Is 'They Might Leave After We Train Them' the Right Question? — The ROI and Limits of Safe-to-Fail Environments

• Target audience: Engineering managers, team leads, HR professionals, and engineers interested in organizational learning
• Prerequisites: None
• Reading time: ~10 minutes

Overview

In Build, Break, Repeat: Why Hands-On Failure Is the Ultimate Learning Accelerator, I explored how individuals can use AI to supercharge a “build, break, learn” cycle. Outside of work, psychological safety is a given — you can fail as much as you want.

But should companies be the ones providing that environment?

Intuitively, the answer feels like “yes.” Edmondson’s research demonstrates that psychological safety enhances learning outcomes¹, and Google’s study of over 180 teams identified it as the single most important factor in high-performing teams². Yet business decisions are never that simple. Higher learning outcomes do not automatically translate to higher profits. Trained employees might leave. Is the return worth the investment?

The data paints a surprising picture. The real problem is not “they might leave after we train them” — it is that when you do not train them, untrained people stay and accumulate costs indefinitely. According to Gallup, Japan’s employee engagement rate stands at just 6%. A full 24% of workers are “actively disengaged,” and the resulting opportunity cost was estimated at 86 trillion yen (approximately $560 billion) for 2023 alone³. To put that in perspective, this single-country figure rivals the entire GDP of mid-sized European economies.

The risk of a trained employee leaving is a one-time cost. The cost of untrained, disengaged employees staying is a continuous drain — paid at full salary, every single day.

That said, providing a “safe-to-fail” environment does not solve everything. Many corporate innovation labs fail to deliver expected results (the widely cited “90% failure rate” lacks a clear methodological source, though the failure patterns themselves are well-documented⁴), and “innovation theater” — claiming to embrace failure while doing nothing of the sort — is pervasive. This article examines the evidence for investing in safe-to-fail environments and the realistic limits of that investment.

Do Trained Employees Actually Leave?

The Data Shows the Opposite

“If we invest in our people, they will just upgrade their skills and jump ship.” This concern is what economists call the “poaching externality.” Theoretically sound. But large-scale survey data consistently shows the reverse.

LinkedIn’s 2024 Workplace Learning Report (surveying thousands of L&D professionals and employees) found that providing learning opportunities is the most effective retention strategy organizations can deploy. Companies with strong learning cultures see retention rates improve by 57% and internal mobility rise by 23%⁵.

Meanwhile, the Work Institute’s 2024 Retention Report (analyzing over 20,000 exit interviews) repeatedly identifies the lack of career development opportunities as a primary driver of turnover. One in three new hires begins looking for another job early on, citing inadequate onboarding experiences⁵.

In other words, people do not leave because you trained them. In many cases, they leave because you did not.

The Cost Asymmetry

There is a fundamental asymmetry in the cost structure.

When a trained employee leaves:

• Loss of training investment (one-time)
• Recruitment and onboarding costs (roughly 33% of annual salary)
• Total: training cost + ~1/3 of annual salary. A one-time expense.

When an untrained, disengaged employee stays:

• Full salary paid while productivity steadily declines
• Knowledge and skills atrophy over time
• Negative impact on surrounding team morale (disengagement is contagious)
• This cost accrues monthly, yearly, indefinitely

A one-time risk versus a daily, compounding drain. The math is not close.

The Cost of Not Investing Is Larger Than You Think

The Global Disengagement Problem

Gallup’s 2026 State of the Global Workplace report (based on 2025 data) reveals that global employee engagement has fallen to 20% — the lowest since 2020 and the first consecutive two-year decline in the survey’s history³.

The estimated productivity loss from low engagement in 2024 reached $10 trillion worldwide — roughly 9% of global GDP³.

Japan as an Extreme Case — and What It Tells Us Globally

Japan’s numbers are especially stark, but they illustrate a pattern relevant everywhere³:

• Engagement rate: 6% (among the lowest globally)
• Actively disengaged: 24%
• Disengaged employees outnumber engaged ones 4 to 1
• Estimated opportunity cost in 2023: 86 trillion yen (~$560 billion, Gallup estimate)

These figures are not just a Japanese problem. They represent the extreme end of a global spectrum. Wherever engagement is low — and Gallup’s data shows it is low nearly everywhere — the same cost dynamics apply. The question is one of degree, not kind.

Given these numbers, is “they might leave after training” really the top concern?

Does Psychological Safety Generate Profit?

The Effect on Team Learning Is Well-Established

Edmondson’s 1999 study (51 manufacturing teams, peer-reviewed) demonstrated that psychologically safe teams exhibit more learning behaviors, which in turn mediate team performance¹. Google’s Project Aristotle (180+ teams) confirmed psychological safety as the most important factor in high-performing teams².

In the software development context, the DORA State of DevOps Report (2019) found that psychological safety predicts software delivery performance⁶.

Direct Causation to Profit Remains Unproven

However, to be honest: no randomized controlled trial (RCT) has established a direct causal link from psychological safety to profit.

The evidence chain looks like this:

Psychological safety -> Learning behavior activation (proven¹) -> Team performance improvement (proven¹²) -> Company profit improvement (correlated, causation unproven)

Ellinger et al. (2002, peer-reviewed, n=208) found a positive correlation between learning organization dimensions and financial performance, but it is a correlation — not causation — and the sample was predominantly manufacturing⁷.

“Psychological safety promotes learning” is solid. “Learning improves performance” is solid. But the direct express route from “psychological safety to profit” remains academically unconfirmed. With that intellectual honesty in mind, let us consider what we can say.

Safe-to-Fail Initiatives — What Works and What Is Theater

Companies deploy many initiatives to create “safe-to-fail” environments. The problem is that the gap between what actually works and what is merely for show is enormous.

Examples That Worked

Atlassian ShipIt Days — A quarterly 24-hour hackathon. It started with 14 developers and now involves over 4,000 participants across 20+ cities in 11 countries. Jira Service Management was born from a ShipIt project⁴. The key to success: ideas have a realistic path to becoming actual products.

Google’s 20% Time — Produced Gmail, Google News, and AdSense. However, according to Laszlo Bock (former SVP of People Operations at Google), actual usage was closer to 10%, and the program has since evolved into structured hackathons and innovation weeks⁴. The “concept” of 20% mattered more than rigid enforcement.

The Failure Pattern — “Innovation Theater”

A concept named by Steve Blank in a 2019 Harvard Business Review article⁴. It looks like innovation is happening, but the substance is missing.

Common symptoms:

• Pilot purgatory: Projects remain stuck in testing phases forever, never reaching production
• Activity metrics as success: Reporting “number of hackathons held” or “ideas submitted” as outcomes
• Disconnected from the core business: Innovation teams are siloed away from business units
• Say-do gap: Claiming to “embrace failure” while never actually rewarding anyone who takes risks

McKinsey’s research found that successful organizations are more than twice as likely to actually reward appropriate risk-taking⁴. Saying “it is okay to fail” is not enough. Unless people who take risks are recognized and rewarded through formal systems, no one will believe it.

What Functioning Initiatives Have in Common

Innovation Theater	Functioning Practices
“Generating ideas” is the goal	“Shipping to production” is the goal
Disconnected from business units	Embedded in business units
Dependent on external consultants	Builds internal capability
Measures “hackathons held”	Measures “proposals implemented”

AI Changes the Equation

Into this discussion enters a new variable: AI.

AI dramatically reduces the cost of providing safe-to-fail environments. A randomized controlled experiment with GitHub Copilot showed a 55.8% improvement in developer task completion speed⁸. McKinsey reports 35-45% time savings on code generation tasks with generative AI.

What does this mean in the safe-to-fail context?

Before AI: Trying a new architecture takes weeks -> Failure is expensive -> Cultural transformation needed to tolerate failure -> Transformation is hard -> Nothing changes

With AI: AI generates prototypes rapidly -> Experimentation costs drop dramatically -> Even failures involve smaller investments -> The cultural barrier to “tolerating failure” also drops

A 2025 perspective in California Management Review describes AI-powered experimental sandboxes as a new organizational form that enables limited experimentation without disrupting core operations⁸.

Companies do not need a full-scale cultural transformation. They can start by creating a small experimentation space powered by AI. Like Atlassian’s quarterly ShipIt, begin with limited, periodic experimentation events. If results materialize, expand. No need to change company-wide culture first. Start small, break things, learn — the very theme of this article.

Yet Even This “Small Investment” Gets Rejected

Let us face reality. AI tool subscriptions range from a few dozen dollars per month for basic plans to several hundred dollars per user per month for full-featured ones. For a team of ten, that is tens of thousands of dollars per year. Not trivial.

But in practice, many organizations will not even provide this. Under the banner of “cost reduction,” they decline to adopt AI tools, leaving employees to either pay for personal subscriptions or go without.

Is this rational? If a ten-person team’s annual AI subscription costs tens of thousands of dollars, that team’s total compensation runs into the millions. If disengagement-driven productivity loss is even 10%, the resulting waste is orders of magnitude larger. The tens of thousands “saved” by withholding AI licenses versus the opportunity cost that those tools might have prevented are not in the same ballpark.

Of course, deploying AI tools does not automatically create a safe-to-fail culture. As we have seen, both systems and culture are required. But while cultural transformation is difficult, providing tools is straightforward. AI licenses are one of the lowest-barrier investments in safe-to-fail environments.

Conclusion — How Much Should Companies Invest?

“How much should companies invest in employees’ safe failures?” Here is an evidence-based answer.

First, invert the question. Not “they might leave after we train them,” but “what happens if we do not train them?” Japan’s 6% engagement rate and 86 trillion yen in opportunity costs³ represent a global extreme, but the pattern is universal. This is the current price of “not investing.” And it is being paid every single day.

However, safe-to-fail initiatives are not a silver bullet. The reality that many innovation labs fail to produce results cautions against the illusion that “just build the program and success will follow”⁴. Token gestures of “embracing failure” change nothing.

The evidence points to three conditions for effectiveness:

1. Experiments have a path to production: Not just ideation workshops — a mechanism that takes ideas through to implementation
2. Risk-taking is actually rewarded: “It is okay to fail” functions as institutional policy, not just a slogan
3. It is possible to start small: Not company-wide cultural transformation, but quarterly hackathons or AI sandboxes — limited experiments that can scale

The advent of AI has dramatically lowered the third barrier. What once required significant budget and executive buy-in just to experiment can now be started at the team level with AI tools. You do not need to change the entire company. You can start by breaking something small.

One clarification: this article does not argue that providing the right environment will make everyone learn. For people who lack the motivation to learn in the first place, or who want answers without engaging in the thinking process, no amount of environmental design will activate the cycle. That is a question of motivation, not environment, and lies outside this article’s scope. What we are discussing is what companies can do when motivated people are unable to learn because of their environment.

“They might leave after we train them” — that is true. But Gallup’s data showing 86 trillion yen in opportunity costs means the price of “not training” is already visible³. And the cost of “training” has never been lower than it is in the age of AI.

Build, Break, Repeat covered the individual “build and break” learning method; this article examined how much companies should invest in creating that environment. For a deeper analysis of organizational psychological safety frameworks, see Why “Boiling Water” Training from Japan’s Ice Age Generation Does Not Reach Gen Z.

Related Articles

Explore more articles on this theme:

• Build, Break, Repeat: Why Hands-On Failure Is the Ultimate Learning Accelerator - Companion article. The individual “build and break” learning method
• Only the Hot Water You Choose Makes You Stronger - Individual-level growth conditions
• Why “Boiling Water” Training from Japan’s Ice Age Generation Does Not Reach Gen Z - Psychological safety 2x2 framework
• The More You Use It, the Less You Can Do — The AI Deskilling Paradox in Empirical Research - Organizational impact of AI dependency risk

References

References are listed in order of citation number as they appear in the article.

1. Psychological Safety and Learning Behavior in Work Teams - Edmondson, A., Administrative Science Quarterly (1999). DOI: 10.2307/2666999. Study of 51 manufacturing teams. Peer-reviewed. Demonstrated a mediation model: psychological safety -> learning behavior -> team performance. [Reliability: High] ↩︎ ↩︎² ↩︎³ ↩︎⁴

2. Project Aristotle - Google Re:Work (2016). Internal study of 180+ teams. Identified psychological safety as the most important factor in high-performing teams. Internal research; methodology is partially undisclosed. [Reliability: Medium-High] ↩︎ ↩︎² ↩︎³

3. State of the Global Workplace 2026 - Gallup (2026). Global engagement rate at 20% (2025 measurement, lowest since 2020). Annual productivity loss estimated at $10 trillion. Japan: 6% engagement, 24% actively disengaged, 86 trillion yen (~$560B) opportunity cost (2023). Large-scale survey, though GDP-equivalent cost figures are estimates. [Reliability: High] (Survey scale and methodology are robust; cost conversion is an estimate) ↩︎ ↩︎² ↩︎³ ↩︎⁴ ↩︎⁵ ↩︎⁶

4. Multiple sources on corporate safe-to-fail initiatives. (1) ShipIt Days - Atlassian. Quarterly hackathon scaling from 14 to 4,000+ participants. Origin of Jira Service Management. (2) Google’s 20% Time - Produced Gmail, AdSense, etc., but has since evolved into structured events. (3) Steve Blank, “Why Companies Do ‘Innovation Theater’ Instead of Actual Innovation”, Harvard Business Review (2019). Coined the concept of innovation theater. (4) When failure is an option - McKinsey. Successful organizations are 2x+ more likely to reward risk-taking. [Reliability: Medium-High] (Integration of case studies and surveys. The “90% of innovation labs fail” figure is widely cited but lacks clear methodological sourcing) ↩︎ ↩︎² ↩︎³ ↩︎⁴ ↩︎⁵ ↩︎⁶

5. Multiple surveys on training investment and retention. (1) 2024 Workplace Learning Report - LinkedIn Learning (2024). Learning opportunities identified as the most effective retention strategy. Companies with strong learning cultures: 57% higher retention, 23% higher internal mobility (relative improvement vs. baseline). (2) 2024 Retention Report - Work Institute (2024). Analysis of 20,000+ exit interviews. Lack of career development consistently cited as a primary driver of turnover. [Reliability: Medium-High] (Large-scale surveys, but self-report based) ↩︎ ↩︎²

6. DORA State of DevOps Report 2019 - DORA/Google Cloud (2019). Investigated performance factors in software development teams. Confirmed psychological safety as a predictor of software delivery performance. [Reliability: Medium-High] ↩︎

7. Ellinger, A. D. et al. (2002). The Relationship Between the Learning Organization Concept and Firms’ Financial Performance. Human Resource Development Quarterly, 13(1). Peer-reviewed. n=208. Confirmed positive correlation between learning organization dimensions and financial performance. Correlation, not causation; sample predominantly manufacturing. [Reliability: Medium-High] ↩︎

8. Research on AI-driven reduction of experimentation costs. (1) The Impact of AI on Developer Productivity - Peng, S. et al. (2023). Randomized controlled experiment. GitHub Copilot users completed tasks 55.8% faster. (2) California Management Review (2025). Discussed AI sandboxes as a new form of organizational experimentation. [Reliability: Medium-High] (Copilot experiment is a robust RCT, but generalizing to organizational impact has limitations) ↩︎ ↩︎²