The trust deficit
Three independent studies converge: workplace wellbeing interventions do not work. What produces measurable outcomes is autonomy, psychological safety, and non-toxic culture — structural factors that require changing the work, not the worker.
The evidence against individual interventions is definitive
In January 2024, Dr. William Fleming of the Oxford Wellbeing Research Centre published the largest matched-pair study of workplace wellbeing interventions ever conducted. The dataset covered 46,336 workers across 233 UK organisations and eleven intervention types — mindfulness, resilience training, stress management, relaxation classes, wellbeing apps, time management, financial wellbeing, sleep improvement, and more.1Fleming, 2024, Industrial Relations Journal. 46,336 workers across 233 UK organisations (Britain's Healthiest Workplace surveys, 2017–2018). Matched pairs at the same company — one using a wellness programme, one not — across multiple subjective wellbeing indicators. The single exception: charity and volunteering programmes showed a positive association.
None of them showed positive effects on worker wellbeing.
The two most rigorous RCTs in the field converge on the same conclusion. The BJ's Wholesale Club/Harvard-Chicago trial (Song & Baicker, 2019, published in JAMA) — a cluster-randomised clinical trial across 160 worksites — found no statistically significant effect on employee tenure, absenteeism, clinical measures, or healthcare spending. The Illinois Workplace Wellness Study (Jones, Molitor & Reif, 2019, published in the Quarterly Journal of Economics) produced null estimates and its 95% confidence intervals ruled out 84% of the positive effects reported in 112 prior studies — demonstrating that the entire prior evidence base was likely driven by selection bias.2Song & Baicker, "Effect of a Workplace Wellness Program on Employee Health and Economic Outcomes," JAMA, 2019. Jones, Molitor & Reif, "What Do Workplace Wellness Programs Do?," Quarterly Journal of Economics, 2019.
The scale of misallocation: Companies with over 20,000 employees spend an average of $11 million per year on wellbeing programmes (Deloitte). The US workplace wellness industry covers over 50 million workers. EAP utilisation rates average under 10%, with some programmes as low as 2–5%. 68% of workers report not using these resources because accessing them is "too time-consuming, confusing, or cumbersome."
Structural factors drive wellbeing — and they are all about job design
The evidence on what actually improves employee wellbeing converges from multiple independent research traditions on the same conclusion: it is the design of work itself, not interventions layered on top of bad work.
Autonomy is the most consistently validated driver. The Hackman & Oldham Job Characteristics Model (validated across 200+ empirical articles) identifies five core characteristics — skill variety, task identity, task significance, autonomy, and feedback — with autonomy as a multiplicative factor: if autonomy is zero, total motivating potential is zero regardless of other characteristics.3The Karasek demand-control model, supported by three recent meta-analyses, confirms that high demands combined with low decision latitude predict psychological strain, while high demands combined with high control predict engagement and learning.
Psychological safety emerged as the number-one predictor of team effectiveness in Google's Project Aristotle study (180+ teams, 250+ variables, two years). The finding: who is on a team matters far less than how the team works together. Only 47% of employees worldwide describe their workplaces as psychologically safe.
Manager quality accounts for at least 70% of the variance in employee engagement scores across business units (Gallup meta-analysis). Only 1 in 10 people possess high natural talent to manage, yet organisations routinely promote based on tenure rather than aptitude.
The five toxic attributes — disrespectful, non-inclusive, unethical, cutthroat, and abusive — affect approximately 1 in 10 US workers and cost US employers nearly $50 billion per year in turnover alone. The most powerful short-term retention lever identified was not pay increases but lateral career opportunities (2.5× more powerful than compensation).4MIT Sloan, analysis of 34 million employee profiles and 1.4 million Glassdoor reviews. Gallup's 2025 State of the Global Workplace report: only 21% of global workers engaged — at an estimated cost of $438 billion in lost productivity. Reaching best-practice engagement (70%) would add $9.6 trillion to the global economy.
Trust is eroding where it matters most
The 2025 Edelman Trust Barometer shows "my employer" remains the most trusted institution at 75%, but this is a 3-point decline from the prior year and is conditional. Associates (entry-level and non-managers) are 32 points less trusting of institutions than executives. Job level is a stronger predictor of trust than income.5PwC 2024 Trust Survey: 86% of executives think employee trust is high, but only 67% of employees agree — an 18-point gap. 69% of employees believe businesses should disclose AI governance frameworks, but only 33% of leaders do so.
On AI specifically, trust in AI companies has declined from 62% to 54% over five years. In the US, trust in AI is just 32% (versus 72% in China). 59% of global employees fear job displacement due to automation. 80% of employees say it is important for their CEO to speak publicly about ethical use of technology.
The Alienational contract: Braganza et al. (2021, Journal of Business Research, N=232) found that AI adoption significantly weakens the positive effect of psychological contracts on job engagement, and proposed a third contract type — the "Alienational" contract — where AI detaches workers from decent work regardless of whether the underlying contract is transactional or relational.
Surveillance accelerates the damage
Employee monitoring is a particularly destructive force. The monitoring software market reached $587 million in 2024 (projected $1.4 billion by 2031). 78% of companies now use some form of monitoring.
| Metric | Monitored employees | Unmonitored peers |
|---|---|---|
| Plan to leave within a year | 42% | 23% |
| Say monitoring improves productivity | 28% | — |
| Say digital tracking damages trust | 59% | — |
| Lost trust in their organisation | 75% | — |
Arizona State University researchers found excessive monitoring actually reduces productivity. Amazon France was fined €32 million by CNIL in January 2024 for an "excessively intrusive" scanner surveillance system.6Gartner reports monitoring adoption at approximately 70% by 2025. The outcomes are consistently self-defeating: the $1.4 billion monitoring industry is generating negative ROI by its own metrics.
Mutual-investment employment models
Companies that explicitly structure the employment relationship around long-term capability development demonstrate dramatically better outcomes.
| Company | Model | Key metric | Industry average |
|---|---|---|---|
| Costco | $30+/hour, employer health insurance for 90%+ | 94% retention after year one | 30–40% retention |
| Patagonia | Values-driven, long-term investment | 6–7% turnover | 60%+ turnover |
| Mondragon | 100 cooperatives, 81,000+ workers | 97% cooperative survival rate | — |
At Costco, 85% of newly promoted warehouse managers started as hourly workers. Mondragon, during economic downturns, primarily adjusts remuneration rather than employment — the opposite of conventional firms.7US ESOPs cover 14.9 million employees across 6,358 companies. ESOP participants aged 28–34 have 53% longer median job tenure. A 2025 update found median tenure three years longer overall. Employee owners report higher annual earnings and give higher ratings on management-employee relations and trustworthiness.
The John Lewis model: democratic governance at scale
The John Lewis Partnership (80,000 "Partners") demonstrates democratic governance at scale: 5 of 15 Board members are elected by workers, and the Partnership Council can dismiss the Chairman by vote. When the company redefined its purpose, approximately 20% of Partners (~13,800 people) were involved in consultation. The UK employee ownership sector is growing at approximately 10% per annum.
The regulatory divergence
Three interlocking EU instruments create the most advanced regulatory architecture for AI in employment anywhere in the world.8EU AI Act (Regulation 2024/1689, entered force August 1, 2024). EU Platform Work Directive (Directive 2024/2831, entered force December 1, 2024, transposition by December 2, 2026). GDPR Article 22, expanded by the CJEU SCHUFA ruling (2023).
The EU AI Act classifies all AI systems used in employment — recruitment, performance monitoring, promotion, termination, task allocation — as high-risk. Emotion recognition in workplaces is banned outright (effective February 2, 2025). Penalties reach €35 million or 7% of global turnover for prohibited practices.
The Platform Work Directive creates a rebuttable employment presumption for platform workers, prohibits processing emotional or psychological data, and bans account suspension based solely on algorithms. It affects an estimated 43 million platform workers.
The SCHUFA ruling and its implications
The landmark CJEU SCHUFA ruling (2023) held that AI-generated scores can constitute decisions under GDPR Article 22 even when a human formally makes the final decision based on the score. This means algorithmic scoring of candidates and workers triggers data protection obligations even with nominal human review.
The European Parliament's Employment Committee published a draft report in June 2025 recommending a dedicated Directive on Algorithmic Management in the workplace.
Where enforcement fails
Null compliance: A Cornell/CAT Lab study of New York City's Local Law 144 found that of 391 employers surveyed, only 18 published bias audit reports and only 13 posted transparency notices. The enforcement agency received 2 complaints in two years. The Illinois AI Video Interview Act has produced zero compliance data.
Spain's Rider Law (2021) required algorithmic transparency from all employers. In practice, enforcement has been described as "largely ineffective." Glovo resisted compliance for years and only began formally employing riders under criminal prosecution of its CEO.9The US revoked its most comprehensive federal AI order (Biden EO 14110) on January 20, 2025. The UK has no AI-specific legislation. The regulatory weight has shifted decisively to the EU, creating compliance asymmetries for multinational employers.
Who captures AI productivity gains
The most consequential unanswered question — acknowledged explicitly by Brookings — is: to what extent will workers benefit from AI-supported productivity gains, and via what mechanisms?
PwC's 2025 Global AI Jobs Barometer finds AI-skilled workers command a 56% wage premium, up from 25% the prior year. But research across 238 regions in 21 European countries shows AI innovation correlated with declining labour income shares — gains flowing to capital rather than labour.10Humlum & Vestergaard, NBER (September 2025): linking ChatGPT adoption to administrative records across 11 AI-exposed occupations in Denmark, found essentially zero aggregate effects on earnings or hours through 2024 — consistent with a "productivity J-curve." OECD SME survey: 83% of generative AI adopters reported no change in staffing levels.
The window: The disruption is real but its labour market manifestation lags adoption by years. This gives organisations a window to redesign — if they act on the structural evidence rather than defaulting to the reskilling playbook.
References
- Fleming, W., "Employee Wellbeing Outcomes from Individual-Level Interventions," Industrial Relations Journal, January 2024. N=46,336, 233 organisations.
- Song, Z. & Baicker, K., "Effect of a Workplace Wellness Program on Employee Health and Economic Outcomes," JAMA, 2019. 160 worksites.
- Jones, D., Molitor, D. & Reif, J., "What Do Workplace Wellness Programs Do?," Quarterly Journal of Economics, 2019. 95% CI ruled out 84% of prior positive findings.
- Hackman, J.R. & Oldham, G.R., Job Characteristics Model, validated across 200+ empirical articles. Karasek, R., Demand-Control Model.
- Google, "Project Aristotle," 2015. 180+ teams, 250+ team-level variables.
- Gallup, manager engagement meta-analysis. MIT Sloan, "Toxic Culture Is Driving the Great Resignation," 2022. 34 million profiles, 1.4 million Glassdoor reviews.
- Edelman, "2025 Trust Barometer." PwC, "2024 Trust Survey."
- Braganza, A. et al., "Productive Employment and Decent Work: The Impact of AI Adoption on Psychological Contracts," Journal of Business Research, 2021. N=232.
- Gartner, monitoring adoption data, 2025. Amazon France / CNIL fine: €32 million, January 2024.
- Costco, Patagonia, and Mondragon retention and ownership data from company reports and NCEO (National Center for Employee Ownership).
- EU AI Act (Regulation 2024/1689). EU Platform Work Directive (Directive 2024/2831). CJEU SCHUFA ruling, C-634/21, December 2023.
- Cornell/CAT Lab, "NYC Local Law 144 Compliance Study," 2024. NYS Comptroller audit, December 2025.
- PwC, "2025 Global AI Jobs Barometer." Humlum & Vestergaard, NBER Working Paper, September 2025.
Share this post