How we designed the studies and how we handle your data
Three linked studies designed to measure different facets of AI's impact on knowledge work. Each study is independently valuable. Together they produce a comprehensive AI Position Map that reveals how AI vulnerability, organizational readiness, and structural friction interact for a specific individual and team.
The studies are designed for knowledge workers in regulated industries, large enterprises, and any organization where autonomous AI meets high-stakes decisions and regulatory oversight. Each study takes 5-6 minutes to complete.
Five question types are used across the three studies, each chosen for a specific measurement purpose:
Tradeoff questions present a forced choice between two valid approaches. This eliminates "agree with everything" bias and reveals genuine preferences when both options have merit.
Scenario questions describe realistic workplace situations and ask what you would do. They measure actual behavioral patterns rather than stated beliefs or aspirations.
MaxDiff questions ask you to rank-order preferences from a set, identifying what matters most and least. This produces interval-scale data from ordinal choices, giving sharper differentiation than simple rating scales.
Likert scales measure agreement with specific statements. Used sparingly and always alongside behavioral questions to calibrate the accuracy of self-reported attitudes.
Confidence calibration compares your stated confidence about a topic with your measured exposure or readiness. The gap between confidence and reality reveals blind spots that are invisible in conventional surveys.
The AI Vulnerability Study asks each question under two conditions: normal work and when pressure increases (deadlines, resource constraints, organizational stress). The gap between these two responses reveals behavioral patterns that single-condition surveys miss entirely.
Under normal conditions, most knowledge workers describe their ideal operating mode. Under pressure, behavior shifts toward established habits, risk aversion, or compensating strategies. This gap is diagnostic: it reveals whether a person's AI resilience is structural (consistent across conditions) or situational (dependent on favorable circumstances).
Dimensional scoring rather than single-number reductions. Each study measures multiple independent dimensions. A person who scores high on role clarity but low on governance readiness is fundamentally different from someone with the opposite pattern, even if their total scores are identical.
Profiles are derived from dimensional patterns, not averages. Each profile represents a distinct combination of strengths and gaps. Close-second profiles capture nuance: the adjacent pattern that is nearly as strong as the primary, indicating where a person or organization could shift with targeted development.
Anonymous by default. Sessions are tracked by browser token, not by identity. No email is required to complete an assessment or view results. Email capture is optional and gates nothing.
Participants choose whether to contribute their responses to the research dataset or keep results for personal use only. Research opt-in is explicit. Personal-use responses are quarantined from aggregate analysis.
The research dataset contains no personally identifiable information. Session tokens, emails, and names are stripped before any aggregate analysis. The platform supports GDPR and CCPA withdrawal requests.
Aggregate benchmarks appear only when they are statistically meaningful:
| Benchmark | Minimum completions required |
|---|---|
| Overall profile distribution | 200 completions per study |
| Industry benchmark | 30 completions per industry per study |
| Role-level benchmark | 30 completions per role type per study |
| Organization size comparison | 50 completions per size band |
| Cross-study correlation | 100 participants with two or more studies completed |
| Longitudinal trends | 50 retakes per study |
When a threshold is not met, the corresponding benchmark section simply does not appear. There are no "coming soon" placeholders or empty sections.
All three studies are grounded in the diagnostic framework developed in Proxy.Me: Agentic AI Digital Apprentices (Jackson, 2026), which identifies four structural problems that current AI deployment strategies fail to address:
Structural inertia (Chapter 1): The coordination tax and activation energy required to initiate knowledge work. AI accelerates individual tasks but does not reduce the overhead of assembling information, permissions, and people before work can begin.
Knowledge at rest (Chapter 2): Expertise locked in individuals rather than available as an organizational asset. AI cannot surface knowledge that was never captured, structured, or made accessible.
Opaque decision-making (Chapter 3): Decisions treated as events rather than reusable, auditable assets. AI deployed in an environment of invisible decision logic accelerates choices that will be relitigated when the reasoning is lost.
The AI Productivity Trap (Chapter 4): The observation that AI reduces the cost of creation while increasing the cost of coherence. AI makes individuals faster while making organizations slower when deployed without structural coordination.
The theoretical framework generates specific, testable hypotheses about how different types of knowledge work interact with AI tools, which the three studies are designed to investigate.
The research program uses a continuous cross-sectional design with optional longitudinal follow-up. Studies remain open indefinitely, allowing the dataset to grow over time. Participants who provide an email address are invited to retake studies at six-month intervals, enabling within-person longitudinal analysis of profile stability and change.
The design is observational and correlational, not experimental. The studies measure self-reported work patterns, AI use, and friction experiences. They do not manipulate any variables. Causal claims are not made from the data; associations and typological patterns are the primary outputs.
This is a convenience sample recruited through multiple channels: the Corvair website, social media distribution (LinkedIn, Twitter/X), professional network sharing, academic partnerships, conference and event promotion, and book readership (references in Proxy.Me).
The sample is not a probability sample and should not be interpreted as representative of all knowledge workers globally. We report sample demographics alongside all findings so that readers can assess generalizability for themselves. We explicitly acknowledge this limitation in every published report.
Any individual who currently performs, is preparing to perform, or has recently performed knowledge work is eligible. The studies adapt their question framing based on work status to accommodate working professionals, students, career changers, those between roles, retirees, and academic researchers.
Each study uses four primary question types plus supplementary measures:
Tradeoff pairs (ipsative forced-choice). Two statements are presented. The respondent selects which better describes their experience on a five-point scale: Definitely A, Mostly A, Equal, Mostly B, Definitely B. Each pair maps to a specific measurement dimension. This format reduces social desirability bias because neither option is obviously "correct" and produces ipsative data suitable for within-person profiling and cluster analysis.
Scenario vignettes (situational judgment). A brief, realistic workplace situation is presented (2 to 3 sentences). Three questions follow: a diagnostic question ("What is the real problem here?"), a behavioral question ("What would you do?"), and a future-oriented question ("How does this evolve?"). Each scenario question is followed by a confidence probe: "How easy was this choice?" with three options: Easy, Hard to decide, Neither fits.
Best-worst scaling (MaxDiff). Four statements are presented. The respondent selects which is "most like my experience" and which is "least like my experience." This produces ratio-scale preference data with higher discriminative power than Likert ratings. Each MaxDiff block yields a complete preference ordering from a single interaction.
Rapid rating items (Likert scale). Single statements rated on a five-point scale from Strongly disagree to Strongly agree. These capture intensity and frequency measures that forced-choice items cannot. They serve as calibration and validation variables for the primary profile assignments.
Open text (optional). Each study includes one optional free-text prompt (e.g., "In one sentence, what is the biggest AI-related challenge you face at work?"). These are not scored but provide qualitative intelligence for research interpretation.
The first demographic question (Current work status) determines which variant of the assessment question text participants see. Ten work statuses are supported. Question text adaptations change surface language only. The underlying measurement dimensions, response scales, and scoring formulas remain identical across all work status variants.
Examples: Working professionals see "In your current role..." and "In a typical week..." Students see "In the career you are preparing for..." and "Based on your internship, project, or coursework experience..." Between roles see "In your most recent role..." and "Thinking about your last position..."
| Study | Tradeoff Pairs | Scenario Qs | MaxDiff | Likert | Confidence Probes | Total Scored | Est. Time |
|---|---|---|---|---|---|---|---|
| Study 1: AI Vulnerability | 12 | 6 | 1 | 5 | 6 | 25 | ~6 min |
| Study 2: AI Adoption | 10 | 6 | 2 | 4 | 6 | 23 | ~6 min |
| Study 3: Structural Friction | 10 | 6 | 2 | 4 | 6 | 23 | ~6 min |
Dimension Scoring. Tradeoff pair responses are coded on a scale from -2 (Definitely A) to +2 (Definitely B). Dimension scores are computed as the mean of each dimension's constituent tradeoff pairs, plus weighted contributions from relevant scenario vignettes (0.3 weight per scenario question). Scores are normalized to a 0-to-100 scale.
MaxDiff Utility Scores. Each item in a MaxDiff block receives a utility score: +1 for "most like me" selection, -1 for "least like me" selection, 0 for unselected. When sufficient data accumulates, hierarchical Bayes estimation will be used to compute individual-level utility scores with greater precision.
Likert Calibration. Likert items provide absolute intensity and frequency measures. They do not contribute directly to dimension scores but serve as calibration variables that validate profile assignments and flag potential inconsistencies.
Study 1 assigns one of 18 profiles across four categories (The Exposed, The Transitioning, The Durable, The Paradoxes) based on the four-dimension score pattern. A composite Vulnerability Index (0 to 100) is computed as a weighted sum of dimension scores.
Study 2 assigns one of 30 profiles organized across six categories (Power Users, Team Players, Frustrated, Specialized, Cautious, Paradoxes), using a decision-tree model that classifies first on three primary axes, then refines using scenario responses, Likert modifiers, MaxDiff barrier preferences, and the future orientation modifier.
Study 3 assigns one of 15 profiles across four categories (Single-Friction Dominant, Dual-Friction Patterns, System-Wide, Paradoxes) based on the dominant friction type or combination. Paradox profiles are identified when tradeoff-pair scores and Likert absolute measures diverge.
Three behavioral signals are collected invisibly during participation. These are disclosed in the consent notice.
Every item records response time in milliseconds. Under 4 seconds indicates rapid/automatic response. 4 to 15 seconds indicates normal deliberation. 15 to 30 seconds indicates extended deliberation on a personally salient item. Over 30 seconds may indicate confusion, distraction, or deep conflict.
Six confidence probes per study (one after each scenario question) capture a three-point response: Easy (the scenario is familiar and the respondent's orientation is settled), Hard to decide (the respondent is genuinely torn, indicating a tension point and high engagement), or Neither fits (the framework does not capture the respondent's reality, providing valuable feedback for instrument refinement).
When the interface provides a back button, answer revisions are tracked. A revised answer indicates that a later question prompted the respondent to reconsider an earlier choice. Revision patterns are among the richest behavioral signals in assessment research and can indicate reflection depth, item interaction effects, and profile instability.
Immediate Feedback. Every participant receives their full profile results on screen the moment they complete a study. This includes profile name and category, visual representation, written interpretation, personalized recommendations, MaxDiff preference summaries, blind spot alerts (where applicable), and benchmark comparison (once sufficient data exists). No email address is required to see full results.
Team Engagement. After viewing results, participants can share a link with colleagues. When three or more participants from the same organization complete studies, a team comparison view unlocks showing the distribution of profiles across the group.
Optional Email Report. Participants who choose to provide an email address receive an extended report including detailed dimension breakdowns, benchmark comparisons, book chapter references, and an expanded action plan. Email addresses are stored separately from study responses.
Longitudinal Re-Participation. Participants who provide an email address are invited to retake studies at six-month intervals. Profile comparison over time is delivered as a personalized change report. Longitudinal matching uses an email hash for privacy-safe identification.
Reliability. Internal consistency (Cronbach's alpha) will be computed per dimension after the pilot phase. Target: alpha greater than 0.70. Test-retest reliability will be assessed using longitudinal re-participation data.
Validity. Content validity: all items are grounded in the theoretical framework of Proxy.Me. Construct validity will be assessed through confirmatory factor analysis. Criterion validity will be assessed by examining whether profiles predict self-reported outcomes such as AI tool adoption, career satisfaction, and organizational friction experiences.
Multi-Method Triangulation. The combination of ipsative (tradeoff pairs), ratio-scale (MaxDiff), normative (Likert), and behavioral (scenario vignette) measurement methods enables convergent and discriminant validation within each study. Where methods agree on a profile assignment, confidence is high. Where they diverge, the respondent's situation is genuinely complex, and both the primary and "close second" profiles capture this complexity.
| Analysis | Minimum Sample | Purpose |
|---|---|---|
| Exploratory factor analysis | 100 per study | Validate theoretical dimension structure |
| Internal consistency | 100 per study | Assess measurement reliability |
| Cluster analysis | 500 per study | Identify natural respondent groupings |
| Confirmatory factor analysis | 500 per study | Test theoretical model fit |
| Cross-tabulation (Study 2 x Study 3) | 300 with both | Test whether friction predicts adoption patterns |
| Hierarchical Bayes MaxDiff | 500 per study | Compute individual-level preference utilities |
| Regression modeling | 500 per study | Test demographic and friction predictors of adoption |
| Industry benchmarking | 50+ per industry | Produce industry-level profile distributions |
| Longitudinal tracking | 100 retakers | Assess profile stability and change over time |
| Response time analysis | 500 per study | Correlate deliberation patterns with profile types |
Quarterly benchmark reports will be published starting when any single study reaches 500 participants. Reports will include aggregate profile distributions segmented by industry, role type, geography, organization size, and work status.
Academic papers will be developed where the data supports it. Planned topics include: the relationship between structural friction and AI adoption success; a typology of AI adoption patterns among knowledge workers; the distribution of AI vulnerability across knowledge work domains; and the psychometric properties of best-worst scaling for AI readiness measurement.
This research program has several known limitations that are acknowledged in all published findings:
Convenience sample. Participants self-select into the studies. The sample is not representative of the global knowledge worker population and likely over-represents individuals who are actively thinking about AI and their careers.
Self-report bias. All measurements rely on participant self-report. The multi-method design and ipsative forced-choice format mitigate but do not eliminate this bias.
Cross-sectional limitations. The primary design is cross-sectional. Causal relationships cannot be inferred from correlational data. Longitudinal follow-up data will partially address this limitation.
Cultural and linguistic context. Studies are currently available in English only. Question wording references workplace norms that may not translate uniformly across cultures.
Theoretical grounding. The instruments are grounded in a specific theoretical framework (Proxy.Me). The studies test and operationalize this framework; they do not test it against competing frameworks.
Profile assignment precision. With current item counts, some profile assignments will be uncertain, particularly for respondents near dimension midpoints. The "close second" profile feature partially addresses this by acknowledging ambiguity rather than forcing a single classification.
Corvair welcomes collaboration with academic researchers. Potential collaboration models include:
Instrument use. Researchers may use the study instruments in their own research with appropriate citation. Contact research@corvair.ai for instrument documentation and scoring specifications.
Data access. Anonymized, aggregate datasets can be made available to academic researchers under data-sharing agreements. Individual-level data is available only under agreements that include IRB approval and privacy protections.
Co-authorship. Researchers with relevant expertise in psychometrics, organizational behavior, AI adoption, or related fields are invited to propose co-authored analyses.
Replication and critique. We encourage independent analysis and critique of our methods and findings. The instruments, scoring logic, and sample characteristics are documented in sufficient detail for external review.
Contact: research@corvair.ai
When referencing this research program in published work, please use:
Corvair Pte Ltd. (2026). Corvair AI and Knowledge Work Research Program: Methodology. Retrieved from https://corvair.ai/research/methodology/
When referencing the theoretical framework, please cite:
Jackson, C. (2026). Proxy.Me: Agentic AI Digital Apprentices.