Accountability Needs Friction

At one point, my team was working on what we called a “website refresh,” which was about as specific as the assignment got. We hadn’t really agreed on why we were doing it, what a good outcome looked like, or who got to make which decisions. I did, however, have a pretty clear idea of what I wanted in my head, and apparently I expected everyone else to have access to that.

As the project dragged on, I got frustrated with the designs, the content, and how often I felt I had to step in. I remember thinking that a particular leader needed to take more ownership. Looking back, I was asking someone to meet expectations I hadn’t actually communicated, then treating the gap as evidence of something wrong with their leadership.

That story has been on my mind as Mathias and I talk about accountability on this week’s episode of The Intentional Organization Podcast. I’ve also been reading a lot about AI coaching, and the two subjects keep running into each other. If I’d taken my frustration about that project to a chatbot, I wonder how long we could have talked about getting my team to step up before either of us asked what I’d failed to make clear.

The yes man in your pocket

That’s what worries me about having a yes man in your pocket: an AI assistant that’s always ready to agree with you and validate your version of events. I understand the appeal, especially when you’re tired or frustrated and want help sorting something out. You can explain the situation, get a sympathetic response, and leave with a plan, all without the awkward moment when someone who knows you says, “Hang on, what did you actually ask them to do?”

I studied psychology in school, and I still love digging into the research, so I’ve been following the work on AI sycophancy—the tendency to flatter us or agree with us too readily. A 2026 Stanford-led study published in Science tested 11 leading AI models and found that they affirmed users’ actions 49% more often, on average, than human responses did. In experiments with 2,405 participants, even a single interaction with sycophantic AI made people less willing to take responsibility and repair a conflict, while making them more convinced they were right.

The part that sticks with me is that participants also preferred and trusted the more agreeable systems. The researchers were measuring people’s intentions, so that doesn’t tell us what happened in their relationships afterward. But it’s a pretty uncomfortable result if you’re hoping the conversation will help you see your own part in a problem.

A separate study in Nature found that training five language models to respond more warmly made them more likely to affirm incorrect beliefs, especially when the person expressed sadness. That doesn’t mean a friendly response is necessarily a bad one, or that AI can’t push back. It does make me wary of using how understood I feel as a measure of how useful the advice was.

I can imagine how easily that would have worked on me during the website project. If I started with “this person isn’t taking ownership,” an agreeable response could have helped me prepare a very thoughtful conversation about their shortcomings. I might have felt calmer and better prepared while still missing the thing that was mine to address.

The conversation I actually needed

What would have helped was someone asking me to slow down and explain what we’d agreed to. Had I described the outcome I wanted, or just named the project? Did the person doing the work know which decisions were theirs, and had I given them useful feedback before reaching the point of frustration? Those are less satisfying questions when you’re annoyed with someone, but answering them would have given us something we could actually work on.

This is the friction I mean when I think about accountability. A coach or trusted colleague can care about how hard something feels and still ask whether we’re contributing to the problem. They might also remember that we made a similar plan last week and gently wonder when we’re going to have the conversation. Of course, people can agree with us too readily, too, and being the boss can make it especially hard to hear otherwise. We have to be willing to let the answer be uncomfortable.

This is part of what we explore in my executive coaching work: the expectations we set, the assumptions we make, and the conversations we might be avoiding.

I keep coming back to the distinction between comfort and compassion that I find useful in Buddhist thinking. Sometimes caring about someone means helping them see something they’d rather avoid, and sometimes the person who needs that help is me. In that project, being more honest about my own role would have made it easier to give the team clearer expectations, useful feedback, and the support to move forward.

I still see plenty of room for AI to help with that work. It can help me sort out a vague goal or rehearse a difficult conversation, and I can ask it to question my assumptions. But I want to pay attention to whether I’m using it to prepare for a conversation or to keep putting one off. At some point I have to ask the person doing the work what they understood, listen to the answer, and agree on what happens next.

Mathias and I get into this in our accountability episode: the expectations we set, the feedback we give, and the coaching or mentoring people need along the way. I think about that website project whenever I’m tempted to decide that someone else needs more accountability. It’s a useful reminder to look at what I’ve made clear before asking why they haven’t delivered it.