I want to examine whether psychological safety mediates the relationship between AI adoption and employee job satisfaction, and whether this relationship differs across occupational groups. What research design, measures, sampling strategy, and statistical methods would be appropriate?

Recommendation. Use a three-wave, multi-occupation survey analyzed with latent-variable structural equation modeling (SEM). Measure AI adoption at Time 1, psychological safety at Time 2, and job satisfaction at Time 3, while testing occupational group as a moderator through multigroup SEM or moderated mediation. This design establishes temporal ordering more convincingly than a single cross-sectional survey, although it still does not establish causality as strongly as a quasi-experimental or panel design with stronger controls.

1. Research design

A suitable model is:

$$\begin{aligned} &\text{AI adoption} \\ &\quad \rightarrow \text{psychological safety} \\ &\quad \rightarrow \text{job satisfaction} \end{aligned}$$

with occupational group moderating either the path from AI adoption to psychological safety, the path from psychological safety to job satisfaction, or both. The central indirect effect for group $g$ is:

$$ab_g = a_g b_g$$

where $a_g$ is the effect of AI adoption on psychological safety and $b_g$ is the effect of psychological safety on job satisfaction for that occupational group.

A three-wave time-lagged design is preferable

• Time 1: AI adoption, occupational group, demographic and job controls.
• Time 2: psychological safety, ideally four to eight weeks later.
• Time 3: job satisfaction, ideally another four to eight weeks later.

The intervals need to be long enough to reduce same-source consistency effects but short enough that major organizational changes do not intervene. If resources permit, collect job satisfaction at both Time 1 and Time 3 so that the model predicts change in satisfaction, controlling for baseline satisfaction.

The theoretical logic should specify how AI adoption affects psychological safety. Adoption may increase safety when employees experience AI as supportive, competence-enhancing, and compatible with their work. It may reduce safety when employees interpret it as surveillance, a threat to employment, or evidence that mistakes will be exposed. This ambivalence is consistent with evidence that AI adoption can produce both self-efficacy benefits and anxiety-related costs, depending partly on whether employees view AI as helping or replacing humans [1]. It is also consistent with evidence that AI adoption may weaken trust and the positive effect of psychological contracts on engagement [2].

Do not treat “AI adoption” as a purely technological variable. Distinguish at least three dimensions: use intensity—how frequently employees use AI; implementation exposure—the extent to which the organization has integrated AI into work processes; and perceived consequences—whether employees perceive AI as supportive, threatening, autonomy-enhancing, or surveillance-oriented. A model based only on frequency of use may confound voluntary experimentation with compulsory organizational implementation.

2. Measures

2.1 AI adoption

Use a multidimensional measure combining behavioral and perceptual indicators.

Behavioral adoption should include frequency, duration, task coverage, and dependence on AI tools. Example items can ask how often employees use AI for drafting, analysis, decision support, communication, scheduling, or customer-facing work. If objective usage logs are available, use them to validate self-reports.

Organizational adoption should measure the extent to which AI is embedded in the employee’s work environment: whether AI is formally deployed, whether supervisors expect its use, whether work processes have been redesigned around it, and whether AI affects performance evaluation or staffing decisions.

Perceived adoption consequences should be measured separately rather than folded into adoption itself. Relevant dimensions include perceived AI support, perceived replacement threat, AI-related job insecurity, autonomy, and training adequacy. This separation is important because two employees may report identical AI use while experiencing very different psychological conditions. Studies of AI adoption have found that job insecurity can be the mechanism linking adoption to depression, while corporate social responsibility attenuates that pathway [3]. Training is similarly important because technostress is associated with poorer adoption and well-being, whereas training can mitigate these effects [4].

Use a validated scale where possible, adapt items to the occupational setting, and report the adaptation process. If the study includes generative AI specifically, identify the tools and tasks rather than using “AI” as an undefined umbrella term.

2.2 Psychological safety

Measure psychological safety as a shared perception that employees can ask questions, admit mistakes, express concerns, seek help, and challenge practices without interpersonal punishment or humiliation. The classic short psychological-safety scale associated with Edmondson is appropriate, provided that its wording is adapted carefully to AI-related work.

You may add AI-specific items, such as whether employees can question an AI-generated recommendation, report an AI error, or disclose difficulty using an AI system without damaging their reputation. However, keep the general psychological-safety scale separate from the AI-specific supplement. Otherwise, the mediator may become conceptually indistinguishable from the predictor.

Psychological safety should be measured at the level implied by the theory. If the hypothesis concerns the immediate work team, ask respondents to evaluate their team climate. If it concerns the organization, use organization-referent wording. Combining team and organizational referents in one score would create interpretive ambiguity. Evidence from hospital staff indicates that psychological safety is associated with affective commitment through job satisfaction and burnout, with perceived organizational support weakening the harmful burnout pathway [5]. This supports treating psychological safety as a relational work condition rather than merely an individual personality perception.

2.3 Job satisfaction

Use a multidimensional or global job-satisfaction instrument with established reliability and validity, such as the Job Satisfaction Survey, Minnesota Satisfaction Questionnaire, or a short global measure if questionnaire length is a serious constraint. Decide in advance whether job satisfaction means:

• overall satisfaction;
• intrinsic satisfaction, such as meaningfulness and achievement;
• extrinsic satisfaction, such as pay and promotion; or
• satisfaction with AI-affected aspects of the job.

A global measure is suitable for the primary outcome, but an intrinsic–extrinsic distinction may be theoretically useful because AI can enrich task content while simultaneously threatening autonomy or employment security. Occupational conditions have a direct relationship with job satisfaction, and the relevant conditions can differ across workers [6]. This makes it important not to interpret a group difference in satisfaction as an AI effect unless baseline job characteristics are controlled.

2.4 Controls and boundary conditions

Include controls that are theoretically justified rather than a large undirected set. At minimum, consider age, gender, education, organizational tenure, job tenure, managerial status, employment security, workload, prior AI experience, digital competence, remote-work status, organization, and baseline job satisfaction.

Potentially important covariates include perceived organizational support, supervisor support, AI training, data-governance quality, and whether AI use is voluntary. Do not control for variables that are consequences of AI adoption if doing so would block the mechanism being estimated. For example, if job insecurity is theorized as a pathway through which adoption reduces psychological safety, controlling for it in the primary mediation model may remove part of the effect of interest.

Use the same response scale where practical, but do not rely on scale uniformity as a remedy for common-method bias. Procedural separation across waves, confidentiality assurances, careful item ordering, and objective or supervisor-reported indicators are more important.

3. Sampling strategy

Define occupational groups before data collection using a substantive classification rather than post hoc statistical convenience. Suitable comparisons might include:

• knowledge-intensive professionals;
• administrative and clerical employees;
• frontline service employees;
• technical or production workers;
• healthcare professionals; and
• managers or supervisors.

The final groups should differ in AI exposure, task interdependence, discretion, and employment consequences. Avoid creating groups that differ only by job title but perform similar work.

Use stratified, multisite sampling. First, recruit organizations from sectors with meaningful AI implementation. Within each organization, sample employees from each target occupational group. If employees are nested within teams and organizations, retain organization and team identifiers. A purely convenience sample of AI users from one occupation cannot test occupational moderation convincingly.

For a basic two- or three-group comparison, a practical target is approximately 150–200 usable respondents per occupational group after attrition. A more defensible number should be determined through a Monte Carlo power analysis because indirect effects are often smaller and less precisely estimated than direct effects. Simulate the expected path coefficients, group proportions, factor loadings, intraclass correlations, and attrition across waves. If the study includes five groups, a total sample of 750–1,000 at Time 1 may be necessary to retain adequate cases in the smallest group at Time 3.

If organizational clustering is substantial, inflate the sample for the design effect:

$$\text{DE} = 1 + (m-1)\rho$$

where $m$ is the average number of employees per team or organization and $\rho$ is the intraclass correlation. Recruitment should also anticipate longitudinal attrition; oversample at Time 1 and use retention reminders and matched anonymous identifiers.

A multilevel design is particularly appropriate if psychological safety is conceptualized as a team climate. In that case, sample enough teams—not merely enough individuals—to estimate between-team variation. A few large organizations with many respondents do not substitute for a sufficient number of independent teams or organizations.

4. Statistical analysis

4.1 Preliminary analysis

Begin with data screening, missing-data analysis, attrition comparisons, descriptive statistics, reliability estimates, and correlations. Compare respondents retained at Time 3 with those lost to follow-up on observable Time 1 variables. Use full-information maximum likelihood or multiple imputation rather than listwise deletion when the missingness assumptions are defensible.

Test whether the data are clustered by team or organization using intraclass correlations and design effects. If clustering is meaningful, use multilevel SEM or cluster-robust standard errors.

4.2 Measurement model

Estimate a confirmatory factor analysis before testing structural paths. The measurement model should distinguish AI adoption, psychological safety, and job satisfaction, as well as any theoretically distinct AI-consequence dimensions such as job insecurity or AI anxiety.

Assess factor loadings, internal consistency, convergent validity, discriminant validity, and model fit using multiple indices rather than a single cutoff. Because occupational groups are being compared, test measurement invariance sequentially:

1. Configural invariance: the same factor structure across groups.
2. Metric invariance: equivalent factor loadings.
3. Scalar invariance: equivalent intercepts.
4. Residual invariance, if required for especially strict comparisons.

At least metric invariance is needed to compare relationships across groups, and scalar invariance is generally needed to compare latent means. If full invariance fails, test partial invariance and report which items differ. Without this step, an apparent occupational difference may reflect different interpretations of psychological safety or job satisfaction rather than a substantive difference.

4.3 Mediation

Estimate the indirect effect of AI adoption on job satisfaction through psychological safety using bootstrapped confidence intervals, preferably with at least 5,000 bootstrap resamples. Do not infer mediation merely because the total effect is significant; an indirect effect can exist even when the total effect is weak or nonsignificant because positive and negative pathways may offset each other.

The primary model should include

• the direct path from AI adoption to job satisfaction;
• the path from AI adoption to psychological safety;
• the path from psychological safety to job satisfaction;
• baseline job satisfaction, if available;
• theoretically justified covariates; and
• organization or team clustering adjustments.

A cross-lagged panel model is preferable if repeated measures are available. A stronger design would measure all three constructs at multiple waves and estimate autoregressive and cross-lagged paths. This would help distinguish whether AI adoption precedes changes in psychological safety and whether psychological safety precedes later satisfaction, rather than merely correlating with them.

4.4 Occupational differences

Use multigroup SEM when occupational groups are categorical and substantively defined. Estimate the mediation model separately in each group and compare:

• the $a$ path from AI adoption to psychological safety;
• the $b$ path from psychological safety to satisfaction;
• the direct effect;
• the total effect; and
• the indirect effect $ab$.

Test equality constraints formally rather than comparing whether one group’s coefficient is significant while another’s is not. The correct test is whether the coefficients or indirect effects differ significantly across groups.

If occupation is continuous or represented by several correlated attributes—such as task complexity, discretion, or degree of customer contact—use latent moderated structural equations or an interaction model rather than forcing arbitrary categories. If employees are nested in organizations and groups are distributed unevenly across organizations, use multilevel moderated mediation or include organization fixed effects where appropriate.

A useful extension is to test whether occupational differences operate through AI exposure rather than through occupation itself. For example, occupational group may moderate the relationship because groups differ in autonomy, replacement threat, task complexity, or access to training. Recent evidence from healthcare professionals suggests that job complexity can weaken the relationship between AI use and satisfaction of autonomy and competence, while the relationship with relatedness may remain unaffected [7]. This provides a plausible mechanism for occupational heterogeneity, but it should be tested rather than assumed.

5. Validity, robustness, and interpretation

Common-method bias is a serious risk because all principal variables may come from employee self-reports. Temporal separation reduces but does not eliminate that problem. Strengthen the design with objective AI-usage indicators, supervisor or team-level measures of AI implementation, and independent indicators of job satisfaction or turnover intention where feasible.

Because the literature reports both beneficial and harmful effects, specify competing mechanisms in advance. AI may improve satisfaction through autonomy, competence, and task support, while reducing it through anxiety, insecurity, impaired organizational support, or alienational psychological contracts. Evidence on generative AI task support shows that employees may perceive ChatGPT support as providing less organizational support and opportunity to perform, while increasing job insecurity and reducing satisfaction [8]. Conversely, AI literacy has been associated with greater fulfillment of autonomy, competence, and relatedness among university faculty, with these needs contributing to work–life balance and satisfaction [9]. Your model should therefore avoid assuming that the coefficient from AI adoption to psychological safety must be positive.

Finally, describe the study as testing temporal mediation or evidence consistent with mediation unless the design includes stronger causal identification. A three-wave survey improves temporal ordering, but unmeasured organizational change, self-selection into AI use, reverse causality, and occupational differences in organizational context can remain. The strongest feasible extension would combine the survey with a pre/post implementation design, matched comparison organizations, or an instrumental-variable or difference-in-differences strategy.