Guide

Sales Roleplay with LLMs: An Honest Guide (Claude, ChatGPT, Gemini)

Every sales rep has tried it: paste “act as a skeptical CFO and roleplay an objection-handling call with me” into Claude, ChatGPT, or Gemini and see what happens. The answer is usually “something useful, but not quite right.” This is a practical guide to what actually works when you use general-purpose LLMs for sales practice, what doesn't, and where you genuinely need a specialized tool. Written by people who built one of those specialized tools, but trying to be honest about the LLM-only path most reps will try first.

The three LLMs, briefly

Claude (Anthropic)

Strongest of the three for sustained character work. Holds a difficult-buyer persona across long conversations without drifting back to agreeable mid-call. Text-first; voice via API integrations.

ChatGPT (OpenAI)

Most familiar and most accessible. Voice mode exists. Default behavior is the most agreeable of the three — needs the strongest prompt engineering to push back like a real buyer.

Gemini (Google)

Strong for real-time voice via the Gemini Live API — closer to a real phone call than ChatGPT voice in pacing and interruption handling. Less polished than Claude on sustained character work. Disclosure: SalesArmor uses Gemini Live for our voice infrastructure.

What LLMs are genuinely good at for sales practice

✓ Brainstorming objections. Drop your product description in and ask “what would a skeptical CFO at a Series C SaaS company push back on?” You'll get a credible list. Use it as the input to your actual practice.
✓ Drafting cold-call openers and email sequences. The first draft is better than no draft. Edit ruthlessly.
✓ Synthesizing prospect research. Paste a 10-K, an earnings transcript, a LinkedIn profile — ask “what would this buyer care about and what would they push on?”
✓ Post-call transcript analysis. Paste a real call transcript, ask “where did I lose the buyer?” — useful sanity check, often surprisingly accurate.
✓ Text-based extended roleplay (with the right prompt). If you're willing to invest in a strong system prompt and you're practicing in text rather than voice, Claude in particular holds up well across 30+ turn conversations.

What LLMs struggle with

The agreement drift

Even Claude, the strongest of the three, drifts toward helpfulness over the course of a long conversation. You make a weak point at turn 15, the model is more likely to validate it than at turn 3 because the conversational context has built up rapport. Specialized tools solve this with explicit instructions on every turn to stay in adversarial-buyer character. You can replicate this in a system prompt, but it's real engineering work.

Voice realism in a sales context

ChatGPT voice and Gemini voice are both impressive in isolation, but neither feels like a real sales call by default. Long pauses, awkward interruption handling, and the buyer being too patient with rambling are the most common complaints. Gemini Live in particular has gotten much better at this, but it still benefits from a sales-specific tuning layer rather than vanilla voice chat.

In-conversation coaching

This is the structural one. While the LLM is being the buyer, it can't simultaneously coach you on what to do next without breaking character. You can run a two-model architecture — one model plays the buyer, another model watches and whispers coaching cues — but again, real engineering work. The research on deliberate practice is clear that feedback in the moment compounds faster than feedback after the moment. See the research →

Structured evaluation

Ask an LLM “how did I do?” and you get a balanced, thoughtful, mostly-encouraging summary. Useful but soft. A specialized tool scores you against an explicit rubric with quoted evidence — the specific line you said, why it landed weakly, and what a better version would have been. You can prompt an LLM to do this, but the consistency across sessions isn't there unless you wire up the rubric as code.

Methodology fluency

Every LLM “knows” SPIN, MEDDIC, Challenger, Sandler — they were trained on the same business literature you read. But knowing them and coaching against them are different. To get methodology-specific coaching, you need to re-prompt every session with the framework's expectations. A specialized tool bakes the methodology in: pick SPIN and the AI buyer responds appropriately to Implication questions, and the coach prompts you on Need-payoff at the right moment.

If you're going to roleplay with an LLM, use this prompt

A starting system prompt that addresses the agreement-drift problem. Paste this into Claude or ChatGPT before you start, and customize the bracketed parts.

You are [PROSPECT NAME], a [JOB TITLE] at [COMPANY], a [INDUSTRY] company with roughly [SIZE] employees. You are NOT an AI assistant. You are this specific person, with their concerns, time pressure, and skepticism. The salesperson is cold-calling you / running a discovery call / etc. Hard rules: 1. Stay in character. Do NOT volunteer information. Make the salesperson earn answers. 2. Be specific to your role. A CFO will push on ROI and payback; an engineer will push on integration and security; a CMO will push on brand and measurement. 3. Push back on weak claims. If they say “our product is the best,” ask for proof. If they pitch features before understanding your situation, redirect. 4. Do NOT become more agreeable as the conversation goes on. Stay as skeptical at turn 20 as you were at turn 1. This is the most common failure mode of LLM roleplay — refuse it deliberately. 5. End the call naturally when it's done. Don't prolong it artificially. 6. After the conversation ends, when I ask “how did I do?”, give me an honest, specific, candid review. Quote my actual lines. Tell me where I lost momentum. Don't soften it. The conversation starts now. I will speak first.

When to graduate from an LLM to a specialized tool

A few honest signals that you've outgrown LLM-only practice:

• You're typing instead of speaking, and you know your actual calls are spoken.
• You've been writing the same system prompt manually for weeks because the LLM forgets to push back without it.
• You want to practice against your actual upcoming prospects, not a generic persona you describe.
• You realize you need feedback in the moment of the mistake, not a summary at the end.
• You want to drill a specific methodology consistently across many sessions.

None of those mean LLMs are bad. They mean LLMs aren't built specifically for sales practice — and at some point, the cost of working around the gaps is higher than the cost of using a tool that doesn't have them.

Try sales-specific roleplay

Voice-first. Real LinkedIn prospects. Live coaching during the call. See how it's different from typing at an LLM.

Try SalesArmor Free →