When I started thinking seriously about safely scaling AI, I assumed the focus areas would be familiar ones: governance, architecture, security, data and process. I still believe all of those matter. What has changed is my view of what underpins these.
My hypothesis is that psychological safety is the foundation. Not because AI requires everyone to agree. Not because standards should be lowered. Psychological safety, as Amy Edmondson describes it, is about feeling safe to speak with candour. Accountability and performance management still matter. In fact, I suspect they will matter more.
Agentic AI asks an organisation to become unusually honest about itself. People need to say that a process does not work the way the procedure suggests. A leader may need to admit they cannot describe how their process performs. A subject matter expert must be able to challenge an agent’s output that appears convincing but feels wrong. Someone whose role is changing needs room to describe what that change means for their confidence, status and professional identity.
None of this is comfortable. AI will make work more visible. It will expose hand-offs, delays, weak measures, undocumented judgement and unclear ownership. That transparency can create learning, but it can also create defensiveness. If people believe candour will be punished, they will protect themselves. They will quietly correct poor outputs, hide failures, avoid experimentation, and defend the current process even when they know it is broken.
I am not yet presenting psychological safety as a proven formula for AI value. It is a hypothesis I want to test, supported by emerging research into psychosocial safety climate, innovation capacity, and attitudes towards change. My experience is increasingly pointing in the same direction.
A psychologically safe organisation is not protected from difficult decisions. Roles will continue to evolve. Expectations will rise. Some work will disappear and new work will emerge, as it always has. The difference is whether people can speak honestly while that happens.
The more capable AI becomes, the more candour the organisation will need. That is why I am beginning to see psychological safety not as the people component of an AI programme, but its underlying foundation that everything else is built on.