching logic route users into trust scales and non-users into access and concern items; sample targeting and quotas follow industry, job function, and whether tools are enabled; and open-ended and interview notes keep original text traceable during synthesis. Measuring real usage rather than claimed usage demands stricter respondent screening; see screener design.

The companion employee generative AI usage and trust survey template contains these dimensions and representative items and can be generated directly. Whatever the tooling, keeping availability, behavior, trust, and avoidance in separate layers prevents an access problem from being reported as an employee-attitude problem.