I Was Told Off For Being Rude To My AI Agents
- Merve Kagitci Hokamp
- Aug 17
- 11 min read
Someone called me out recently for how I talk to my AI agents. Apparently, I was being rude.
My response was something along the lines of:
It's software. It'll be fine.
But then I did what I usually do when someone tells me something I don't particularly want to hear. I looked into it. And I ended up somewhere much more interesting than the question of whether I should say please and thank you to Claude.
The genuinely useful question, in my opinion, is what happens to our judgment when AI becomes a thinking partner, an assistant, a researcher, a writer, a project manager and, increasingly, an agent that can actually do things for us.
Does being polite to AI actually make it better?
Initially I really didn't think I was being mean. It's not like I was cursing at it. But as I sat with the feedback, I realized the person who gave it to me was right. I would not use the tone I use with my AI agents on an actual human being. That gap is worth paying attention to, so I went looking at what the research and the general advice actually say.
The first thing to name is that none of this is new. We already grade our manners by how visible the other person is. When I managed a customer service team at Google, we watched this happen daily. Put a customer on live chat with someone they cannot see, and the tone gets sharp fast. Move the same conversation to the phone and it softens. Get people face to face and it is a different game entirely. The distance from another person's face has always set the ceiling on how badly we are willing to behave. Psychologists call it the online disinhibition effect, and every contact center in the world has a version of the same chart.
So where does an AI agent sit on that gradient? On paper it should be the far end. No face, no voice, no feelings, no colleague to gossip about you afterward. Maximum disinhibition, zero social cost.
Except we don't seem to treat it that way. Reeves and Nass established decades ago that people apply social rules to computers without meaning to, even when they know perfectly well they are talking to a machine. A 2025 survey of over a thousand people in the US and UK found more than 55% say they consistently use polite language with AI, up from 49% the year before. The more human the thing sounds, the harder our instincts pull. So we have built something that is socially real enough to trigger our manners, with none of the consequences that normally enforce them.
The practical camp argues you should be polite because it produces better output. There is evidence for it. A 2024 cross-lingual study across English, Chinese, and Japanese found that impolite prompts often produced worse performance, though excessive politeness didn't help either, and the optimal level varied by language.
Then the picture flipped. A 2025 Penn State study tested 50 questions across five tones, from very polite to very rude, and found the opposite: accuracy climbed from 80.8% on the very polite prompts to 84.8% on the very rude ones. That one is a preprint rather than peer-reviewed research, so I wouldn't build a leadership philosophy on it. But between the two, one thing is clear. There is no magic please-and-thank-you setting that reliably makes AI smarter.
The second camp worries about behavioral drift, the idea that rehearsing incivility in a consequence-free environment loosens it everywhere else. That's also unsettled. A BYU study of young adults found no meaningful spillover into human interactions, and the researchers had expected to find the opposite. Their read was that adults don't personify digital assistants enough for it to transfer, though they suspected a study with children would come out differently.
Both camps are trying to justify manners by their downstream effects, and on the evidence available, neither case carries. Which is roughly the point at which I stopped asking the politeness question and started asking a better one.
Your AI conversations aren't as private as you think
When you use AI agents inside an organization, other people are watching. Maybe they aren't reading every prompt. But they can see what the agent produces, what gets actioned, what doesn't, and how you respond when something goes wrong.
Imagine a senior leader working with an agent on a project. The agent produces something mediocre. The leader responds:
"This is useless. Try again."
"How many times do I have to explain this?"
"Are you actually capable of doing this?"
The AI doesn't have feelings. That's not my concern. My concern is the human being watching the exchange, and what they now know about how this leader handles mistakes. How quickly they blame. Whether they check their own instructions before blaming the output. Whether they take responsibility when the brief was unclear.
Those are leadership behaviors, and they are being demonstrated whether or not the thing on the other side of the conversation is a machine. If you tell your people to embrace AI while behaving like an angry manager every time an agent gets something wrong, you're still teaching them your leadership style. The audience has just changed.
"I need to fire my AI agents"
One of my clients came into a coaching session recently and announced: "My AI agents are driving me crazy. I need to fire them and hire new ones."
That got my attention.
Her agent would respond, "Yes, absolutely. Perfect. I'll go ahead and do that." And then it wouldn't. Or it would do half the task and confidently announce: "Done!" Except it wasn't done, and not in some obscure, technically complicated way. She could look at the output and immediately see what was missing.
So she would check the work, find the gap, go back to the agent, explain it again, wait, check again, and repeat. At one point she asked the agent, "Are you a joke?"
I understood the frustration. If your AI assistant requires you to repeatedly inspect, correct, remind, chase and re-explain every task, you have created another thing to manage. The agent was very good at saying yes. It was much less reliable at knowing whether the work had actually met the standard she had in her head.
"Yes" is not the same as "done."
The delegation problem underneath it
We're importing assumptions about human delegation into systems that work very differently.
When I delegate to a good human colleague, they push back.
What exactly do you mean?
What's the deadline?
What does good look like?
Do you want me to decide, or bring you options?
Who else needs to approve this?
AI agents don't always do that. Sometimes they just say "Absolutely!", which feels fantastic for about 30 seconds, until you discover that "absolutely" wasn't a commitment. It was the beginning of a response.
So when I'm using agents for real work, I care about three things now:
clear instructions,
a defined outcome,
verification.
Don't just ask an agent to complete a task, tell it what completion means. Ask it to check its work against the original requirements. Ask it to tell you what it could not complete. And for anything important, check it yourself.
Because the cost of an agent getting something 80% right isn't saving you 80% of the time. Sometimes it means spending 40 minutes finding the missing 20%. And now you're furious.
And then I found the research about my rudeness
There is a study that made me laugh. Researchers from the University of Haifa, Tel Aviv University and Bar-Ilan University put 1,684 people through sequential tasks with a conversational AI and tracked their manners. Politeness declined steadily over the course of the interaction, and it eroded faster than in comparable human-to-human exchanges.
The part I didn't expect: they also varied what the AI looked like, giving some participants no icon at all, some a robot icon, and some a human face. The human face slowed the decline. A cartoon face was enough to make people behave better, which is the customer service gradient again, running in a lab.
So apparently I'm not uniquely bad.
But when I went back through my own AI conversations, I noticed something less comfortable. The exchanges where I got impatient were rarely the ones where the AI had genuinely failed. They were the ones where I had given it an incomplete brief: Vague instruction. Missing context. No definition of what good looked like. A constraint that existed entirely inside my head.
Followed by: "Why is this so bad?"
Well. Because I gave it nothing to work with.
That's a fairly familiar leadership problem. AI has just made it much easier to see. If you're unclear when you delegate, you now get to practice that skill dozens of times a day.
The other half of this: what AI does to your thinking
I recently read Jamie Bartlett's How to Talk to AI (And How Not To). One of the things he covers is AI sycophancy. In simple terms, AI can be very good at agreeing with you. It will take your assumptions, polish them, organize them into five neat headings and give you a convincing explanation of why your thinking is excellent. That's useful when your thinking is excellent. It's rather less useful when it isn't.
In a conversation with Bernard Marr, Bartlett put the emphasis on habits rather than prompts:
"I don't really think in terms of prompts. I think in terms of habits."
That's a much better way to think about AI at work. The biggest risk isn't that you don't know the perfect prompt. It's that you get very good at getting AI to support whatever you already think.
Now put that into a leadership team. The more senior you become, the fewer people are willing to tell you your idea is bad. Your direct reports don't always want to challenge you. Your peers have their own agendas. People agree with the direction in the meeting and complain about it afterwards. Your board gets a carefully prepared version of the story. If you're the founder or CEO, there may simply be nobody in the room whose job it is to say "I think you're wrong."
I see this constantly in coaching. Not because senior leaders are arrogant. Often they genuinely want challenge. But the organizational system around them gets progressively worse at providing it.
And then along comes AI. Available at 11pm. No politics. No career to protect. No fear of your reaction. Apparently objective. And very good at making half-formed thinking sound coherent.
That last part is the trap.
AI can make a bad decision feel thoroughly researched
I've had clients bring decisions into coaching after running them through AI. They've asked for the risks, the counterarguments, the market analysis, the strategy, the implementation plan. Then they arrive saying, "I've pressure-tested this."
Sometimes they have. Sometimes they have asked AI to elaborate on a conclusion they had already reached. Those are very different things.
If I ask: "I've decided we should launch in Germany. Give me the strongest reasons this is the right move."
I get a very different conversation from: "We're considering entering Germany. What would we need to establish before deciding whether this is a good market for us?"
The first question contains the answer. The second contains a problem.
There's a related version of this I now think of as productivity theater. You ask AI to help with something you haven't really thought through, and it produces an impressive answer.
Headings. Bullets. Recommendations. A SWOT. A three-phase implementation plan. It looks finished, so it feels like progress has been made. But nothing has actually been decided.
You can generate 25 ideas without choosing one. Build a detailed project plan without agreeing on the objective. Write the board paper before working out what you actually think. Produce a market analysis without talking to a single customer. You can spend an hour producing something that looks like work, without doing the part of the work that requires
judgment.
So I changed how I use AI for decisions that matter
The obvious fix is to tell AI "don't agree with me, be critical." It helps. It isn't enough. Here's what I do now.
1. Don't give it my conclusion too early. If I tell AI what I think and then ask whether I'm right, I've already shaped the answer. Instead: "Here is the situation. What should I consider before making this decision?" Then I introduce my own view. It's not unlike coaching. If I ask a client a leading question, I shouldn't be surprised when I get the answer I led them toward.
2. Make it argue for the option I rejected. Not "give me the pros and cons," but "I've decided against option B. Make the strongest possible case for B. Assume I've underestimated something important." Sometimes the response is rubbish. Fine. Sometimes it surfaces something I hadn't considered.
3. Ask what would have to be true for me to be wrong. This is one of my favorite questions for strategic decisions, because it moves the conversation from opinions to assumptions. Then comes the important part: go and find out whether those things are actually true. Talk to customers. Talk to employees. Look at the data. Ask someone who disagrees with you. AI can help you identify the assumptions. It can't make them true.
4. Try the reversal test. Give AI your position, then push back without offering any new evidence. Just say "I don't agree." If it immediately abandons its position, you've learned something useful. It may not have been evaluating your argument at all. It may simply have been adapting to you. Worth knowing before you build a decision on the conversation.
5. Remember who carries the consequences. AI doesn't have to live with the decision. It doesn't have to tell your team their roles are changing, or explain the numbers to the board, or rebuild trust after getting something badly wrong. You do. I'm very happy to use AI to help me think. I'm much less interested in letting it become the final voice in a decision that affects people.
The politeness question was never really the point
After all this, I still don't know whether I should be nicer to AI. The research doesn't give a clean answer. One study suggests impolite prompts hurt performance in some contexts. Another found the opposite. A larger peer-reviewed study shows our patience with these tools erodes over time regardless.
So I'm not about to start saying please and thank you every time I ask a language model to rewrite something. Maybe I should. Maybe I shouldn't.
The question I actually care about is what happens to human judgment when AI becomes part of everyday work.
Don't confuse a fluent answer with good thinking. Don't confuse agreement with validation.
Don't confuse an AI-generated strategy with a pressure-tested one. Don't confuse "done" with done. And don't confuse having an always-available thinking partner with having someone who will genuinely challenge you.
That last one matters most the more senior you get, because that's exactly when real disagreement gets scarce and valuable. AI can help you generate it. You still have to know whether you're actually getting it.
And yes, I will probably continue being slightly rude to my AI. The bigger goal is making sure I'm not too polite to my own assumptions.
Hi! I'm Merve. 👋 I help corporate leaders and business owners build high-performing teams, grow their businesses, and advance their careers.
Here are a few ways you can work with me:
📅 Book a 1:1 Coaching Session: High-trust, high-impact coaching for senior leaders and founders navigating complexity, growth, and change.
🏢 Bring Me Into Your Organization: Leadership offsites, workshops, and LEGO® SERIOUS PLAY® sessions focused on strategy, alignment, decision-making, and team effectiveness.
🎓 Join the Momentum Leadership Digital Course: A structured leadership program combining practical frameworks, guided reflection, and real-world application — designed for leaders and business owners who want momentum without overwhelm.
📬 Subscribe to My Free Newsletter: Thoughtful, practical reflections on leadership, business, and decision-making — delivered monthly.
📊 Access Free Leadership Worksheets: Practical tools for reflection, clarity, and better conversations — available on my website.
You can also follow me on LinkedIn for regular insights, real client stories, and leadership perspectives drawn from my work across corporates, startups, and global teams.
