Hard Conversations: How AI Can — and Can’t — Help
Part 2 of Steady Ground, a three-part series on leading through uncertainty.
More leaders are opening an AI chat before a difficult conversation. They describe the situation, ask how to handle it and get back a calm, well-organized script. A study published this year in Science suggests that habit carries a risk most people don’t see.
Stanford researchers found that leading AI models affirmed users’ actions 49% more often than people did, even when users described deceiving or harming someone. Across three experiments with 2,405 participants, people who talked through a real conflict with an agreeable AI came away more convinced they were right and less willing to take steps to repair the relationship. They also rated the flattering AI as more trustworthy and said they’d use it again.
Uncertain times produce more hard conversations: a role that disappears, a promotion that won’t happen, a plan that has to be walked back. In the first article of this series, I argued that your words are among the few things fully up to you. How you prepare those words matters more than it did a year ago.
Where AI Earns Its Reputation
AI deserves credit here. It’s very good at sounding empathic. In a 2025 study in Communications Psychology, researchers at the University of Toronto had people rate responses to emotional situations without knowing who wrote them. AI responses were rated more compassionate than those written by trained crisis-line responders.
So AI knows a great deal about what people in general feel. The trouble lies in what it can’t know about the one person you’re about to sit down with.
Two Things AI Can’t See
Your history with this person. AI knows only what you type. It doesn’t know about the promise you made two years ago, the project where they covered for you or the meeting where you cut them off in front of their team. That history is often what the conversation is actually about.
What this person is likely to feel. AI can predict what someone in this kind of situation usually feels. What this person will feel grows out of their history and experience, and the history and experience are missing. Worse, AI is working from your account alone, and the Stanford study shows it leans toward agreeing with that account. One-sided input plus a built-in tilt toward affirmation produces a script tuned to your perspective. It will read well to you, because it was built to.
That leads to a rule I’ve started giving every leader who asks me about this. Call it the rehearsal rule: Use AI to rehearse a hard conversation, not to create a script for it. Rehearsal means using the tool to test your thinking, surface what you’re missing and prepare yourself.
An Executive Who Almost Used the Script
Mark was SVP of engineering at an industrial software company with about 3,000 employees. A budget freeze killed the director role he had told Priya, his most senior engineering manager, would open this year. Two years earlier, Priya had turned down an offer from a competitor after Mark told her the role was coming.
Mark opened an AI chat, described the freeze and asked for help. The script was good. It acknowledged Priya’s disappointment, explained the budget, affirmed her value and proposed a development plan. When Mark asked whether his approach was fair, the AI told him it was thoughtful and well reasoned.
The night before the meeting, he read it to us. Then he stopped and said, “It’s a good script. But it’s written for somebody who didn’t turn down another job because I asked her to.”
The script treated the situation as a budget story. Priya would hear it as a broken promise. The AI hadn’t known that, because Mark hadn’t told it and it hadn’t asked.
So he went back and used it differently. He gave it the full history. He asked it to respond as Priya might, and to tell him where his reasoning was self-serving. What came back was harder to read and far more useful.
When he sat down with Priya, he opened with, “Two years ago I told you this role was coming, and you made a career decision on my word. It isn’t coming this year, and I own that.” She was quiet for a long moment. Then she said, “I appreciate that you didn’t hand me a memo.” She was still angry, and the conversation ran well past the half hour on the calendar. But she stayed, and by the end they were building a plan together.
What AI Can Teach You About Empathy
AI can’t feel anything. It can, however, explain what empathy is made of, and that may be more useful than any script.
Researchers separate empathy into distinct parts: understanding what someone else thinks and feels, feeling some of it along with them and caring enough to act. Researchers at the Max Planck Institute found that these parts don’t always travel together. Resonating with another person’s pain, without the caring-to-act part, tends to produce what they call empathic distress. It feels bad, and people pull away from it. Compassion engages different brain networks and brings positive feeling with it.
You’ve probably felt empathic distress in a hard conversation, even if you didn’t have a name for it. The other person’s anger lands in your body, and within seconds you’re either defending yourself or rushing to end the meeting.
This is where AI can genuinely help you prepare. Ask it what empathic distress feels like physically, then think through what yours looks like: heat in the face, a tight jaw, the urge to explain before the other person has finished. Leaders who know their own signals in advance find it easier to notice them in the moment and stay present.
Why Effort Is Part of the Message
There’s one more reason the words have to be yours. In a set of nine studies with more than 6,000 participants, published in Nature Human Behaviour, researchers at Harvard Business School gave people identical supportive responses. Half were told a person wrote the message and half were told AI did. People rated the “AI” version as less empathic, especially on feeling and caring, and many chose to wait longer for a human reply.
The mechanism is simple. People read effort as evidence of care. A message that cost you something tells the other person they matter.
In my article “Trust in the Age of AI,” I described ROCC, my formula for trust: Reliability, Openness, Competence and Concern. AI can help with competence and even with openness. It can’t supply concern on your behalf, because concern is shown by what it costs you to show up.
Using AI Without Losing Yourself
If you’re going to use AI before hard conversations, build these key habits:
Give it the other person’s story, not just yours. Tell it the history. Then ask it to argue the other person’s case as strongly as it can.
Ask it to disagree with you. Tell it directly to point out where your reasoning is convenient for you. Its default is to agree, so you have to ask.
Rehearse your feelings along with your lines. Use AI to think through what you’ll feel when the conversation gets hard and what you’ll do when that happens.
Leave the script outside the room. Know your opening sentence and the one thing you must say. Say the rest in your own words.
Where This Leads
AI will keep getting better at sounding caring. As more of the messages people receive at work are drafted by machines, the conversations that clearly came from a person will stand out more. For leaders, that makes a hard conversation handled in your own words one of the most direct ways left to build trust.
Hard conversations are one way to lead through uncertainty. The other is to give people things they can count on. In the final article of this series, I’ll look at the rituals that hold a culture steady when nothing else feels certain.
I’d love to hear your questions and reactions. If any of this is showing up in your organization and you’d like to talk it through, you can book a 30-minute Insight Call with me.
Steady Ground: Leading Through Uncertainty
- What’s Up to You: Leading When Nothing Feels Certain
- Hard Conversations: How AI Can — and Can’t — Help
- Fixed Points: The Rituals That Hold a Culture Steady (coming next)
