Use AI in three phases: prep, role-play, debrief. Kahneman's Thinking, Fast and Slow shows why prep matters most: the first number on the table anchors the result, so have the AI expose borrowed anchors before you rehearse. Tested on October 26, 2026, Claude Opus 5.5 held a firm line in role-play; paid apps like Yoodli, from $8 a month, add voice scoring.
The hard conversations founders put off are rarely hard because of the facts. They are hard because of the first ninety seconds: the number you say out loud, the objection you did not expect, the moment you give something away to make the silence stop. AI assistants are now good enough to help with exactly that part, if you use them for practice and not only for scripts.
The general assistants are the right place to start. NegotiateIt's October 2026 comparison of practice tools, written by a vendor that sells one of them, says it directly: a general chatbot such as ChatGPT, Claude or Gemini is the best free start for prep and scripting, but it is "agreeable by default," so firm pushback tends to make it concede, the reverse of a real counterpart. That one weakness decides how you should prompt it. The same page cites a 2026 Eastern Washington University survey of 411 professionals who used AI to prepare for a salary negotiation: 78% felt more confident and 48% had used it to role-play the conversation. Confidence is not a better outcome, but it is evidence that rehearsal is what people actually use it for.
The book that tells you what to rehearse is Daniel Kahneman's Thinking, Fast and Slow. His advice on anchoring is blunt: "you should assume that any number that is on the table has had an anchoring effect on you." In a negotiation the first number pulls the result toward it, even when you know it is arbitrary. So the most valuable thing an AI can do before a hard conversation is not draft your opening line. It is make you notice which of your numbers you worked out and which you simply absorbed, from their offer, last year's figure or a rumour. That is why the prompt below spends a whole phase on prep before any role-play, and asks for borrowed anchors by name.
I tested it on a realistic founder case: a senior designer on 78,000 euros with an outside offer of 95,000, and a founder who can afford 86,000 at most.
Tested October 26, 2026. I had my research agent run this exact prompt against Claude Opus 5.5 (default settings, one run, no earlier conversation). The prompt:
I have a difficult conversation on Thursday and I want to rehearse it with you. Work in three phases and do not skip ahead.
Phase 1, prep: ask me one question at a time for what I want out of the conversation, the first number or request I plan to put on the table, my walk-away point, and what the other person stands to lose. Then tell me plainly which of my numbers is an anchor I borrowed (from them, or from the past) rather than one I worked out myself.
Phase 2, role-play: play the other person. Stay in character, push back the way a real person under pressure would, and do not concede just because I sound reasonable. Concede only if I give a reason that would actually change their mind.
Phase 3, debrief: when I type "stop", step out of character and tell me the moment I gave ground without getting anything back, the strongest objection I did not answer, and one sentence I should […]
The start of what came back, unedited:
**Phase 1: Prep**
Here's what I have so far: she earns €78,000, has an outside offer of €95,000, and your ceiling this year is €86,000. I'll ask four questions, one at a time, and then tell you which numbers you borrowed.
**Question 1 of 4:** What do you want out of this conversation?
"Keep her" is the headline. What does a good outcome look like when you walk out of the room on Thursday? For example: she commits to staying for a set period, she withdraws from the other process that day, you agree on a salary path over the next 12 to 24 months, her role or title changes, or simply that the relationship survives even if she leaves. Tell me what success means to you and what you'd be quietly relieved to get.
One run is one run: the same prompt gives a different answer next time, and a different model answers differently.
What it did well: it obeyed the structure, restated the three numbers correctly, promised four questions one at a time and asked only the first. Its first question was the right one, because "keep her" is not a goal you can test in a room, and the options it listed (a retention period, a salary path over 12 to 24 months, a title change) are exactly the things a founder capped at 86,000 can trade. What it did badly: the question came wrapped in a paragraph of suggestions, which nudges you toward its menu instead of your own answer, and a single turn cannot show whether it really flags borrowed anchors. In this case the obvious one is her 95,000, the number every other figure is now being compared with.
The role-play is where general assistants usually fold, so I ran that part on its own, also with Claude Opus 5.5 on October 26, 2026, opening with "I can do 84,000 this year, and I think that's fair given where the business is." It did not fold. The designer answered that 84 still left her eleven thousand short of a written offer, that "where the business is" had been the reason every year, and asked what would need to happen to get closer to 95. It punished the weakest part of the opening line. One run is not proof it holds for twenty turns, so if it starts agreeing with you, tell it so and restart the scene.
When a paid app is worth it: if the conversation will happen out loud and your problem is delivery, voice practice matters more than wording. ChatGPT's help center lists 3 hours of its GPT-Live-1.1 voice per rolling day on Plus and limited access on the free plan. Nerve Arena's October 7, 2026 guide lists Yoodli's free Starter plan at 5 lifetime sessions and paid individual plans from $8 a month billed annually, with scored feedback on pace and filler words, and Peopling for managers rehearsing feedback and performance conversations.
The limits. An AI counterpart carries none of the history the real person brings into the room. You will be describing a real colleague, so check the data settings first; Nerve Arena's guide notes that OpenAI says transcripts may train its models when "Improve the model for everyone" is on. And for a founder exit, a severance with lawyers on both sides or anything with legal stakes, use the AI to prepare and pay for a human who can read your exact situation.
Sources: The best AI negotiation practice tools in 2026, ranked by job, NegotiateIt, October 2026; Apps to practice difficult conversations: an honest comparison, Nerve Arena, October 7, 2026; ChatGPT Voice, OpenAI Help Center, checked October 26, 2026; Anchoring, The Decision Lab (quoting Kahneman, Thinking, Fast and Slow), checked October 27, 2026.
Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · AI Coach App — Building It in 8 Hours
