Can AI Carry Human Conflict?
Jack Dorsey and Roelof Botha published an essay called “From Hierarchy to Intelligence” at the end of March, and the part I find most interesting is not the headline claim about middle management going away. It is a smaller passage near the end, where they describe what humans must still do in their AI-coordinated company.
They write: “The people are on the edge. The edge is where the action is. The edge is where the intelligence makes contact with reality. People reach into places the model can’t go yet. They sense things the model can’t perceive: intuition, opinionated direction, cultural context, trust dynamics, the feeling in a room. They make the calls the model shouldn’t make on its own, especially ethical decisions, novel situations, and high-stakes moments where the cost of being wrong is existential.”
Read that paragraph slowly. They are listing what humans must do that AI cannot: trust dynamics, the feeling in a room, ethical calls when being wrong is existential. They are clear that humans at the edge must do this work. They do not address how humans get good at it.
That gap is what I want to think through here.
Two jobs, not one
Their argument is that hierarchy existed for two thousand years to route information through humans. AI can now do that routing, so the layers can go. They are not lazy about it. They go all the way back to the Roman army to show why hierarchy made sense in the first place. They walk you through the math of how many people one person can manage, from the Roman contubernium to the Prussian general staff to the American railroads to Frederick Taylor’s factories to the modern matrix corporation. The Roman soldier needed a decanus because eight men needed a leader. The same math, escalated, eventually shaped every modern corporation.
On the information routing, they are right. What I notice is that hierarchy was always doing two jobs at once, not one.
The first was routing information.
The second was navigating the conflict that peers could not navigate themselves.
Middle managers are paid full-time for that second job. They decide when their teams cannot agree, mediate when there is room to negotiate, deliver hard messages no one else wants to deliver, and carry the residual tension while the team keeps working. Some of this gets named in job descriptions. Most of it never gets measured. It is the day-to-day work of management as I have observed it across executive teams in Ukraine, China, the US, and France for twenty-five years.
Without that layer, the conflict does not disappear. It lands directly on peers. And in most companies, peers have never been trained for it.
Why do people accept a manager’s call?
There is a question I have been turning over since the Block essay came out. When two engineers cannot agree, and the manager rules one way, the engineer who lost the call accepts it. Often unhappily. The acceptance still holds.
What is doing the work in that moment?
It is not that the manager is necessarily right more often than the peer would be. In my experience, managers are about as right as anyone else. The acceptance is doing something else. It has more to do with who is making the call than what the call says.
Frances Frei, who was brought into Uber after the Kalanick crisis specifically to do trust work, has a useful frame for this. She calls it the trust triangle. People trust someone when they believe they are getting authenticity (the real them), logic (sound judgment), and empathy (real care that you are okay). When trust breaks, you can almost always trace it to one of those three wobbling. AI can simulate logic. It struggles with authenticity, because there is no “real you” inside the model, and it does not have empathy in the way Frei means, where you believe someone cares whether you are okay.
Patrick Lencioni, in The Five Dysfunctions of a Team, makes a sharper version of the same point. He talks about two kinds of trust. There is predictive trust, where you know what someone will do. And there is vulnerability-based trust, where you know they have shown you their fallibility, and you have shown them yours, and neither of you used it as a weapon. Predictive trust is what you have with reliable software. Vulnerability-based trust is what teams need to function under pressure. AI sits squarely on one side of that line. It can be remarkably reliable, and it cannot be vulnerable at all.
David Maister’s trust equation does the math another way. Trust equals credibility plus reliability plus intimacy, divided by self-orientation. Every variable except reliability is built between people. AI can score perfectly on reliability and zero on intimacy. And it has no self-orientation, because there is no self there at all.
What managers carry is the cost of having made the call, and AI does not carry that cost.
Can AI earn that kind of trust?
This is where I find myself most uncertain. What I notice is that people can trust AI in some ways. Edelman’s 2025 trust barometer found that trust in AI runs from 87 percent in China to 32 percent in the US. Some of that is exposure, some generational, some task-specific. People might trust AI to recommend a movie and not trust it to fire someone they care about.
The philosophers ask the question in different ways. The one that interests me most is Mark Ryan. Writing in 2020, Ryan argues that AI cannot really be trusted the way humans trust each other, because it does not have feelings or carry responsibility, and it has no skin in the game. What we feel toward AI, he says, is reliance. We rely on it the way we rely on a reliable car. Calling that trust makes AI sound more human than it is, and lets the people who built it off the hook.
I am not sure he is fully right. The question I find myself asking is this. When an AI delivers a hard message, that you did not get the promotion, your project is being shut down, your colleague’s complaint about you was substantiated, is the person receiving it accepting the verdict, or accepting that they have no other option? That is a very different kind of acceptance, and it is the kind that builds resentment over time rather than resolution.
What happens when companies actually try
We have some real cases to look at now.
Klarna told the world in 2024 that their AI agent had replaced 700 customer service representatives. By May 2025, CEO Sebastian Siemiatkowski was reversing course. He told Bloomberg that “as cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality.” Klarna is hiring humans back. The AI could handle volume, but not the relational work that customer service actually is.
Amazon’s algorithmic management of Flex drivers is darker. Workers have been fired by email with no human contact, with ten days to appeal to another bot. Stephen Normandin, a 63-year-old Army veteran in Phoenix, was terminated this way. France’s CNIL fined Amazon’s French warehouse subsidiary 32 million euros in 2024 for excessively intrusive surveillance. The algorithm was technically right by its own metrics. What it left behind was bad enough to bring fines and union pushback.
The OpenAI board fired Sam Altman in November 2023 because, in their words, he was not consistently candid in his communications with the board. That is the most consequential governance moment we have seen in AI, and it was made by humans on a human read of whether they could trust him. I doubt any algorithm could have made that call, at least not in a way the room would have accepted. And then it was reversed within five days by another deeply human moment: 700 employees threatening to leave.
What I notice across these is that the more human the stake, the harder it was to live with what the AI decided. Klarna found AI could partly mediate its customer relationships. Amazon’s employment cases could not be handled ethically by AI, and the OpenAI board’s governance call could only be made by humans, who then had to come back to each other to figure out what they had done.
The case that should give the AI optimists pause
The closest example to what Block is attempting is not in Silicon Valley. It is Buurtzorg, the Dutch home-care organization that scaled from four nurses to fourteen thousand employees with no middle managers.
Buurtzorg has self-managing teams of about twelve nurses each. They make their own scheduling, hiring, and day-to-day decisions. They do not have managers above them.
What they do have is conflict resolution training in Nonviolent Communication for every nurse. And when a team cannot resolve a conflict on their own, they call in one of about twenty regional coaches, each supporting forty to fifty teams. The coach is not a manager. The coach is a trained human mediator who shows up in person.
The most successful flat-organization story we have kept a layer of humans whose job is conflict. They just renamed the role. The need did not go away when the org chart got flatter.
Remote-first makes this harder still
There is one more piece of the Block design I find difficult to think past. Block is remote-first by design. The world model is built from the artifacts of remote work: code commits, decision logs, performance metrics. That is exactly the kind of data that lets AI take over coordination.
It is also exactly the kind of work that makes peer trust hardest to build.
Trust between peers grows through proximity. It grows in hallway conversations and in the moments you watch a colleague handle a hard situation in real time. It grows when you are in the room as something difficult happens, and you see how someone carries themselves through it. Without those moments, peers have fewer chances to build the trust that handling conflict together actually requires.
Amy Edmondson, who is among the most cited researchers on team trust in the world, published a paper with Jayshree Seth in Harvard Business Review in February 2026, just weeks after the Block essay. They titled it “How to Foster Psychological Safety When AI Erodes Trust on Your Team.” What they have found is this. When companies bring AI in, team performance can decline even when the productivity tools work as advertised. People start second-guessing themselves and trust erodes in ways that are hard to pinpoint. The patterns they describe look like the same team problems we have been seeing for decades.
That is a serious researcher publishing in real time that the substitution Dorsey and Botha are betting on is not working as smoothly as the Block essay suggests.
The function I have not named yet
There is one more thing managers were doing that I have not named. They returned. After every disagreement, every hard call, every meeting that had not gone well, they came back the next day and asked. They repaired what could be repaired and held the relationship together.
AI is not absent from our work lives. It is constant, and yet it does not return in the way that matters. There is no ongoing relationship for it to come back to.
This is where my own work continues. The Loving Conflict framework I write about and bring into companies is built around the idea that conflict, when you stay with it, can deepen connection rather than damage it. Repair after rupture is not optional. It is the practice that makes everything else possible.
I am not sure that practice is something we can hand to a system.
What I am left with
I want to leave a few questions here and hopefully explore them together with you.
If hierarchy was doing two jobs, and we have only solved one, what happens to the other?
If trust between humans is built on vulnerability and skin in the game, can AI build a different kind of trust that serves us in conflict, or only in coordination?
What is the new role we have not yet named? It is not the old middle manager, and it is not the pure individual contributor either. It is someone whose job is to hold and repair the conflict the AI cannot.
I am bringing this work into more companies this year because the answers are going to matter sooner than most of us are ready for. The companies that grow the capacity to handle human conflict fast will be ready for what Block is betting on. The ones that do not will discover that flatness without skill leaves you with a stuck organization, not a flatter one.
I would love to hear what you think. Where is this showing up in your own work? And what do you think AI can do, and cannot do, with human conflict?