Afraid to Ask:
Why AI Underperforms
in Silent Workplaces
Enterprises spent record sums on AI in 2025. Most have little to show for it. The bottleneck is not the technology — it is whether employees feel safe enough to actually use it.
AI budgets have doubled in three years. The productivity gains have not. PwC's 29th Global CEO Survey of 4,454 executives across 95 countries found 56% reporting no measurable revenue gains or cost reductions from AI over the past year. An NBER study of roughly 6,000 executives found nine in ten reporting no firm-level impact on productivity or employment. McKinsey's 2025 State of AI survey puts organizational adoption at 78% of organizations — and material enterprise-level EBIT impact below 20%.
The gap has a mechanism. Perceptyx research found that 70% of employees in organizations with high psychological safety feel confident using AI effectively. In low-safety organizations with equivalent tool access, that figure drops to roughly 50%. The gap is not training, budget, or technology stack. It is trust.
This note maps the adoption-impact gap, explains the three organizational practices most reliably eroding psychological safety in AI programs, and offers six specific leadership moves for closing it — with concrete examples, leading indicators, and a 90-day self-assessment framework any business unit can deploy immediately.
Section 01 — The Adoption Paradox
AI Is Deployed Widely but Failing to Move Earnings
PwC's 29th Global CEO Survey, covering 4,454 executives in 95 countries, found 56% reporting no revenue gains or cost reductions from AI in the past year. Only 30% reported any revenue increase, and a mere 10–12% saw both. The NBER's "Firm Data on AI" study, covering roughly 6,000 executives in the U.S., U.K., Germany, and Australia, found nine in ten reporting no observable impact on their firm's productivity or employment over the prior three years — despite roughly 70% already deploying some form of AI. McKinsey's 2025 State of AI survey found 78% of organizations using AI in at least one function and more than 80% reporting no material EBIT impact at the enterprise level.
Six percent of organizations qualify as AI "high performers" by McKinsey's definition, generating 5% or more of EBIT from AI. They differ from the rest not primarily in their technology stack but in how they pair AI investment with workflow redesign, executive oversight, and deliberate capability-building. The same pattern appears in every major survey of AI outcomes: the gap between adoption and impact is a management problem, not a software problem.
Of organizations using AI, just 6% achieve ≥5% of EBIT from AI. The remaining 94% have deployment without measurable returns at the enterprise level.
The Employee Reality
Section AI's 2026 AI Proficiency Report adds a detail that matters. Fifty-five percent of enterprise employees use AI tools at least weekly. Eighty-five percent of them lack any use case that delivers real business value. Fewer than 3% qualify as advanced practitioners who actually integrate AI into existing workflows at scale. The explanation is not resistance — it is employees using tools they were given, on top of workflows that haven't changed, with no clear signal of what proficiency looks like or what happens to their jobs if they fail to achieve it.
Labor-Market Context
Goldman Sachs Research estimates AI substitution is eliminating roughly 25,000 U.S. jobs per month while augmentation creates around 9,000 net new roles — a drag of 16,000 positions monthly. Gen Z and entry-level workers bear a disproportionate share of those losses, particularly in customer service and junior back-office functions. The International Labour Organization finds that 9.6% of female employment in high-income countries sits in the highest risk category for AI automation, compared with 3.5% for men, because women are overrepresented in clerical and administrative roles. Employees in high-exposure jobs are not wrong to feel uncertain. Generic reassurance about AI as "opportunity" typically makes the problem worse, because it requires people to distrust their own reading of observable facts.
per month
junior back-office
per month
workflow design
Section 02 — The Missing Variable
Psychological Safety Explains the Gap That Technology and Training Cannot
Psychological safety is the shared belief among team members that the team is safe for interpersonal risk-taking: that people can ask questions, admit mistakes, and challenge ideas without fear of embarrassment or retribution. Amy Edmondson at Harvard Business School identified it as the strongest single predictor of team learning and performance. Google's Project Aristotle, which studied 180 internal teams, found it was the top factor by a significant margin. Teams with high psychological safety are 27–30% more likely to outperform their peers.
In AI programs, the connection is direct. Using AI effectively requires employees to ask questions that reveal ignorance about how models work. It requires admitting when AI outputs are wrong or misaligned. It requires experimenting with approaches that will fail before they succeed. And it requires challenging leadership decisions about where AI gets deployed — which very few employees will do if they believe the personal cost is their reputation, their performance review, or their job.
In organizations with high psychological safety, 70% of employees feel confident using AI effectively. In low-safety organizations with similar tool access, that figure barely exceeds 50%. The gap is not closeable with training. Trust and safety conditions are decisive.
When employees don't feel safe, they perform what researchers call "competence theater." They use AI minimally in officially approved ways. They don't report when it fails. They don't share failed experiments. They don't push back on deployment decisions. The organization collects license utilization data that looks healthy while the learning loop that would generate real value quietly stops turning.
"The organization collects license utilization data that looks healthy while the learning loop that would generate real value quietly stops turning." Icarus Asia Research · June 2026
The Safety × Maturity Matrix
Mapping organizations on two axes — psychological safety (low to high) and AI maturity (low to high) — produces four distinct patterns. Most large enterprises have business units distributed across all four. The question is whether leadership is deliberately steering toward the top right or drifting.
"We're figuring it out — together"
Limited AI use, but teams experiment openly, share failures, and co-create use cases. Where the best early-stage pilots happen. Capable of moving right quickly under good leadership.
"AI makes my work better, and I make AI better"
AI is deeply embedded in workflows. Employees understand its limits. Questioning outputs is a norm, not a risk. Errors get reported, not concealed. This is where AI generates sustained returns.
"Another rollout. Let's wait it out."
Limited experimentation, low trust. Employees use tools minimally or privately. Leadership sees thin results and loses confidence. Where most organizations actually sit.
"I won't say anything. I'll just fix it myself."
AI is widely deployed but employees fear challenging outputs or reporting incidents. Efficiency gains may appear on paper while errors, compliance risks, and cultural damage accumulate underneath.
Icarus Asia framework · Rows: psychological safety (high top, low bottom) · Columns: AI maturity (low left, high right)
Section 03 — What Breaks It
Three Management Habits Kill the Feedback Loops AI Needs
When companies deploy AI primarily to monitor employees — tracking keystrokes, active time in applications, and AI usage rates — they send a clear message about the relationship between the organization and its people. The American Psychological Association found that 56% of workers subjected to electronic monitoring report feeling tense or stressed about work most of the time, compared with roughly a third of non-monitored workers. Cornell researchers found that AI-driven monitoring specifically produces more complaints, worse performance, and higher quit intentions.
When the first use case employees experience for AI is a system watching what they do, they do not subsequently trust AI programs designed to "help" them. The rollout sequence matters. Organizations that lead with surveillance create precisely the conditions that prevent AI from ever becoming useful.
Most organizations introduce AI with limited transparency about which roles face the highest risk of task automation, what the organization's plan is for affected employees, and why specific tools are being deployed in specific workflows. Accenture's 2024 research found that while over 90% of workers see potential value in generative AI, nearly 60% are simultaneously concerned about job loss and burnout. Fewer than one-third of C-suite executives believe job displacement is a serious concern for their people — a perception gap large enough to destroy credibility on its own.
When employees don't have accurate information, they fill the gap with worst-case scenarios. And given the Goldman Sachs data on net job losses and the ILO's findings on gender asymmetry in AI exposure, they are not unreasonable to do so. Transparent communication about real risks, paired with specific plans for affected cohorts, is the only response that works.
BCG research found that workers' self-reported productivity increases when they use up to three AI tools, then declines as the number rises to four or more — what BCG's researchers call "AI brain fry." Section AI's data align with this: most employees are stuck in fragmented, basic use cases rather than integrated workflows. A common pattern: a company deploys AI across sales, service, and merchandising simultaneously. Each function uses different tools. None integrates with core systems of record. No legacy reports or approval steps are removed. Six months later, license utilization looks healthy. Frontline employees describe AI as "just another tab."
McKinsey's 2025 survey identified redesigning workflows as the single organizational attribute most correlated with EBIT impact from AI. Only about one in five organizations using generative AI has fundamentally redesigned even some workflows. When AI is piled onto broken workflows, it produces fatigue and skepticism — the precise conditions that prevent AI from scaling.
Section 04 — Leadership Playbook
Six Moves That Close the Adoption-Impact Gap by Treating Safety as a Design Constraint
Leaders who only share AI success stories send a message: "Everyone else gets it but me." Leaders who demonstrate their own AI learning curve — including failures — change the calculus for everyone below them. One large financial institution's COO began monthly AI learning sessions where senior executives showed in real time how they used generative AI for memos, data analysis, and board preparation, including misfires. Over six months, attendance grew and internal surveys showed a double-digit increase in employees agreeing that "leaders are learning alongside us" — which correlated with higher AI experimentation in frontline teams.
When AI is presented primarily as a tool for reducing costs and eliminating roles, employees focus on protecting themselves rather than co-creating new workflows. One global consumer packaged-goods company deployed generative AI in marketing and procurement, generating measurable savings in media spend and contract review time. Instead of banking all savings, the company publicly committed to reinvesting a portion into new digital roles and reskilling programs. Over two years, the firm grew digital product teams and redeployed hundreds of employees from exposed roles into new positions, reducing involuntary exits and increasing internal mobility.
One European insurer selected a single claims process as a "model line" for AI-enabled redesign. The team mapped the entire workflow, introduced AI for document extraction and triage, eliminated redundant approvals, and created a specialist role for handling exceptions. Cycle times dropped by around 30%, error rates fell, and employee satisfaction improved. A comparable insurer that added an AI triage tool without adjusting approval steps found employees now had to reconcile AI recommendations with existing checklists — doing the same work twice. Within three months, the tool had been quietly sidelined.
The operating principle: for every AI capability added to a workflow, remove at least 10–15% of legacy tasks or handoffs. The burden of proof belongs on the workflow, not on employee willingness to adopt yet another tool.
One global technology company established quarterly "AI experiment weeks" where cross-functional teams tested AI ideas on real business problems. Teams were explicitly rewarded for documenting failed experiments alongside successes. The volume of experiments grew over four quarters, and several high-impact use cases emerged from finance and supply chain rather than IT — exactly the areas where AI was most likely to matter commercially and least likely to be explored organically. A lightweight experiment registry (hypothesis, context, outcome, reusable patterns) made the cultural aspiration observable: leaders could see where experimentation was alive and where teams needed more support.
A regional bank conducted a role-by-role AI exposure analysis and found that clerical operations and customer service roles — predominantly staffed by women and younger workers — faced the highest risk of task automation. Rather than issuing a general statement about AI supporting all employees, the bank committed to a multi-year transition plan: role-specific upskilling pathways into relationship management, fraud detection, and data quality roles; protected transition roles with salary guarantees; and joint design workshops with affected staff to shape the workflow changes themselves. The three moves that made the biggest difference were naming the risk, resourcing the transition, and tracking who actually benefited by demographic group.
A large healthcare provider created an AI governance council chaired by the CEO, with representation from clinical leadership, HR, legal, and frontline staff. The council established clear guidelines on appropriate use, human oversight requirements, and escalation paths for AI-related incidents. Employees reported higher trust in AI systems and greater willingness to use them after this governance structure was visible and operational — not because governance was imposed but because it signaled that the organization would not offload blame onto individuals when systems failed. When governance is co-designed with business and frontline teams rather than handed down, it becomes a contributor to psychological safety rather than a constraint on innovation.
Appendix A — Self-Assessment
A 7-Item Diagnostic and KPI Scorecard for Business Unit Leadership Teams
Use a 1–5 scale for each statement: 1 = not at all true, 3 = somewhat true, 5 = consistently true. Complete individually, then discuss as a team in a 90-minute workshop.
7–17: AI program is likely eroding psychological safety. Prioritize leader modeling, error-reporting norms, and clear communication with high-exposure cohorts.
18–27: Mixed picture. Focus on the lowest-scoring items and design targeted interventions within 90 days.
28–35: Strong foundation. Shift emphasis to scaling, equity tracking, and governance as AI maturity increases.
| Metric | What to Measure | Cadence |
|---|---|---|
| Team psychological safety | Edmondson 7-item scale at team level, with AI-specific items added (e.g., comfort questioning AI outputs) | Annual; quarterly in high-change environments |
| AI adoption quality | Active-use rates by role; diversity of use cases (distinct workflows per team) rather than login counts alone | Monthly dashboard |
| Learning and experimentation | AI experiments per quarter; share with documented outcomes in a central registry; cross-functional participation rate | Quarterly review |
| Error and incident transparency | AI incidents reported per 100 active users (target non-zero during learning phases); average time from detection to triage | Monthly |
| Work redesign and impact | Workflows fundamentally redesigned (not just tool-added); before/after cycle time and error rates; employee agreement that "AI has simplified my work" | Quarterly |
| Talent and equity outcomes | Reskilling participation and completion by demographic group; internal mobility from high-exposure roles; differential regrettable attrition between exposed and non-exposed cohorts | Semi-annual |
Appendix B — Primary Sources
- PwC, "29th Global CEO Survey" (2026). Survey of 4,454 CEOs across 95 countries. pwc.com
- NBER, "Firm Data on AI" (Acemoglu et al., 2025). Survey of ~6,000 executives in U.S., U.K., Germany, Australia. nber.org
- McKinsey Global Institute, "The State of AI 2025" and "How Organizations Are Rewiring to Capture Value." mckinsey.com
- Section AI, "2026 AI Proficiency Report" (January 2026). sectionai.com
- Goldman Sachs Research, "The Jobs AI Is Likely to Boost — and Those It May Disrupt" and related labor-market decomposition analysis, 2025–2026.
- International Labour Organization / UNU, gender automation risk analysis. "The AI Gender Trap" (C3, UNU, 2025).
- Perceptyx, "Trust, Not Tech Skills, Predicts Generative AI Adoption Success" (2025). blog.perceptyx.com
- Edmondson, A. C., "Psychological Safety and Learning Behavior in Work Teams," Administrative Science Quarterly (1999). Foundational source for the 7-item scale.
- Google Project Aristotle, as reported in LeaderFactor and Aristotle Performance summaries (2015–ongoing).
- Accenture, "Work and the Potential of Generative AI" (2024). newsroom.accenture.com
- American Psychological Association, "Electronically Monitoring Your Employees? It's Impacting Their Mental Health" (2025 workplace data).
- Cornell University / SHRM, "AI Surveillance in the Workplace Linked to Employee Resistance" (2025).
This research note is prepared by Icarus Asia Research for informational purposes only. It does not constitute investment advice. All survey data and claims are attributed to named primary sources; “Icarus Asia estimate” is applied where figures are derived rather than directly reported. Verify material statistics with primary sources before use.