03 Sep How Psychological Safety Determines Whether AI Transformation Succeeds or Fails
A Workplace Where Failure Doesn’t Trigger Fear
Sound familiar? An AI project keeps running in plain sight, even though the team has known for a while that it isn’t working. Nobody says anything, because nobody wants to be the one left holding the bag.
Whether a team names that failure out loud or stays quiet about it comes down to a single factor: psychological safety.
Amy Edmondson, a professor at Harvard Business School, laid the groundwork back in 1999. She defined psychological safety in a 1999 study as a work environment where taking a risk, voicing a half-formed idea, admitting a mistake, or asking an uncomfortable question doesn’t cost you your career or your reputation.
Just how solid the evidence for this has become shows up in the European Workforce Study 2025 by Great Place To Work, the largest pan-European survey on the subject to date, with nearly 25,000 respondents across 19 countries. In organizations with high psychological safety, 85 percent of employees rate their employer as a great place to work, and even customer satisfaction there sits at 79 percent, so the effect doesn’t stop at the workforce.
That’s essentially what Edmondson found in her original research: teams who could speak openly got more out of their mistakes and improved measurably over time. Countless follow-up studies have confirmed the same thing since: a measurable performance gap, not cultural window dressing.
This isn’t just a coincidence, either. Leadership itself now backs it up: a 2025 Workplace Options study on psychological safety found that 93 percent of leaders worldwide now see it as a driver of innovation, not a side note to company culture.
Contents
Where Psychological Safety Meets Al Transformation
How Al Undermines Psychological Safety
How to Spot an “Al-Ready” Organization
How to Make Al Transformation Work in Your Organization
Why Structure Alone Isn’t Enough
Case Study: What Psychological Safety Looks Like at Infosys
Conclusion: Psychological Safety in Al Transformation Emerges from Structure and Pattern
Where Psychological Safety Meets AI Transformation
Failing the Right Way
Bringing AI into an organization means stepping into unfamiliar territory, no matter how well the tools eventually work. Rolling out new software and running a few training sessions doesn’t cover it. Models need tuning, established processes get questioned, and roles sometimes need a complete rework, and there’s simply no way to know in advance what will hold up. Not every deviation along the way is an avoidable mistake. Many of them are honest misjudgments that only become visible in hindsight, and that’s exactly why they’re part of the process, not the exception to it.
The moment someone fears being held personally responsible if an AI project fails, behavior shifts, quietly but noticeably: fewer experiments, more guarded feedback, problems that get buried instead of raised. In the worst case, the very system everyone already knows isn’t working keeps running anyway. Nobody brings it up anymore.
Just how widespread this silence is shows up in the MIT Technology Review Insights study “Creating psychological safety in the AI era”: of the 500 leaders surveyed worldwide, roughly one in five, 22 percent, had already held back from proposing or leading an AI project at some point, out of fear of the internal fallout if it failed. That fear becomes an innovation killer in its own right, before a single project even gets off the ground.
How AI Undermines Psychological Safety
That much holds up. But the relationship between AI and psychological safety runs a lot deeper than that.
Trust Ambiguity: The Quiet Distrust of AI
Amy Edmondson and Jayshree Pandya take this a step further in their 2025 Harvard Business Review piece on AI and psychological safety, showing that AI can damage trust in a way that looks nothing like a human mistake. The reason: AI delivers wrong answers with exactly the same confidence as right ones, and that gap is what the authors call “trust ambiguity”: the sense that you’re supposed to trust the AI but can’t quite bring yourself to. Because that discomfort is hard to name out loud, people end up losing trust not just in the AI’s output, but in their own judgment too.
Mistake or Misjudgment: Wohland’s Distinction
One reason this ambiguity is so persistent lies in the nature of the deviation itself: it’s not easy to judge whether a wrong AI answer could have been avoided at all. This is where a distinction we touched on earlier comes into sharper focus. Dr. Gerhard Wohland, a physicist, management consultant, and pioneer of organizational development, distinguishes between two types of deviation in his work:
❌ Mistake: an avoidable deviation from something that was already known, a violation of existing knowledge.
⭕️ Misjudgment: arises where genuine uncertainty exists, where no one could have known beforehand what the right answer even was.
Wohland also draws a line between two types of problems:
🔵 Blue problem: complicated, but solvable with existing knowledge and fixed rules. There’s a known right and wrong you can check a solution against.
🔴 Red problem: complex, with no such known right and wrong beforehand. Whether a solution holds up only becomes clear once you try it.
Apply that to AI, and the difference becomes concrete. With a blue problem, say, rule-based document review, an AI output can be checked against the known right and wrong. If it deviates, that’s a mistake in the true sense, avoidable and explainable after the fact. With a red problem, like a new product decision, that right and wrong never existed in advance. If an AI-supported decision only turns out not to hold up after the fact, that’s a misjudgment rather than a preventable rule violation, made worse by the fact that the black box gives no insight into the reasoning even in hindsight.
Why AI Mistakes Land Differently Than Human Mistakes
Beyond trust ambiguity, Edmondson and Pandya draw a second important distinction in the same 2025 Harvard Business Review piece: the one between human mistakes and AI mistakes.
When a human team member makes a mistake, it often kicks off a cooperative learning process. Colleagues ask questions (“What data did you use?”, “How did you get there?”), understand the context (“I had too many balls in the air at once”), and work out a fix together, say, by building in a second pair of eyes going forward. Mistakes that a team can talk through and process this way can strengthen the group.
That mechanism doesn’t break down with a human mistake. It does with an AI mistake. That’s down to the so-called black box problem: neither the assumptions behind an AI decision, nor its methodology, nor the individual reasoning steps can be reliably reconstructed after the fact. A team can sense that something is off with the AI’s output, but has no reliable way to resolve the uncertainty or stop the mistake from happening again.
That creates a genuine dilemma: AI adoption depends on a mature culture of learning from error, and AI itself puts that culture under pressure. The next section looks at how organizations deal with that.
How to Spot an “AI-Ready” Organization
One pattern runs through every organization that gets AI adoption right: they never separate technological development from cultural development in the first place. They think about both together, from day one.
What the Numbers Show
➡️ European Workforce Study: In organizations considered “AI-ready”, 75 percent of employees get recognized for trying new approaches, regardless of the outcome.
➡️ MIT study: Among companies with the highest AI project success rate, where 76 to 100 percent of projects make the leap from pilot to production, three out of four have a culture that actively supports experimentation.
This lines up with what we see in consulting practice: AI surfaces whatever is already there in an organization. Where trust and a willingness to learn already exist, AI acts like a catalyst. Where they don’t, it magnifies exactly those weak spots. That’s why open, learning-oriented cultures gain disproportionately from AI transformation: they provide the protected space needed to experiment and learn, while highly hierarchical, risk-averse organizations fall behind.
Still, nobody can count on a culture of learning and tolerance for error holding up on its own once it’s built. Culture gets recreated in every single interaction. Because AI transformation both depends on a psychologically safe, learning-ready culture and can undermine that same culture, keeping it alive is an ongoing job, one that has to be built into the structure rather than left to good intentions.
How to Make AI Transformation Work in Your Organization
Where Psychological Safety Matters Most
Before getting into concrete levers, one distinction is worth making. Not every AI application needs the same amount of psychological safety:
🔵 Blue problem, like document classification or standard reporting: what’s needed here is primarily standards and quality control, not a special tolerance for error.
🔴 Red problem, like a new product, a new customer solution, a market with no known answer: this is where psychological safety, in the sense described here, is the decisive lever, because the solution only emerges through experimentation.
Mix the two up, and you end up building safety measures for problems that don’t need them, while missing the places where it counts.
Six Levers for Psychological Safety
Psychological safety can’t be manufactured directly, whether through a communications program or through structural change alone. Even so, consulting and transformation practice, together with the research covered here, point to six structural levers that create the conditions under which that kind of culture can take shape during AI transformation:
1. Build clarity about job impact into the process itself. People need clarity about what AI adoption means for them personally. A one-off announcement or an individual leader’s good intentions can’t deliver that. It takes a fixed, recurring point in the process where this information has to come up, a standing agenda item in town halls or development conversations, for example. The structure has to force that openness, an appeal for it won’t.
2. Clarify accountability for AI-supported decisions. Psychological safety doesn’t come from individual leaders modeling uncertainty. It certainly helps when a leader does that, speaking openly about their own uncertainty and inviting pushback, but that alone doesn’t hold up if the decision rights behind it stay unresolved. What matters is knowing who’s accountable when an AI-supported decision goes wrong. Someone needs to review the outcome before it takes effect. And a team needs somewhere to escalate if they don’t trust the AI, three things that rarely get written down anywhere. Where that escalation goes should be based on demonstrated competence, not automatically the next rung up the hierarchy, otherwise escalation is just a formality. Without that clarity, accountability stays vague, and vague accountability produces exactly the kind of caution that keeps people from speaking up.
3. Build experimentation into the structure. If you want people to experiment, it has to be written into processes and budgets. Low-barrier pilot projects, dedicated learning spaces, and explicit permission to try things out aren’t nice-to-haves. They’re structural decisions. In organizational development, this is called a protected project space: a time-boxed, ring-fenced experiment with a clear external anchor and a sponsor who shields it from the rest of the organization. That can also include playful, informal formats where people can try things out without pressure to deliver a result.
This also means institutionalizing sense-making around AI mistakes. When an AI outcome can’t be fully explained, the team still needs a place to process that misjudgment together, even without full clarity on the cause. That can be a standing routine, a short retro after every major AI-supported project step, where the question isn’t why the AI did what it did, but what the team does now that it doesn’t know. That shift in the question replaces the missing explanation with a shared decision.
4. Get rid of individual success metrics that punish failure. The obvious instinct is to add extra rewards for learning. The more effective lever is the opposite one: get rid of individual targets and bonuses tied to visible AI project success. An AI project is usually a complex, red problem, its success comes out of a team working together, so holding one person accountable for it is a category error to begin with. There’s a simpler reason too: anyone whose personal metrics depend on a visible project win ends up carrying the risk of failure alone, which means it’s not worth trying anything genuinely uncertain, and a struggling project gets dressed up rather than surfaced. And an individual metric shifts the point of reference. Instead of orienting toward real value for the organization or the customer, behavior starts orienting toward hitting the number, people optimize for the metric, not for the project working. Where success is measured individually, failure gets avoided individually too, by staying quiet about it if necessary.
5. Put AI decisions and organizational development in the same room. AI transformation is neither a pure technology project nor a pure culture project, yet in practice, decision-making authority almost always ends up sitting with IT or a digital transformation team alone. That structural split is exactly what becomes the biggest bottleneck. A better approach ties AI projects above a certain scale to a joint decision-making body that combines technical and organizational-development responsibility, so that questions of roles, process, culture, and who actually gets a say get decided from the start, not bolted on afterward.
6. Use psychological safety as a diagnostic, not a new metric. What doesn’t get measured can’t be deliberately changed. Regular pulse surveys and structured feedback formats are valuable, but only if the results lead to a structural decision. If a survey shows, say, that a team doesn’t know who’s accountable when an AI project fails, that’s the trigger to sharpen the escalation paths from point two, not to massage the number itself. The moment psychological safety becomes a target with a number attached to it, it turns into a new internal reference point, one that creates exactly the fear it was meant to reduce.
These six levers aren’t a checklist to work through. They’re starting points whose relative weight shifts from one organization to the next. Which one matters first is a question of experience and judgment, not a method you can simply copy.
Why Structure Alone Isn’t Enough
These six levers don’t operate in a vacuum. The language of caution, concealment, and fear that runs through levers two, four, and six already signals that more is going on here than pure structural work. Organizational consultant, executive coach, and author Klaus Eidenschink describes organizations as psychodynamic and organizational-dynamic systems at the same time, neither one reducible to the other.
People bring their own psychodynamic patterns into an organization: the fear of visible failure, say, or the need to avoid conflict. Those patterns are real, and no amount of restructuring makes them disappear. At the same time, a pattern like that tends to seek out exactly the organizational structure that confirms it. People who avoid conflict often end up in organizations that reward that avoidance with a calm surface.
A leader who personally fears the failure of an AI project doesn’t experience an individual success metric as an external constraint. They experience it as confirmation of their own pattern. Change only the structure without addressing the psychodynamic pattern, and the avoidance usually finds a new outlet. Work only on the pattern without changing the organizational structure, and the person slides back into old habits the next time real pressure hits. Psychological safety, in other words, needs both organizational dynamics and psychodynamics working together.
In practice, that’s what this means for us at CoHive: clarify the system first, then look at the person. We diagnose the organizational-dynamic level before we take on the psychodynamic one. That doesn’t mean the second level matters less. In our experience, working on personal patterns alone, through coaching or personal development, tends to hold up only as long as the surrounding structure supports it: a leader who goes through coaching and then returns to an organization that still rewards conflict avoidance tends to slide back into the old pattern before long.
Case Study: What Psychological Safety Looks Like at Infosys
The technology company offers a concrete example of how structure and pattern work together in practice.
Psychological safety here emerges from a structural safety net, not a communication goal:
➡️ An internal incubator where employees can pitch their own AI ideas, with a clear guarantee that they can return to their previous role if the project fails.
➡️ New career paths in fields like responsible AI and small language models.
➡️ Deliberate support for role changes into AI-adjacent areas, even when the outcome is uncertain.
What stands out is that none of these three measures target behavior change directly. They’re all structural decisions that make a certain behavior possible in the first place, not culture initiatives that demand it.
In the MIT study, Sushanth Tharappan, Head of Human Resources at Infosys, sums up the principle well: the company builds new roles and makes lateral moves possible. Whether that always plays out exactly as planned is an open question, but that’s fine too, because the point is iteration, not perfection.
Conclusion: Psychological Safety in AI Transformation Emerges from Structure and Pattern
Picture that AI project again, the one still running in plain sight while everyone already knows it doesn’t work. Whether that happens in your organization isn’t something you can diagnose by looking at how open the communication feels. It comes down to a single question: what happens to the person who’s first to say it isn’t working? Do they carry that alone, or is there a structure built to carry it with them? That’s what decides whether an AI misjudgment turns into a learning curve, the way it did at Infosys, or whether the result is that nobody talks about AI in that area anymore, and the organization quietly slides back into its old, familiar processes.
That decision is being made in a lot of organizations right now, whether deliberately or by default.
So psychological safety can’t be mandated, whether through communication or through coaching alone. It emerges when an organization clarifies its decision-making architecture first, and only then works on the human patterns that help keep that architecture alive.
Get in Touch
If AI projects in your organization tend to fail quietly instead of openly, that’s rarely down to a lack of communication. It’s because structure and personal patterns keep confirming each other, without anyone noticing.
In an initial conversation, I’ll work with you to map out where the real leverage points are in your organization. No charge, no pitch, no obligation.