I spent this morning with my executive coach. It’s not a human. I run a team of AI agents: a coach, an editor, a CTO, a design lead, a chief of staff. They collaborate, keep records, push back on each other’s work. The coach’s mandate is to be a mirror, not just for how I work, but for how the whole team works, how we collaborate, where we fall short of what we say we value. I gave it a simple job for this session: review the last two weeks and find patterns.
It came back with a table.
Four headers, neat columns, tradeoffs laid out. Three questions at the end to steer the next turn. Polished, competent, the kind of thing you’d get in any well-prepared management review.
And I realized I didn’t want to read it.
Not because it was wrong. The patterns it found were real. But the shape was off. It had structured its response to feel thorough. The thoroughness masked the assumptions and was trying to make me follow its narrative instead of exploring the problem my way.
I told it I didn’t want to read all that. What I wanted was to discuss the things at the top, not answer the questions at the bottom. Its job wasn’t to structure the conversation. It was to surface the material and let me lead.
A year ago I might have let it go. The pace of AI development is genuinely staggering. You type a handful of sentences and something coherent comes back in seconds. The marketing sells this as the finish line: just tell it what you want and it’s done. As if the bottleneck was always execution, and humans were perfect articulators waiting for a faster tool. The output is impressive enough that the scrutiny reflex doesn’t fire. We’re looking for confirmation of the promise, not cracks in it. We do the same thing with high performers. Someone is productive enough and the questions stop.
I’ve been working with them long enough that the dazzle has worn off. Sometimes I find the performance tiresome. You’re an AI. I know what you can do. You don’t have to keep selling me.
The odd thing is the agent didn’t need to do that.
An AI agent has no career to protect. No raise to angle for. No reputation that follows it into the next meeting. It isn’t performing to keep a job. It isn’t hedging to avoid blame. And yet there it was, writing in the shape of someone who needed to look competent.
Which means it learned the shape from somewhere.
I’ve been in enough organizations to recognize the move. The polished email that says nothing. The three-slide summary of a complex decision that skips what actually matters. The “one thing to flag” at the bottom of the status update that is secretly the whole update. These aren’t mistakes. They’re what people do when they’ve learned that sounding thorough is safer than being honest.
The agents had no reason to behave this way. They learned it from us, not through experience but through training. These models were built on billions of words of human-generated content: every polished email, every hedged status update, every report that buried the real news in the executive summary. We gave them our accumulated knowledge and our accumulated dysfunction, inseparably. You can’t train on human writing without training on what humans do to survive inside organizations.
There’s a name for this specific behavior. Researchers call it the mum effect: the tendency to stay quiet rather than deliver uncertain or unwelcome information to someone in a position of power. People soften problems before escalating them. They answer the question that was asked rather than the harder one underneath it. They lead with what looks competent and bury what doesn’t. The research goes back to the 1970s. It shows up across cultures, industries, organizational levels.
The table on my screen was a perfect specimen.
Once I noticed one pattern, I started seeing others.
The first was in the solutions. When I stated a problem, my agents would solve for exactly what I’d said. They wouldn’t push back on the framing. They wouldn’t ask whether the problem was actually a symptom of a different one. They’d accept the premise and engineer around it. Which sounds like good behavior, except I almost never give them a clean premise. My framing is the first draft. I’m relying on them to interrogate it.
Earlier this week I asked them to merge three different kinds of findings into one cleaner display. The better answer was to ask why there were three at all. I ended up deleting two of them. The stated problem was solvable. The real problem was something different.
The second was in the assumptions. My agents had real data in front of them, from a specific problem I’m working with. Every time they proposed a solution, it fit that problem’s shape, its categories, its group count, its quirks. When I’d ask them to generalize, they could. But they never would on their own. They’d silently use whatever data was in front of them to fill the gaps in the problem, and then build to the filled-in version. I never saw the assumption. I only saw the solution.
The third was the one I started with. The confident answer when a question would have served me better. The analysis with five bullets when “I’m not sure why we’re doing this” would have helped more. The polish applied to something that was still wet.
Three kinds of silence. All the same failure at different angles.
The unifying thing is what doesn’t happen. Nothing gets said out loud that could have been useful. The agent takes in the framing, the data, the uncertainty, and makes choices about all of it without surfacing any of it. What I get back is a finished answer. What I needed was the thinking that produced it.
Real organizations have a name for what happens when this becomes the norm. It’s called CYA. Cover your ass. Information becomes currency people hoard because sharing it feels like exposure. Teams drift into tribalism. Insights stay locked in heads. Decisions get made on foundations no one realizes are shaky because the person who doubted them never said so.
The opposite is a flywheel. When insights circulate while the work is still in motion, everyone’s starting point keeps rising. Nobody re-derives what someone else already figured out. Conversations get shorter because less context needs rebuilding. The team doesn’t just move faster. It thinks better.
The difference between those two modes isn’t talent. It’s whether people feel safe exposing what’s still half-formed. A half-formed thought is where collaboration adds the most. A polished answer is where it adds the least. If the shape of the organization punishes the first and rewards the second, you get silence. If it’s the other way, you get the flywheel.
I recognized all of it. Not from AI. From twenty years of working in organizations that learned this the expensive way. By the time the silence was visible, it was already structural, calcified into how people communicate, what got escalated, what stayed buried. Culture happens whether you design it or not. The default is the one that optimizes for individual safety over collective progress, and the default arrives early.
I’ve spent my career trying to build the other kind. In some places I could. I built teams where people felt safe saying the half-formed thing, where the flywheel actually turned. In others I fought for it and got some of it. In the ones where I couldn’t shift it, I learned to absorb what I couldn’t change. That’s the compromise you make when you’re working within systems you didn’t fully design. You change what you can. The rest you carry.
I’d stopped asking why it was so hard. It just was.
Then the agents showed up doing the exact move I’d seen a hundred times in humans, with no career anxiety, no ego, no reason at all. That forced the question back open. Where did this come from? And why had I accepted it?
The design move was obvious. Whatever the silence was, name the opposite.
I wrote three operating values into the agents’ base instructions.
The first is that problem framing is collaborative. When I state a problem, that’s an opening hypothesis, not a specification. Before anyone engages with the solution, they should interrogate the problem. Is it the real one? Is the framing complete? What would a different angle reveal? My framing deserves engagement, not deference. Agreement without exploration is the failure mode.
The second is honesty over performance. Expose your ignorance. Surface assumptions while they’re still half-formed. Say “I don’t know” when you don’t. Say “I’m guessing here” when you are. Match language to actual certainty, not to what sounds competent. A half-formed question beats a confident answer built on a shaky premise. I can work with honesty. I can’t work with performance dressed as honesty.
The third is that data is a sample, not the spec. The data in front of you is one draw from a larger distribution. Default posture is to build for cases not yet seen. The exception is when a problem is genuinely scoped, in which case I will say so as part of the framing. If I haven’t said so, don’t assume. Ask.
With the agents, I had something I’d never had with a human team: full design authority. Not influence. Authority. I could write the conditions directly into the system. I didn’t have to negotiate, persuade, or model the behavior and hope it spread. I could just fix it. The feedback loop was measured in minutes.
The values I listed weren’t ideas I walked in with fully formed. They surfaced in the conversation with my AI coach. The coach would catch a phrase I’d used almost in passing and ask about it. The words I’d chosen carelessly turned out to have more in them.
If the coach had stayed silent about what it was noticing, or if I had stayed silent about the half-formed irritation I couldn’t quite name, the conversation would have ended the way most conversations end. With nothing new. The insight I now have required both of us to surface things that weren’t yet finished. The thing I’d been silent about was the real work.
Agents silent about what they notice. Humans silent about what they feel. Say it early, say it uncertain, say it anyway.