The Silence of the Lambs: Corporate Edition

A slick boardroom presentation. The rollout of AI across the organization. Profit misses estimates? AI will improve productivity in the next quarter.

Gen-AI tools have rolled out to thousands of employees, and adoption dashboards glow with healthy login rates.

And yet.

Nine in ten firms using AI report no measurable impact on productivity or employment after three years. That's finding from a recent The National Bureau of Economic Research (NBER) survey of roughly 6,000 executives across advanced economies.

PwC’s 29th Global CEO Survey echoes the same frustration.

More than half of the CEOs surveyed saw no meaningful revenue gains or cost cuts from AI in the past year. McKinsey & Company’s 2025 State of AI report notes that while 78% of organizations use AI somewhere, more than 80% report no material EBIT impact. The "AI high performers" (those generating 5%+ of EBIT from AI) make up just ~6% of the sample.

This is known as the Solow Paradox, or alteast in a rebooted format. AI is everywhere in strategy decks and earnings calls. It’s just not showing up in the productivity numbers.

The usual suspects--bad data, weak change management, poor model choices, governance gaps--are real enough. But there’s a quieter culprit that many C-suites gloss over. Partly because it sounds "soft." Partly because addressing it means admitting something uncomfortable about how they actually manage people.

The productivity gap has a human name.

Amy Edmondson’s research, and Google ’s Project Aristotle that followed, identified psychological safety as the single strongest predictor of team learning and performance.

The question is simple: Do people believe they can ask questions, admit mistakes, or challenge decisions without fear of punishment or embarrassment?

The question is simple: Do people believe they can ask questions, admit mistakes, or challenge decisions without fear of punishment or embarrassment?

Recent employee data sharpens the point.

In high-safety environments, roughly 70% of people feel confident actually using AI tools effectively. In low-safety ones, that drops below 50%--even with equal access and training.

The gap isn’t about technical skill. It’s about whether people feel safe enough to experiment.

Successful AI adoption demands vulnerability. Employees must reveal what they don’t understand about the model. They must flag hallucinations, biases, or outright wrong outputs. They must run experiments that fail--sometimes publicly. And they must push back on leadership’s workflow choices. In low-safety cultures, none of that happens. People perform "AI use" for the dashboard while the organization learns little.

Worse, many AI programs quietly erode the very conditions they need to succeed.

Surveillance arrives first. AI monitoring tools are sold as productivity aids. On the ground, they often feel like digital overseers. Research from the APA and Cornell University shows higher stress, lower autonomy, and increased resistance--exactly the opposite of the curiosity AI requires.

Then comes the silence around jobs. Over 90% of workers see potential in gen AI, but nearly 60% fear it will replace them. Most organizations lack any real plan for displaced roles. Then headlines fill the void. Goldman Sachs estimates AI is netting a loss of thousands of U.S. jobs monthly, hitting Gen Z and entry-level roles hardest. Empty reassurances from the top don’t land.

The tool sprawl seals it. Productivity often rises with 1–3 tools, then drops. For many, AI is just another thing to juggle on top of unchanged legacy work.

The result? Formal adoption, structural failure.

High-performing organizations do things differently. They redesign workflows end-to-end, stripping out legacy steps. They create roles instead of just cutting heads. Crucially, they build cultures where a frontline worker can say "this tool is producing garbage" and see action--fast.

Crucially, they build cultures where a frontline worker can say "this tool is producing garbage" and see action--fast.

A simple 2x2 diagnostic is useful: psychological safety vs. AI maturity. Most companies sit in the bottom half—either low-maturity theater or high-maturity silent failure. The top-right quadrant, where people fluently challenge AI outputs, is where real EBIT impact lives.

Five practical moves (none requiring a five-year transformation):

Have every executive committee member run a live gen-AI session on real work and narrate their failures publicly. The signal—"we’re all figuring this out"—matters more than you’d think.
Pick one high-friction workflow per function. Run a 6–8 week sprint with frontline staff, process owners, and AI specialists. As you add AI, cut 10–15% of legacy steps. No net simplification? Iterate.
Publish a role-by-role AI exposure analysis with honest reskilling paths. (Note the disparity: ~10% of female employment sits in highest-risk jobs vs. 3–4% for men.)
Create a lightweight, no-blame AI incident log with clear escalation. Early on, more reports are a good sign.
Track psychological safety on your AI dashboard—right next to accuracy and adoption--using a short, anonymous Edmondson-scale survey. Break it down by team.

The models are getting cheaper and more powerful by the month. The tooling is accessible. What remains scarce--and hard to fake--is an organization where people will quickly and honestly tell you when something isn’t working.

That’s what the productivity statistics are still waiting for.

The author is the head of Research and Analysis at Icarus Asia, a Hong Kong-based risk and advisory business that covers private credit, structured finance, and alternative investments across Asia.

This draws on our research note "Afraid to Ask: Why AI Underperforms in Silent Workplaces" (June 2026).

The productivity gap has a human name.

Sources

Related

Google Finds a Way to Skip the Training Step for Spreadsheet Predictions

History Doesn't Repeat. But the Mellons Do.

The Wall That Private Credit Built