Psychological safety is the real AI adoption bottleneck (not another Copilot license)

A team can have the licence, the prompt sheet and a channel full of impressive screenshots, and still go quiet when a manager asks who has changed a workflow. That quiet is worth taking seriously. People have usually learned that a rough draft, a wrong number or a killed experiment can cost them, so they wait until the work looks finished. Another seat on the same product does not change that calculation.

Amy Edmondson, the Harvard professor who put psychological safety into the management literature, describes it as a shared belief that the team is safe for interpersonal risk. Google's Project Aristotle later found the same pattern in its own teams. People speak up, admit a miss and try a half formed idea when they believe they will not be punished for it. AI adoption is full of those risks, which is why another seat on a copilot product does not move a quiet room.

Why the tool makes the room quieter

Using a model in real work asks a person to do three socially expensive things. They show a draft that might be naive. They admit they do not know whether the answer is right. They challenge a colleague's output, sometimes a senior colleague's output, in a meeting where speed is being praised.

If the last person who showed unfinished work got a joke at their expense, they will not show the next draft. If the last person who said I think this number is invented got labelled as negative, they will paste the number. If the only stories in the channel are flawless wins, everyone else learns to keep their experiments off the record. The licence does not fix any of that. It gives people a private place to hide the experiment, which is how you get shadow use and a public story that adoption is low.

Safety here is not niceness, and it is not a lower bar. Edmondson's point, and the one we use when we work with teams, is that high safety with low standards is comfort. High standards with low safety is anxiety. The useful zone is both. People can say the output is weak, and they are still expected to make it good.

What leaders do in the first month

Leaders set the tone in small, repeated moves. A town hall announcement that it is safe to experiment is not one of them. The moves are local.

1. Show your own rough draft, including one place the model was wrong, before you ask the team to show theirs.
2. Ask what the model missed, in the same meeting where you ask what it sped up, so a critique is part of the update and not a confession.
3. Respond to a caught error with a question about the check, not with a speech about carelessness.
4. Keep a visible example of a stopped workflow, so killing a bad use is as respectable as launching one.
5. Separate the experiment from the performance review for a defined period, and say the dates out loud.

The last one is the one people listen for. If a clumsy first month can show up as a lack of capability at review time, rational people will wait until they are fluent in private. You will then conclude that the team is slow, and buy more training. They were not slow. They were managing risk.

What L&D can actually run

A learning team cannot install safety by adding a module called trust. You can, though, design the AI sessions so the social risk is practised, not preached.

Put intact teams in the room with their manager, not a random slice of the company who will never see each other again. Use their work. Require every person to bring one output they would not yet send. Spend more time on the critique than on the prompt that produced it. End with a norm the team writes themselves: how we show unfinished work, how we challenge a number, how we admit a miss. Three lines is enough. A culture deck is not.

The psychological safety workshop exists for the team habit underneath the tools. It is the companion to AI training, not a substitute, and it is wasted if leaders send the team and stay outside. The person with the most status has to do the first risky thing.

Signals that are better than licence counts

Stop reporting adoption as the number of people who logged in. Logins measure access. Ask instead whether people will put a weak draft on the table, whether a junior can challenge a senior's AI assisted slide, and whether a killed experiment gets airtime. Those are observable in a fortnight if a manager is looking.

A simple pulse is enough. Four questions, anonymous, asked before an AI push and again six weeks later. Can I show unfinished work here? Can I say I do not know if this answer is right? Will I be blamed for a good faith miss? Do we talk about what we stopped? If those scores are low, more seats will not help. Fix the room, then teach the workflow.

Keep a hard edge on the standard. People should be able to show unfinished work and to say an output is weak, and they are still expected to make it good enough to send. Safety is not permission to ship an unchecked draft, and it is not permission to ignore the new way of working. If a thoughtful attempt is punished, the next attempt happens in private. If the standard is dropped, the team gets faster at sending work it has not checked. The useful room does both: the argument about the draft happens where the manager can hear it, and the customer still gets a checked answer.