Most companies are now buying AI faster than they are getting anything back from it.
Gartner surveyed 1,303 leaders at companies turning over more than fifty million dollars, and only 22% had managed to scale AI across more than one part of the business. MIT looked at a few hundred enterprise deployments and found very few delivering anything measurable. Microsoft's latest Work Trend Index puts roughly two thirds of the difference down to organisational factors rather than individual skill.
Read those together and the conclusion is uncomfortable. The technology is not the hard part.
What is actually happening is subtler, and it is happening inside teams rather than inside systems. AI is quietly removing the small interactions that leadership used to run on. Nobody notices at the time, because each one looks like a saving.
Here are four of them.
1. Decisions arrive finished, so nobody sees the thinking
A manager used to reason out loud. They would sit with a problem, work through it badly for ten minutes, change their mind once, and land somewhere. Everyone in the room watched that happen.
That was the training. Not the answer, the working.
Now the answer arrives already formed. It is better structured than it would have been, it is faster, and it is often right. But the reasoning is invisible, so the people around that manager stop learning how the judgement was made. Three years later you have a team that can produce excellent documents and cannot tell you why one option beat another.
The fix is deliberate and slightly awkward. Say the reasoning out loud even when you did not need to. Tell people what you nearly did instead. Show the version you rejected. It costs five minutes and it is the only way the next layer learns to think rather than to prompt.
2. The hard conversation becomes an easy message
Giving someone difficult feedback is uncomfortable. That discomfort has a function. It is the thing that makes you pause and ask whether this should be a conversation rather than a message.
AI removes the discomfort. The awkward note is now well written, measured and fair in about nine seconds. So it gets sent, because sending it is suddenly easy.
The problem is that the person receiving it has no idea what it cost you to write, and that cost was part of the message. We have written before about what makes feedback actually land, and almost none of it survives being typed.
The fix is a rule rather than a judgement call. If the message is about someone's performance, their behaviour or their future, it is a conversation. Draft it if you like. Then go and say it.
3. Everybody's work looks the same, so you cannot see who is struggling
Managers used to spot trouble in the output. The report that came in late. The slide deck that was clearly rushed. The email with a tone that told you something was going on at home.
Those signals have largely gone. The person who is drowning now produces work that looks much like the person who is thriving, because both of them are using the same tool to polish it.
This is the one that worries us most, because it removes the early warning rather than the symptom. By the time you notice, it is usually a resignation rather than a conversation.
The fix is to stop reading the work as evidence and go back to reading the person. More one to ones, shorter. Ask what they are finding hard rather than how it is going. And pay attention to the things AI cannot smooth over, like whether someone has stopped speaking in meetings.
4. The five minute favour disappears
A lot of trust in teams was built in moments that had nothing to do with the task. Can I borrow you for five minutes? Does this read alright to you? What would you do here?
Those questions now go to a machine, because the machine is faster and it never sighs.
Each individual time that happens, it is a good decision. Cumulatively it removes most of the low stakes contact between people who need to trust each other when the stakes are high. This is the same mechanism we described in why company culture dies at every doubling, just arriving faster and for a different reason.
The fix is to put some of it back deliberately. Make asking a colleague the default for one category of decision, even where the tool would be quicker. The inefficiency is the point.
What this means for the people leading
None of this is an argument against the technology. It is very good, it is not going away, and the companies that use it well will beat the ones that do not.
But look again at what those four things have in common. Deciding what good looks like. Telling someone their work is not there yet. Noticing that a quiet person is not a happy one. Making it safe to ask a colleague something obvious.
They were always the hard parts of leading. They are now the parts that decide whether any of the investment pays off, and they are the parts nobody is training for.
Terry Russell has spent twenty five years in leadership teams doing this work. His view 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. AI has not changed that. It has just removed a lot of the interactions.
The companies getting this right are not the ones with the best tools. They are the ones where somebody noticed which human moments were disappearing and put them back on purpose. That starts with turning values into behaviours people can actually see, because a behaviour is something you can coach and a value is not.
Common questions
What can AI not do in a leadership role? It cannot decide what good looks like in your organisation, tell someone their work is not good enough, notice that a quiet team member is struggling, or make it safe for people to admit a mistake. All four require someone to carry a social cost, and carrying that cost is what makes them work.
Is AI bad for team culture? Not inherently. The risk is that it removes the small, low stakes interactions that trust is built from, because each individual removal looks like a saving. The damage is cumulative rather than sudden.
How should managers lead AI-augmented teams? Make reasoning visible rather than just sharing conclusions. Keep performance and behaviour conversations spoken. Read the person rather than the output. And deliberately protect some interactions that would be faster to automate.
The question worth taking into your next leadership meeting
Which of the moments that used to teach people how we do things here have quietly stopped happening?
If nobody around the table can answer that, it is worth half an hour of the agenda. The answer is usually two or three specific things, and they are usually fixable.
If two or three of them turn out to be real, that is the work, and it gets done with the leaders themselves.
Sources: sample sizes and dates are in the three linked reports. Gartner surveyed 1,303 leaders at organisations over $50m revenue between January and April 2026. Microsoft surveyed 20,000 knowledge workers across 10 markets. MIT NANDA interviewed 52 organisations, surveyed 153 senior leaders and analysed 300 publicly disclosed AI initiatives, and describes its own figures as directional.
