How to Use AI to Prepare for Difficult Conversations at Work.

In 1971, NASA introduced a training technique for Apollo astronauts called emergency scenario simulation. The premise was straightforward: before a mission, astronauts would mentally walk through failures they hoped would never happen. A burst oxygen tank. A stuck thruster. A communication blackout. Not because these events were predictable in their specifics, but because the cognitive shock of an unexpected crisis — the gap between what you expected and what’s actually happening — degrades decision-making faster than the crisis itself does.

What NASA discovered is that the brain under surprise doesn’t just respond more slowly. It responds less precisely. It reverts to familiar patterns even when those patterns are wrong for the situation. The fix wasn’t more technical training. It was mental rehearsal of the uncomfortable scenarios, repeated until the surprise wore off and the reasoning could stay online.

What’s interesting is how directly this maps to workplace management — a domain where the uncomfortable scenario isn’t a thruster malfunction but a performance conversation that has been delayed for three months, or a termination that has been rewritten seven times but never delivered. The manager who has mentally rehearsed “what if they cry” or “what if they say they’ll quit” handles those moments with a clarity the unprepared manager cannot access. The difference isn’t emotional toughness. It’s preparation.

The conversation didn’t change. The preparation did.

What the delay actually costs

Most managers know they’re avoiding a conversation. What they underestimate is the cost of the gap.

The behavior being discussed continues in the interim — the missed deadlines, the friction with colleagues, the disengagement that’s quietly spreading. Other team members notice. They watch whether the manager is going to say something, and when nothing happens, they draw their own conclusions about what standards actually mean. Resentment builds on both sides: the employee who senses something is wrong but hasn’t been told, and the manager who has now accumulated three months of frustration and finds the conversation carrying weight it never should have had.

When the conversation finally happens — and it always does eventually — it arrives with all of that accumulated context. The manager is no longer addressing one incident. They’re delivering a summary of a pattern, and the employee is experiencing the full weight of it without having seen any of it coming.

AI doesn’t eliminate the discomfort of difficult conversations. But it can eliminate the unpreparedness that makes avoidance feel like the safer choice. When you’ve already heard the worst-case response in a practice session, the real conversation loses some of its power to derail you.

The 3-step AI prep protocol

This protocol takes roughly 40 minutes. It works for any difficult workplace conversation — performance issues, salary decisions, conflict mediation, hard feedback. The tool to use is Claude (claude.ai) or ChatGPT. No account configuration or plugins required.

Step 1: Clarify what you actually want (10 minutes)

Most managers walk into difficult conversations without having separated two things that feel identical but aren’t: what they need (the business outcome) and what they’re afraid of (the emotional reaction). This conflation is what makes preparation feel impossible — you’re trying to plan a conversation whose objective you haven’t actually defined.

Open Claude and use this prompt:

“I need to have a [type] conversation with [role] about [issue]. My goal is [outcome]. Help me separate what I’m trying to achieve from what I’m afraid of. What am I assuming about how this will go that might not be true?”

The output you’re looking for: a one-sentence objective, three assumptions to examine, and a working definition of what success looks like. Most managers discover they conflated the business need — a change in behavior, a documented warning, a clear decision — with a personal discomfort about being the source of bad news.

Clarifying those separately doesn’t make the conversation easier to have. It makes it easier to plan.

Step 2: Roleplay the real responses (20 minutes)

Not the cooperative version. The realistic one — where the person pushes back, becomes defensive, or starts to cry.

Use this prompt:

“I’m preparing for a difficult conversation. Here’s the situation: [brief description]. Play the role of the other person — realistic, not hostile but not compliant. I’ll give you my opening line. Push back the way a real person would. Start when I give you my opening.”

Then send your planned opening. Run three to five exchanges. Ask Claude to play different versions of the response: cooperative first, then defensive, then emotional. What this reveals is not primarily what the other person will say. It reveals where your own language breaks down. You’ll notice when your opening was too vague to get a real response, when your framing was so apologetic that the feedback didn’t land, and — most importantly — where you would have backed down instead of holding firm.

The purpose of this roleplay isn’t to predict the actual conversation. It’s to surface the moments where your preparation hasn’t reached yet.

Step 3: Build your key phrases (10 minutes)

After the roleplay, ask Claude:

“Based on this conversation, what are the 5 most important things I need to say clearly? Give me specific, non-defensive language I can actually say out loud. Focus on behavioral observations, not personality judgments.”

The distinction between behavioral observations and personality judgments is one of the most practical in management communication. “You’ve missed the last three deadlines” is observable and specific. “You’re not reliable” is a character assessment that the other person will spend the rest of the conversation disputing. AI is good at flagging when your draft language crosses that line, and at offering alternatives that stay behavioral.

The phrases you generate here should be said aloud before the meeting — not just read. The gap between what you plan to say and what you actually say under pressure is closed by rehearsal, not by notes.

Forty minutes. That’s the whole preparation.

Five workplace scenarios with specific prompts

The protocol above applies to any difficult conversation. What changes is how you frame the context for the AI — the specifics of the situation and the particular dynamics that make this scenario hard.

1. Performance issue (PIP conversation)

What makes this hard: the manager must communicate seriousness without the employee feeling ambushed. If they feel surprised by the severity, the conversation becomes about fairness rather than improvement.

Prompt: “I’m a manager giving a performance improvement conversation. The employee has missed deadlines on 3 consecutive projects. My goal is for them to understand the seriousness without feeling blindsided. Play the employee.”

See also: AI for performance conversations for a deeper treatment of how AI can support the documentation and follow-up process.

2. Salary negotiation decline

What makes this hard: the employee is disappointed and may be actively considering leaving. Retaining them requires acknowledging the disappointment without making promises that can’t be kept.

Prompt: “I need to tell a strong performer their raise request has been declined this cycle. They will be disappointed. I want to retain them. Play the employee.”

3. Final warning before termination

What makes this hard: the manager must be unambiguous about consequences without the conversation becoming a confrontation. Clarity here protects both the employee and the organization.

Prompt: “I’m delivering a final warning before possible termination. The employee has been warned twice. I need to be clear about what happens next without being threatening. Play the employee.”

4. Team conflict mediation

What makes this hard: you’re meeting with one person who believes the other person is acting in bad faith. Their version of events is sincerely held and probably incomplete.

Prompt: “Two team members have a recurring conflict. I’m meeting with one of them first. They believe the other person is deliberately undermining them. Play that team member.”

5. Hard feedback to a high performer

What makes this hard: high performers often receive less direct feedback, which means they’re more likely to be surprised — and more likely to be defensive — when they finally hear it.

Prompt: “I need to tell a strong performer that their communication style is creating friction with the team. They’re unaware of the impact. Play the high performer — confident, somewhat defensive.”

This type of conversation sits at the intersection of AI for internal communications and people management — AI can help you draft the feedback and practice the delivery, but the message itself has to come from clear observation.

What AI cannot do here

This is worth being direct about, because the preparation can start to feel like a substitute for the real thing.

AI doesn’t know the person. It doesn’t know the history between you, the power dynamics in your organization, or whether this employee has other things happening outside work that are relevant context. The roleplay is a structural exercise, not a personality model.

AI-generated scripts sound prepared. Don’t read them verbatim in the actual meeting — the person across from you will notice, and it will make the conversation feel staged when it needs to feel human. Use the phrases you developed as anchors, not as a script.

For terminations and performance improvement plans: always run the framework by your HR team and legal counsel before the meeting, not just by an AI. Compliance requirements, documentation standards, and legally appropriate phrasing vary by jurisdiction and employment type. AI can help you get your thinking clear. HR and legal are what make the conversation sound.

And finally: the preparation helps with words. It doesn’t eliminate the anxiety of sitting across from someone and delivering hard news. That’s appropriate. Some discomfort in a difficult conversation is a sign that you’re taking it seriously. The goal isn’t to feel nothing. The goal is to be clear.

Use the preparation to clear your head. The conversation still requires your presence.

Try this today

Think of the one conversation you’ve been avoiding. Write three sentences: what you need to say, what you’re afraid they’ll say back, and what you’d do if they did. Paste those three sentences into Claude with: “You are the person I’m about to talk to. Respond to this opening: [your opening line]. Be realistic, not compliant.”

Most people discover in under 10 minutes that their planned opening was either too vague to prompt a real response, or so apologetic that the other person wouldn’t take the feedback seriously. That discovery, by itself, is worth the 10 minutes — because it means you can fix the opening before the meeting, not during it.

The conversation you’ve been putting off hasn’t gotten easier in the delay. It’s gotten heavier. Better to hear the hard response in practice, where you can pause and reconsider, than in the room, where you can’t.

Improving the quality of difficult conversations is one of the highest-leverage things a manager can do for employee engagement — not because hard conversations feel good, but because people consistently report that honest, direct feedback is what they wanted more of. They just rarely get it prepared.