Rehearsing a Difficult Employee Conversation With AI

Why this matters

Most hard conversations with a tech go wrong in one or two sentences: the one that turned a performance problem into a character judgment, or the one that promised a consequence you had not decided on. You usually find those sentences while you are saying them, which is too late. A chat assistant can play the other side of the conversation as many times as you like, and it is good at one specific job: pushing back on exactly the sentence that gives it room. Used that way, it finds your weak lines before the real conversation does. Used as a judge of whether you are right, or whether a discipline decision is lawful, it is worse than useless, because it will tend to agree with however you framed the question.

The rule that makes it work: the model plays the employee, not the judge

You are rehearsing your sentences, not getting a verdict. Two consequences follow, and every step below serves one of them:

• It plays one role and stays in it. A realistic employee who has their own view of what happened. It does not coach mid-conversation, soften its pushback to be helpful, or tell you how you did until you ask.
• It never decides anything. Whether the discipline is fair, whether you can let someone go, whether something the "employee" said in rehearsal creates a legal problem: none of those questions go to the tool. They go to your own judgment, and at the point marked below, to an employment attorney or HR adviser.

The steps

1. Write one purpose sentence and the facts, before opening the tool. The purpose sentence says what you need to change and what you want to learn. The facts are dated and observable, things a camera or a job record would show, not interpretations. Without these the rehearsal drifts into a general argument, and so will the real conversation.

2. Strip what identifies the person. No name, no customer names or addresses, no medical information you know of, no pasting of the personnel file or a prior warning. Keep the shape: "has had one prior written warning, about a year ago" carries what the role-play needs. A rehearsal needs a realistic pattern, not this person's record. See related: What Not to Paste Into an AI Tool, where employee records are one of the categories that stay out.

3. Set the role in the first message. A prompt that works:

You are playing a residential service technician, five years in the trade, generally skilled, who is about to be spoken to by the owner about the problems listed below. Respond as a real person in that position would, including defensiveness, excuses, or raising things the owner does not know about. Do not coach me, evaluate me, or step out of character until I type PAUSE.

The last sentence matters most. Without it the model tends to break role after two exchanges and become a helpful adviser, which ends the rehearsal.

4. Run it three times, with a different temperament each time. Before each run, add one line: "This time you are defensive and feel singled out." "This time you go quiet and give one-word answers." "This time you raise something about your situation that the owner does not know." A real employee may do any of the three, and each tests different sentences: the defensive run tests your facts, the quiet run tests whether your questions are real questions, and the third run tests whether you can stop and listen.

5. After each run, ask for your weak sentences only. Type PAUSE, then:

Quote back, word for word, the three sentences of mine that gave you the most room to push back, and say in one line each why.

Ask for quotes, not a review. "You could be more empathetic" is useless; "you said 'everyone's noticed', and I could ask who" is a sentence you can rewrite. Ask for exactly three so it ranks rather than lists everything.

6. Rewrite those sentences and rerun the hardest round. Change the sentence, not your position. Then run whichever temperament gave you the most trouble once more and see whether the new line holds.

7. Take the facts into the real conversation, not the script. A memorized script sounds memorized and breaks on the first unexpected answer. Bring the purpose sentence, the dated facts and your rewritten opening line. Afterwards, write down what was said, what was agreed and the date you will check back.

The fork: when the rehearsal raises something you should not handle alone

This is not legal advice. Some things an employee says change what kind of conversation it is, and the rehearsal's third temperament exists partly to make you practice hearing them. If the employee raises any of these, in rehearsal or for real, stop the performance discussion, do not decide anything in the room, and talk to an employment attorney or HR adviser before any discipline:

• An injury or a medical condition, including one they connect to work. Under the federal floor, the ADA reaches employers of 15 or more (29 CFR Part 1630) and the FMLA employers of 50 or more, with an employee eligible only where 50 work within 75 miles of their worksite (29 CFR Part 825); state law layers on top, often at lower headcounts, and every state's workers' compensation system protects an employee who reports a work injury from retaliation.
• A complaint of discrimination or harassment. Title VII protects complaining from retaliation at 15 or more employees; many states cover employers of any size.
• A pay complaint, or talking about pay with co-workers. The National Labor Relations Act protects most non-supervisory private-sector employees acting together about pay and conditions, union or not.
• Pregnancy, military service, or a family member's illness.

Walk in to that meeting with the dated facts, a note of exactly what the employee said and when, and what you were about to do. The question for the adviser is whether the action you planned is still safe to take, and how.

Worked example: callbacks and late starts

A 12-person HVAC shop. The owner needs to talk to a tech, five years in, normally solid. Purpose sentence: "I need the callbacks to stop and the first call to start on time, and I want to know if something is getting in the way." The facts, from the job records and the dispatch log:

• 3 callbacks in the last 5 weeks, all condensate drains on installs he completed.
• Late to the first call on 4 of the last 10 workdays, late meaning more than 15 minutes past the start of the customer's window.
• One prior written warning, about a year ago, for a different issue.

Round 1, defensive. Part of the exchange, before any rewriting:

OWNER: I feel like you've stopped caring about
       the work lately.
TECH:  That's not fair. I'm running five, six
       calls a day. Who's saying I don't care?
OWNER: Honestly, everyone's noticed.
TECH:  Everyone? Nobody's said a word to me.
       Which jobs are we talking about?
OWNER: Look, this is your last chance.
TECH:  So you're firing me? Over a couple of
       drains?

After PAUSE, the three weak sentences it quoted back:

1. "I feel like you've stopped caring." A judgment about his character, which he can deny and you cannot prove. Rewrite: "You've had 3 callbacks in 5 weeks, all condensate drains on your installs."
2. "Everyone's noticed." Unprovable, and it invites "who?", which turns the meeting into a hunt for the complainer. Rewrite: drop it. The job records are the source; nobody else needs to be in the room.
3. "This is your last chance." Commits you to a consequence you had not decided on, and turns a correction into a threat in the first minute. Rewrite: "I need this fixed, and I want to understand what's happening first."

Round 2, quiet. The rewritten opening held, but the owner found himself filling every silence with more criticism. The weak sentence it quoted: "And another thing, the huddle." The fix was not a new sentence but fewer: ask one question ("What's going on with the drains?") and wait.

Round 3, raises something new. The model, playing the tech, said: "My back's been bad since that attic job in August. I've been slow getting going in the mornings." In the real conversation two days later, the tech did raise his back. The model had not predicted that; it had given the owner practice at hearing something like it. Because he had rehearsed it, he stopped the discussion of late starts, asked whether the back had been reported, learned it had not, and that afternoon filed the injury report his state's workers' compensation process requires. He spoke to his HR adviser before saying another word about the late starts. The callbacks, which had a separate cause (a drain trap detail he had skipped on new installs), were dealt with as a training issue the same week.

Time spent: three rounds of about 10 minutes, 30 minutes, plus about 15 minutes rewriting, 45 minutes in all, for a conversation that would otherwise have opened with the three sentences above.

What the model gets wrong as an opponent

It does not know your tech. A rehearsal tells you which of your sentences are weak, not how this person will react; do not walk in expecting the model's version of them. It also tends toward one of two extremes, too reasonable or theatrically hostile. If a run feels unlike anyone you have managed, say so in the prompt ("less dramatic, more like a tired tech at the end of a long week") and rerun. And it will occasionally step out of role to offer legal or HR opinions ("you should document this for wrongful termination purposes"). Treat that as noise: it is the judge role creeping back, and the fork above is where that question belongs.

References

• 29 CFR Part 1630, ADA employment regulations, and 29 CFR Part 825, FMLA regulations
• 29 U.S.C. 157, National Labor Relations Act, Section 7
• Your state's workers' compensation agency, for injury reporting and anti-retaliation rules
• See related: What Not to Paste Into an AI Tool
• See related: What a Language Model Is Actually Doing When It Answers