How to practice difficult conversations with AI




Almost everyone has a conversation they keep putting off. A manager rewrites the opening line of a feedback talk three times, then moves the meeting to next week. In a Harris Poll survey for Interact, 69% of managers said they’re often uncomfortable communicating with employees.
Rehearsing in your head rarely helps, because nobody in your head interrupts, goes quiet or pushes back. AI role-play gives you a partner to say the hard thing to out loud, before the real person is in the room. Set up well, it lets you walk into that conversation steady and present.
AI role-play means rehearsing a specific conversation out loud with an AI partner that plays the other person and pushes back when you dodge. You practice your words, your pacing and your reaction to the hard moments before the real conversation. Several kinds of tools can play the part, and each senses a little more than the last:
Any of them can run a useful session, and a PAL adds a few extras along the way.
A partner only plays the other person well if you describe them in detail. A vague prompt gets you a generic employee, and you end up rehearsing a talk you’ll never have. A little preparation fixes that.
Gather these before your first session:
With these in hand, your AI partner has a real situation to play.
One run-through shows you where you stumble, and a few rounds with a clear goal fix those spots before the real conversation. Each round works the same way in a text chatbot, a voice assistant or a PAL, so you can start with whatever you have.
Write a short brief on who the other person is and what’s at stake for each of you. Then add a one-sentence goal, such as “Jordan understands the two missed deadlines and agrees to a weekly check-in.” Name the objection you dread most.
Grounding practice in your own situation has research behind it. A Journal of Applied Psychology meta-analysis of 117 studies found that skills carried over to the job best when trainees practiced on scenarios they wrote themselves.
Common mistake: rehearsing a vague “tough talk” with a generic employee, which trains you for a conversation that never happens.
A good prompt gives the partner a persona, a difficulty level, specific pushback and a request for feedback. These three work in any AI chat tool:
Performance review: Play Jordan, a senior analyst who missed two deadlines this quarter and believes they were unrealistic. Difficulty is high. Interrupt when you feel blamed and raise the extra work you took on. Agree to a plan only after I acknowledge your workload and name one specific change. After I say “end scene,” rate my clarity, empathy, listening and whether we reached an agreed next step.
Layoff: Play Priya, a nine-year employee whose role is being eliminated in a restructuring she didn’t see coming. Begin in silence, then ask whether this is about her performance, whether she can move teams and when her last paycheck arrives. After I say “end scene,” tell me where I was unclear or defensive and whether I explained severance and benefits.
Bad news to a patient’s family: Play Maria, the daughter of a 78-year-old patient whose scan shows the cancer has spread, arriving hopeful after a good week. Ask direct questions, cry when the news lands, and push for a treatment I can’t recommend. After I say “end scene,” score me on each step of the SPIKES protocol.
With a PAL, the same prompt goes into PAL Maker, the Tavus studio where you build a PAL without code. There you can also add Objectives, which set measurable goals for the conversation, and Guardrails, which set limits for the whole session.
Common mistake: an overly agreeable partner. If Jordan says “That’s fair” after your first sentence, revise the prompt to spell out when the partner may concede.
Typing gives you time to polish each line, and the real meeting won’t. Use voice mode if your tool has one, because speaking adds pace, interruptions and awkward silences. Keep going when you stumble, since the real conversation won’t pause either.
A PAL goes a step further because it sees you too. Raven-2, Tavus’s multimodal perception system, picks up your tone, expression and posture, and Sparrow-2 handles turn-taking so the partner waits through your pauses. In Tavus’s testing, when Raven-2 detects sarcasm, the language model knows to play along.
Common mistake: stopping to fix a line. It breaks the pressure you came to practice under, so jot a timestamp and keep going.
Score each session against the same four criteria every time, and use the transcript as your evidence.
Feedback is what makes the review worth doing. In a randomized trial of 51 clinicians and trainees, 30 minutes with an AI virtual patient that gave feedback improved empathy and communication more than reading a module.
Most tools save a transcript, so you can find the moment you spoke over Priya’s silence. Read the transcript before you accept any score, and check that the moment it praises or flags really happened. Tavus transcripts also add speaker roles and per-turn timing.
Raise the difficulty once you hit your goal at the current level. Then change how the other person reacts, for example by going quiet or getting angry. Each change tests whether you can adapt to the person in front of you.
Change one thing at a time, so you can tell which variation caused the stumble. Note what tripped you up and start the next round there. Teams can build the same variations into a leadership training program, so everyone practices the same hard cases.
Common mistake: replaying the same easy version until it feels comfortable.
The best conversations to rehearse carry high stakes and an unpredictable reaction. Managers, HR teams and clinicians face them often, and the right scenario depends on which reaction you fear most.
Cast the partner as a report who believes the deadlines were unrealistic and has taken on extra work. Open with the facts, ask what got in the way, and listen before you propose a fix. A good run ends with a specific agreement, such as a weekly check-in.
With a PAL, even the opening gets a reaction. When a manager speaks quickly and looks at the screen edge, the PAL playing the employee picks up the rushed pace and the averted eyes. It answers, “It sounds like you’ve already decided this is my fault.”
Pay conversations go better when the reasons are clear. An HR consultant speaking at SHRM26 advised explaining pay decisions, noting that transparency builds trust even when budgets are tight. The same goes for a performance improvement plan (PIP) meeting, where one expert quoted by SHRM recommends measurable objectives and “immediate and sustained improvement.”
Practice the explanation until you can give it without notes. A PAL will notice if you can’t. When an HR partner reads “Budgets are tight this year” from notes, the PAL playing the employee asks, “Are you reading that to me?”
Clinicians learn SPIKES, a six-step protocol for delivering bad news. In a Journal of Cancer Education survey of 226 patients with cancer, two steps were followed least. They were asking what the patient already understands (27.7%) and how much they want to know (30.4%).
Cast the partner as a hopeful family member, rehearse those two steps first, and ask for a score on each step afterward. With a PAL, a clinician who opens with “The scan shows the cancer has spread” sees the PAL playing the patient’s daughter notice the hurried pace and go quiet. On the next run, opening with what she understands shifts her questions from alarm to next steps.
An AI partner never tires of another round, but it has limits. Its feedback can miss things, and what you tell it deserves the care you’d give any sensitive work detail.
Treat AI feedback as a draft, because both its judging and its sensing have limits. A Nature Machine Intelligence study found that language models rated empathic communication almost as reliably as human experts, though they can miss nuance an expert would catch. Emotion is harder, since a 2019 review found that people express it very differently across cultures and situations.
The partner has limits too. At low difficulty, sessions can feel scripted, so raise the settings early. For protected leave, prior complaints, layoff notice rules or anything highly charged, bring in a human coach or counsel.
Remove names, employee IDs and dates from your notes, and remove every patient detail. HIPAA, the US health privacy law, has a Safe Harbor identifier list worth using as a checklist. It includes geographic units smaller than a state, dates other than the year and full-face photographs.
Check your HR policy as well, since SHRM advises that a strong AI policy should say what information never goes into an AI system. Before using real health information with any platform, confirm its terms against your own contract. Tavus holds SOC 2 certification, and HIPAA compliance is available on Enterprise plans.
Hard conversations put someone’s livelihood, health or trust on the line. The person across the table remembers how you said it long after the details fade. Rehearsing lets you bring presence to the conversation, so the other person feels you’re paying attention.
With a Tavus PAL, you can rehearse face to face with a partner that sees, hears and responds to how you show up.
See it for yourself. Book a demo.
A partner that perceives delivery can react to interruptions and long pauses, and it can stay quiet while you search for words. It can’t reproduce the shared history behind a real employee’s anger. Write that history into the prompt, for example “she was passed over for promotion last year.”
An AI partner is available at 11 p.m. and lets you fail without anyone at work watching. A colleague knows your organization’s politics and the real person you’re about to face. Use AI for the first rounds and a trusted colleague for the final run-through.
No study sets a count for one specific conversation. Fact-recall research in Psychological Science put the best review spacing at 20% to 40% of the time before a one-week test, roughly one to three days. Treat that as a rule of thumb, and stop when you can reach your outcome against two different reactions.
Memorizing lines tends to make delivery stiff, so rehearse your intent, your first two sentences and your response to the hardest objection. Varying the partner’s reactions forces you to adapt in the moment. That’s what keeps you sounding like yourself when the real conversation goes off script.