Almost every large organisation has now put its managers through something on AI. Licences, a platform, a session on prompting, a policy document with a section about what should not go into a public model.
That training answered a question about tools. The questions managers are actually stuck on are different, and none of them is technical.
EY surveyed more than 1,100 US desk workers at companies with over a billion dollars in revenue, asking about agentic AI specifically. 53% of people managers said they worry they may not be good at supervising AI-augmented teams. 82% expected managing AI agents to make their role harder. The third figure is the one worth sitting with: 63% of non-managers said they are now hesitant to step into management at all because of it.
Read those together and you have a capability problem arriving at the same time as a pipeline problem. The people doing the job are unsure, and the people who would replace them are watching and deciding against it.
What follows is what we think they are actually short of, based on the work rather than the survey.
1. A view on what AI is for here
Ask a manager what their team is allowed to use AI for and most will give you a version of "use your judgement". Ask them what they would say if someone sent a client a document the model mostly wrote, and the answer gets vaguer.
That vagueness travels. In the absence of a position from their manager, people invent one, and they invent it privately, which means you end up with six different standards inside one team and nobody comparing notes.
What helps is narrower than a policy. It is a manager willing to say out loud: here is what I am happy for you to use it on, here is where I want a human to have thought about it first, and here is what I want you to tell me about. Three sentences. They can be wrong and revised later, which is considerably better than absent.
2. The ability to judge work they did not watch being made
Managers used to assess work partly by watching it happen. You saw the thinking, the drafts, the thing somebody got stuck on at four o'clock on Tuesday.
A finished piece of work now arrives without any of that visible. It is well structured, it is plausible, and the manager has to form a view on whether it is right, which is a harder skill than most of us admit.
This is judgement rather than knowledge, and it is learnable. It looks like asking what the person considered and rejected. It looks like being able to say this is good enough to send and this is not, and explaining which. We wrote about the same muscle in the context of feedback that actually lands, because it is the same muscle.
3. A way to see who is struggling
This is the one that worries us most, and we covered it at length in what AI can't do.
Managers used to spot trouble in the output. The late report, the rushed deck, the email with a tone that said something was going on. Those signals have largely gone, because the person who is drowning and the person who is thriving are both polishing their work with the same tool.
The replacement is duller and it works: more one to ones, shorter, and questions about the person rather than the status. What are you finding hard at the moment is a different question from how is it going, and it gets a different answer.
4. Permission to protect the slow bits
A lot of trust in a team was built in exchanges that had nothing to do with the task. Can I borrow you for five minutes. Does this read alright. What would you do here.
Those questions now go to a machine, because the machine is faster and never sighs. Each individual time, that is a sensible decision. Cumulatively it removes most of the low stakes contact between people who need to trust each other when the stakes are high.
A manager can put some of it back, and most will only do so if they believe the organisation will back them for being slower on purpose. That permission has to come from above them.
Why the tools training did not touch any of this
Microsoft's 2026 Work Trend Index, which surveyed 20,000 knowledge workers across ten markets, attributes roughly two thirds of AI impact to organisational factors: culture, manager support and talent practices. Individual skill and mindset account for about a third.
Almost all of the money has gone into the third. Very little has gone into the two thirds, and the two thirds are where the manager sits.
That gap also explains the finding we wrote about in why leaders overestimate their teams. Executives answer a question about what has been provided. Employees answer a question about what happened the last time something went wrong. Both are honest, and only one of them is describing behaviour.
What to do in the next ninety days
Three things, in order, none of which needs a new platform.
Ask every manager to write their three sentences, the ones in section one, and share them with their team. It takes twenty minutes and it is the single highest-return thing on this list, because it converts a vague anxiety into a stated position that can then be argued with and improved.
Change what gets reviewed in one to ones for a quarter. Fewer status updates, one standing question about what they have found hard. Managers will need to be told this is a legitimate use of the time.
Then at ninety days, ask the people around those managers what is different. Not the managers themselves, and not a completion rate. The people who report to them. If nobody can name anything, nothing changed, whatever the dashboard says.
Terry Russell's line on this is that culture gets built through thousands of small interactions, and that in each one leaders teach people three things: what matters here, what gets rewarded here, and what gets tolerated here. A manager with no view on AI is still teaching all three. They are teaching that nobody knows, that it does not matter much, and that you are on your own with it.
Common questions
What skills do managers need to lead AI-augmented teams? Judgement rather than technical fluency. Being able to state what AI is for in their team, assess work they did not watch being produced, notice when someone is struggling without the usual signals, and protect some of the slower human interactions that trust depends on.
Is tools training a waste of time? It’s both necessary and insufficient on its own. People do need to know how to use what they have been given. The gap is that almost nothing has been built on top of it for the people leading the work.
How do we know whether manager development has worked? Ask the people who report to those managers what is different ninety days later. If they cannot name a specific change, the programme changed nothing, whatever the completion rate says.
The question worth taking into your next leadership meeting
If we asked every manager in this business to write down, in three sentences, what AI is for in their team, how many different answers would we get?
The number itself is the finding. A business where everybody writes roughly the same three sentences has a culture. A business where everybody writes something different has a policy document.
Run it before you commission anything else. It costs twenty minutes and it will tell you more about where your managers are than another round of tools training will. If the answers worry you, come and talk to us about what to do with them.
Sources: sample sizes and dates are in the linked reports. EY's Agentic AI in the Workplace survey, October 2025, covered more than 1,100 US desk workers at companies with revenues above one billion dollars, and the figures above relate to agentic AI specifically rather than AI generally. Microsoft surveyed 20,000 knowledge workers across ten markets with Edelman Data x Intelligence for the 2026 Work Trend Index.
