Evidence-based management

Evidence-based management asks whether organizational claims are supported by the best available, relevant evidence — and whether the evidence is strong enough for the decision being made.

2014-briner-employee-engagement-evidence-based applies that stance to employee engagement. Briner argues that the most important practical questions are causal: whether engagement increases lead to performance increases, and whether engagement interventions increase engagement and then performance. Cross-sectional correlations cannot answer those questions, especially when the construct itself is unclear and measures overlap with older attitudes.

This is directly relevant to AI workforce transformation. Claims about AI adoption, productivity, well-being, skill atrophy, or human-AI collaboration need evidence matched to the claim: descriptive usage data can show uptake, but not causal impact; correlations can suggest associations, but not prove intervention effects.

algorithmic-assessment is a useful stress test for evidence-based management because several kinds of evidence travel under the same “AI assessment” label. 2016-chamorro-premuzic-new-talent-signals cautions that novelty, speed, and vendor enthusiasm are not evidence; new talent signals still need construct, validity, and utility standards. 2023-landers-machine-learning-psychometric-assessment is a strong example of matching method to claim: simulation evidence can test design and scoring contingencies, but it should not be read as observed organizational adoption or applicant outcome evidence. 2023-nyc-automated-employment-decision-tools-faq shows a different evidence category: compliance can require a bias audit even though the audit does not by itself establish construct validity, job relatedness, or practical utility.

2026-tomei-what-jobs-can-ai-learn shows why evidence-based AI workforce planning needs to match the metric to the decision. A forward-looking RL Feasibility Index may be useful for scanning where future automation pressure could emerge, but it should not be read as observed displacement or as a complete labor-market forecast. Its job-posting analysis is suggestive; the index itself is a diagnostic for task learnability.

2025-mazeika-remote-labor-index is useful because it moves the evidence unit from model capability to economically accepted work output. The RLI automation rate asks whether an AI deliverable would be accepted by a reasonable client compared with a human gold-standard deliverable, which is closer to a management claim about substituting for real work than scores on narrower benchmark tasks.

2022-strah-diversity-issues-job-analysis is a reminder that evidence quality depends on the upstream definition of the thing being evidenced. A selection system, performance tool, or compensation structure can be evidence-based in a narrow procedural sense while inheriting a job analysis that failed to capture diverse work experiences. For management claims about fairness, legal defensibility, or inclusion, the evidence chain has to include how the job domain was defined, who had voice in that definition, and whether aggregation erased meaningful subgroup differences.

2008-icf-competency-modeling-job-analysis adds a method-fit warning: job analysis, strategic job analysis, cognitive task analysis, personality-oriented job analysis, and competency modeling answer different evidence needs. Evidence-based HR should not ask which method is best in general; it should ask whether the method fits the decision, specificity level, legal risk, time horizon, and unit of analysis.

2009-sanchez-levine-competency-modeling-job-analysis adds a use-case warning: the same competency model can be evidence-relevant for strategy communication and weak evidence for selection or assessment. Evidence-based management has to ask what decision the model is supporting before deciding what rigor, specificity, and validation evidence are required.

1997-goh-benchmarking-learning-capability adds a classic survey-feedback case. The OLS can help managers diagnose learning-enabling conditions and target organizational-development interventions, but the evidence claim should stay bounded: reliability and cross-sectional differentiation do not by themselves show that learning capability causes performance or that OLS-guided interventions improve outcomes. That distinction is directly relevant to AI transformation, where organizations may want diagnostics for learning from AI experimentation before they have outcome evidence.

Connections

• Relates to construct-validity because evidence quality depends on whether the measured construct is coherent enough to interpret.
• Relates to employee-engagement as a practitioner domain where advocacy often outran evidence.
• Relates to ai-adoption because adoption interventions will need designs capable of separating access, discovery, workflow redesign, and performance effects.
• Relates to automation-and-substitution because exposure measures can inform planning, but only if leaders keep exposure, adoption, productivity, and employment effects analytically separate.
• Relates to algorithmic-assessment because the evidence ladder must distinguish validity evidence, adverse-impact evidence, compliance evidence, implementation evidence, and decision utility.
• Relates to job-analysis because many HR evidence claims begin with a documented job domain; that documentation is itself evidence that can be inclusive, incomplete, or biased.
• Relates to competency-modeling because competency models often become evidence artifacts for HR decisions, even when their intended use and validation standard are underspecified.
• Relates to organizational-learning-capability because learning surveys can guide intervention only if leaders distinguish diagnostic evidence from causal evidence about changed routines and performance.

Contradictions & open questions

• Evidence-based management should discipline claims without becoming paralysis. The wiki should track where imperfect evidence is still decision-relevant and where claims are too weak to operationalize.
• AI transformation evidence will often be fast-moving and proprietary; the public-only boundary means the wiki may underrepresent some internal organizational evidence even when such evidence exists.
• 2025-landers-ethical-use-ai-iop adds a practical tension: by the time evidence on a new AI system is peer-reviewed, the technology may already be obsolete, yet organizations still make worker-facing decisions. Evidence-based management should not become an excuse for either paralysis or reckless adoption; it should specify what evidence is available, what remains uncertain, which ethical principles govern the gap, what scope limits are warranted, and what monitoring, safeguards, or rollback mechanisms are required before scale.
• For occupational exposure claims, the evidence ladder should distinguish annotation-based indices, observed tool deployment, productivity changes, job-posting shifts, and actual employment or wage effects. 2026-tomei-what-jobs-can-ai-learn mostly sits at the index-plus-suggestive-validation stage.
• 2025-mazeika-remote-labor-index strengthens the evidence base for autonomous remote-work completion, but it still should not be read as evidence about assisted productivity, within-job augmentation, labor demand, wages, or adoption. Those are different outcomes requiring different designs.
• For algorithmic hiring claims, the open question is often not whether evidence exists, but what kind: predictive validity, construct validity, bias-audit metrics, applicant reactions, legal compliance, implementation fidelity, and business utility answer different questions.
• 2022-strah-diversity-issues-job-analysis raises an evidence paradox: job analysis is often used as the fairness foundation for HR systems, but the inclusiveness of the job analysis may itself be an untested assumption rather than established evidence.
• 2008-icf-competency-modeling-job-analysis suggests that strategic flexibility and evidentiary rigor can pull in different directions: broad competency models travel across jobs, but job-specific decisions often require more specific and reliable evidence.
• 2009-sanchez-levine-competency-modeling-job-analysis offers a bridge but not a full solution: use competency modeling to signal strategic behavior, then use job-analysis methods to derive behavioral indicators in specific jobs or job families. The open question is how much evidence is enough when the competency is partly strategic, partly behavioral, and partly measurement-facing.
• 1997-goh-benchmarking-learning-capability raises the same evidence ladder problem for learning organizations: a reliable climate/capability diagnostic may be useful for survey feedback, but claims about AI integration or organizational performance still require longitudinal intervention and outcome evidence.