Psychological safety is not a HR initiative dressed up in technology language. It is a direct financial lever. Organisations that create environments where employees can experiment, fail, and speak honestly about what isn't working deploy AI tools faster, waste less budget on shelf-ware, and see meaningfully higher adoption rates than those that simply mandate new software from the top and wait for the magic to happen.

If your transformation programme is stalling, the technology is almost never the problem. The people — and more specifically, how safe they feel engaging with that technology honestly — almost always are.

Why does forcing new technology on people never actually work?

There is a pattern I have seen repeat itself across organisations of every size, sector, and level of IT budget. A senior leadership team returns from a conference — or, increasingly, from a conversation with a vendor who bought them a very nice lunch — convinced that a particular platform will solve a problem they have not yet fully defined. The procurement process moves swiftly. The contract is signed. The implementation partner arrives.

Then nothing happens. Or rather, something happens, but it is not adoption. It is performance. People use the tool in front of their manager and revert to spreadsheets the moment they leave the room.

This is not laziness or resistance for its own sake. It is a rational response to an environment where admitting that something does not work, or that you do not understand it, carries professional risk. If raising your hand to say "this AI output is wrong" might make you look like you're obstructing the transformation agenda, most people will not raise their hand. They will nod, close the browser tab, and carry on as before.

According to McKinsey's research on digital transformation outcomes, roughly 70% of change programmes fail to meet their stated objectives, and workforce resistance consistently ranks among the top three causes. You can spend seven figures on a platform and lose most of it to the entirely human problem of people not feeling safe enough to use it properly.

What actually is psychological safety, and why does it matter here?

Psychological safety, a term coined by Harvard Business School professor Amy Edmondson, describes a team climate in which members believe they will not be punished or humiliated for speaking up with ideas, questions, concerns, or mistakes. In the context of AI adoption, it has a very specific and practical application.

When an employee uses a generative AI tool and gets a confidently stated, completely wrong answer — which will happen, regularly, because that is how large language models currently behave — they need to feel safe enough to flag it. To question it. To tell someone the tool is not fit for the task they were given.

In an organisation where the message from leadership has been "AI is our future, and anyone who struggles with it is struggling with the future," that employee will not flag anything. They will either quietly correct the output and say nothing, or — more worryingly — submit the AI-generated nonsense because they assumed everyone else found it useful and they did not want to be the one who didn't get it.

This is how shadow IT takes root. It is also how perfectly avoidable AI-related errors end up in customer communications, financial reports, and operational decisions.

What is "shadow AI," and how does it connect to this?

Shadow AI refers to the unsanctioned use of AI tools by employees — typically consumer-grade generative AI products — outside the organisation's approved technology stack and governance framework. It is the direct descendant of shadow IT, which gave us employees using personal Dropbox accounts to share sensitive client files because the official file-sharing system was too painful to use.

Shadow AI is not primarily a technology problem. It is a psychological safety and communication failure. When employees feel their legitimate needs are not being heard, or that the official tools are inadequate but they cannot say so without professional consequences, they solve the problem themselves with whatever is available. Usually something free, usually something that processes their data on someone else's servers, usually something that would make your Data Protection Officer require a sit-down.

Gartner estimated that by 2025, more than 40% of employees would be using AI tools that their IT department has no visibility into. The antidote is not surveillance or blanket bans — both of which create resentment without solving the underlying problem. The antidote is building an environment where people can say "the approved tool doesn't do what I need" without fearing the conversation.

How do you actually measure the financial impact of psychological safety?

I understand the scepticism. "Culture" as a boardroom topic has a long and distinguished history of being discussed at length and measured never. But psychological safety, when approached with discipline, produces indicators that translate directly into numbers that finance directors recognise.

Reduced time-to-adoption

Organisations with high psychological safety see faster genuine adoption of new tools — not the performative kind, but the kind where people actually change how they work. Google's Project Aristotle, which studied hundreds of internal teams over several years, found psychological safety to be the single most important factor in team effectiveness. Teams that felt safe to take risks were more productive, more innovative, and — relevant here — more willing to experiment with new processes.

In practical terms: a slower adoption curve means your investment sits idle for longer. If a £500,000 platform takes eighteen months to reach genuine operational use because people are nervous about engaging with it honestly, you have lost the return on six months of potential value. That is not a soft cost.

Lower employee turnover during transformation

Digital transformation programmes are stressful. They change job roles, sometimes eliminate them, and require people to learn new skills in real time whilst still doing their existing jobs. Turnover spikes during transformation are common and expensive — the Chartered Institute of Personnel and Development estimates average replacement costs in the UK at between £3,000 and £30,000 per employee depending on seniority and specialism.

Organisations where leaders communicate honestly, acknowledge uncertainty, and create space for employees to voice concerns during transformation consistently see lower voluntary turnover during those periods. The maths is not complicated.

Faster identification of what isn't working

Perhaps the most undervalued financial benefit: when people feel safe to say a project is going wrong, you find out earlier. In my experience working with enterprise transformation programmes, the projects that run catastrophically over budget are almost never surprises at the project manager level. People knew. They just did not feel they could say so without career consequences.

A culture where bad news travels fast is worth more than any project management software on the market.

How does agentic AI change the psychological safety equation?

This is where it gets genuinely interesting, and genuinely complicated. Until recently, the psychological challenge of AI adoption was primarily about learning to use a tool. Now, with the rise of agentic AI — systems that do not just respond to prompts but autonomously plan and execute multi-step workflows — the challenge has shifted to something more fundamental: learning to trust a system that acts on your behalf without constant supervision.

An agentic AI system might autonomously draft and send communications, update records across platforms, or trigger operational processes. The human role shifts from executor to orchestrator — reviewing, approving, and occasionally overriding. This is a meaningful change in the nature of work, and it requires a different kind of psychological safety.

Employees need to feel safe asking: "What exactly did this agent do, and why?" They need to feel safe overriding an AI decision without it being framed as obstructing progress. And they need honest leadership acknowledgement that human judgement remains essential, particularly in high-stakes decisions, rather than the corporate messaging that implies AI is simply better and the human role is to get out of the way.

Leaders who get this wrong will find their agentic AI deployments either rubber-stamped without proper oversight (creating risk) or quietly circumvented by people who do not trust them (creating waste). Neither outcome is in the business case.

What does a growth mindset actually look like in practice during AI transformation?

The term "growth mindset" has been through the corporate motivation machine so many times it has largely lost its meaning. Let me be specific about what it looks like in the context of AI adoption, because the abstract version is not useful.

It looks like leadership being visibly uncertain

When a CDO or CTO stands in front of their organisation and says "we are implementing this AI platform and it will transform how we work" with complete confidence, they are — generously — overstating what they know. AI implementations are experimental by nature. Pretending otherwise does not inspire confidence; it erodes trust the moment the first thing goes wrong, which it will.

Leaders who say "we believe this will significantly improve how we do X, we are going to test it carefully, and we want to hear what you find" create an environment where honest feedback is possible. That is not weakness. That is the only approach that generates the information you actually need to make the implementation work.

It looks like celebrating useful failure

Not failure for its own sake — the Silicon Valley "fail fast" mantra has produced some genuinely expensive failures that could have been avoided with basic diligence. But celebrated learning from a contained experiment that did not work as expected is a powerful signal to an organisation about what is actually valued.

When a team runs a pilot, finds that the AI tool produces outputs that require too much human correction to be efficient, reports that clearly, and is thanked for the quality of their evaluation rather than blamed for the negative result — that is what a growth mindset looks like in practice. It is also how you avoid scaling something that does not work.

It looks like making the unofficial official

If your employees are already using AI tools — and they are — the psychologically safe response is to bring that into the open rather than pretend it is not happening. Create structured opportunities for people to share what they are experimenting with, what is working, and what concerns them. You will learn more about your organisation's actual AI readiness in one of those sessions than in any vendor-led capability assessment.

Four practical steps to de-stigmatise failure and accelerate genuine adoption

1. Establish a formal "what we learned" practice. Every AI pilot, however small, should conclude with a structured retrospective that is shared beyond the immediate team. Frame the output as organisational intelligence, not a performance review. The goal is to make it genuinely useful for the next team attempting something similar.
2. Create protected experimentation time. Asking people to explore new AI tools while simultaneously delivering their existing workload is a reliable way to ensure neither gets done properly. Even a modest allocation — a few hours a fortnight — signals that experimentation is a legitimate work activity, not something done on personal time.
3. Make escalation easy and consequence-free. If an employee encounters an AI output they believe is wrong, or a workflow they believe is creating risk, there must be a clear, simple, and genuinely safe route to raise it. This means not just having a process but actively demonstrating — repeatedly — that using it does not result in being labelled as obstructive.
4. Tie leadership communication to honesty, not confidence. Brief your senior leaders to communicate with calibrated uncertainty. "We are confident in the direction; we are still learning about the specifics" is more credible than false certainty, and it gives the organisation permission to be honest in return.

Psychological safety vs. standard change management: what's the difference?

What I have actually seen this look like in practice

I worked with an organisation — a mid-sized public sector body, since you ask — that had invested significantly in a document processing AI platform. Adoption at six months was, on paper, reasonable. Usage logs showed the tool was being accessed. The project was being reported as on track.

What was actually happening: a small number of highly conscientious employees were running everything through the tool, correcting the outputs manually, and then submitting the corrected version. The majority of the team had quietly concluded the tool was not fit for their specific document types and had reverted to their previous process without telling anyone, because the transformation programme had been presented as non-negotiable and they did not feel safe saying it was not working for them.

The platform was not fundamentally broken. It needed configuration adjustments for the specific document types this team handled. That was a three-week fix. It took seven months to surface because nobody felt safe raising it.

The cost of those seven months — in wasted licences, in staff time correcting outputs, in delayed efficiency gains — was considerably more than the cost of having created an environment where someone could have said "this isn't working for us" in week three.

That is the ROI of psychological safety. Not as an abstraction. As a line item.

Frequently asked questions

Is psychological safety just another way of saying "be nice to your team"?

No, and conflating the two causes real problems. Psychological safety does not mean avoiding difficult conversations, lowering standards, or tolerating poor performance. It means people can raise concerns, flag errors, and share honest assessments without fear of disproportionate consequences. High-performing teams are often direct and demanding — they are also psychologically safe, which is precisely what enables them to be direct without it becoming destructive.

How long does it take to build psychological safety in a transformation programme?

There is no honest universal answer, but the direction of travel matters more than the timeline. Consistent, visible behaviour from senior leaders — repeatedly demonstrating that honest feedback is welcomed and acted upon — produces measurable shifts in team behaviour within months. Sporadic gestures produce nothing. The mechanism is trust, and trust is built through repeated evidence, not announcements.

What if leadership is genuinely not open to hearing that the transformation isn't working?

Then you have a governance problem that sits above the psychological safety intervention. No amount of cultural work at the team level compensates for senior leaders who have tied their professional identity to a particular outcome and cannot hear contrary evidence. In that scenario, the honest advice — which I appreciate is uncomfortable — is that the transformation will fail, and the question is whether it fails quickly enough to preserve budget for something better, or slowly enough to consume the entire programme.

Can you measure psychological safety quantitatively?

Yes, imperfectly but usefully. Amy Edmondson's original research used a validated seven-item survey instrument. Organisations can adapt this to their context and track it over time as a leading indicator. Proxy metrics — voluntary turnover during change programmes, volume of issues raised through official channels versus discovered post-incident, time between problem occurrence and escalation — also provide useful signal without requiring a dedicated measurement framework.

Does this apply differently in the public sector versus commercial organisations?

The underlying psychology is identical. The specific pressures are different. Public sector employees often face additional constraints around job security, political accountability, and the consequences of visible failure that can amplify the barriers to psychological safety. This means the investment in building it is, if anything, more important — and more challenging. It also means that the cultural work cannot be separated from honest engagement with those structural realities.

What is the first thing a leader should do if they suspect their team is not being honest about how a transformation is going?

Stop asking in group settings. The social dynamics of a team meeting, particularly one attended by senior leaders, suppress honest feedback almost by design. Structured one-to-one conversations, anonymous feedback mechanisms, and small-group retrospectives with a neutral facilitator will give you a much more accurate picture of what is actually happening. What you hear may be uncomfortable. That discomfort is information, and it is considerably less expensive than discovering the same information eighteen months later.

Nicholas Hodder is a digital transformation and technology leadership advisor with over two decades of experience across commercial, public sector, and mission-driven organisations. He works with boards, CDOs, and senior leadership teams on the human side of AI adoption — which, in his experience, is most of it. If your transformation programme is producing confident status reports and quietly going nowhere, get in touch about a Leadership Workshop on Culture-First Transformation.