Dear Negotiation Explorer,

Welcome to Week 24 of our NegoAI series.

Two weeks ago, I ran a Maven Lightning Session demonstrating how AI transforms negotiation preparation. I showed two approaches: voice prompting with ChatGPT and a four-agent workflow built in Cassidy. This newsletter captures what happened.

Today's negotiations are more complex. We face shrinking margins, buyers under pressure to cut costs, and deals requiring multiple stakeholder approvals. Cross-border negotiations add culture, regulations, different time zones, and distance. Preparation is rarely comprehensive.

AI changes this by structuring interests, options, alternatives, and trade-offs systematically. It processes large amounts of information and generates multiple what-if scenarios. It's like having a high-level expert negotiator on your team that delivers fast, high-quality preparation to everyone.

More importantly, AI supports integrative negotiations even in low-trust settings.

I conducted research last year with 120 experienced senior executives, comparing traditional negotiation preparation against LLM-assisted preparation using ChatGPT.

The results: AI-assisted negotiators achieved 48% higher individual value compared to counterparts not using AI, and up to 84% higher joint value when both parties used AI.

These results used basic LLMs with wide knowledge but no negotiation focus. Today's agents and assistants would likely produce even higher values.

When there's no trust between negotiators, we tend to revert to distributive behavior. AI allows us to explore solutions in parallel, so we don't have to rely on trust-based information exchange. This enables more integrative outcomes, even in low-trust conditions.

I think this is one of the fundamental outcomes of the research and what we're finding today in AI negotiation.

The first demo used voice prompting with ChatGPT. I took McKinsey's perspective in a renegotiation with PharmaCrest, a $45 billion pharmaceutical company.

The case: PharmaCrest has worked with McKinsey for 10 years, spending $18.5 million per year. But only 42% of McKinsey's recommendations are fully implemented, projects take too long, and most work is done by juniors. PharmaCrest has developed AI tools that replicate much of what McKinsey does at the basic level. Sarah Chen, the chief procurement officer, wants to cut fees by 25%, shift to outcome-based pricing, and have more senior teams involved.

McKinsey, on the other side, has deep institutional knowledge of PharmaCrest's politics, culture, and processes. They have strong relationships at the board level. They know AI tools and cheaper competitors are challenging their position.

The prompt structure followed five components: role, context, instructions, criteria, and output.

Role: "You are Deepak Malhotra, the renowned Harvard University professor."

Context: We're preparing for a negotiation with PharmaCrest, representing McKinsey. The relationship is fundamental.

Instructions: Summarize the case study. Identify and rank the seller's interests. Analyze the seller's BATNA, including explicit and implicit alternatives. Create three distinct scenarios for the potential buyer, each with plausible ranked interests and BATNA. Develop creative options that create value and enlarge the pie.

Criteria: Outline interests beyond basic price. Analyze BATNA by determining both explicit and potential alternatives. Be creative yet realistic with buyer scenarios. Focus on delivering value-added insights beyond what a typical negotiator would do.

Output: Present with clear headings and bullet points for each section.

ChatGPT generated a comprehensive analysis. It identified McKinsey's interests (highest priority: provide strategic operational value, cost efficiency, pricing flexibility, improve implementation success, knowledge transfer). It determined BATNAs and created buyer scenarios. It suggested creative options like a co-created AI-led innovation hub and outcome-based pricing.

Because LLMs are prone to inconsistent outputs, hallucinations, and black box behavior, I've started developing quality assessments of the output.

I asked ChatGPT: "I think some of the things you said are not completely correct, so I would like you to perform a quality assessment of the output. Identify four factors against which to conduct this assessment and assign each factor a weight from 1 to 100%. Then provide a score for each section of your output against each factor. At the end of the quality assessment, please include recommendations for improvement."

ChatGPT identified four factors: depth and accuracy of interest analysis, BATNA analysis, creativity and realism of scenarios, and value-added insight. It provided scores from one to five and gave recommendations for improvement.

This didn't catch that it had reversed buyer and seller in the analysis (we used Instant mode, not Thinking Mode), but it provided useful guidance for refinement.

The second demo used a four-agent workflow in Cassidy, a no-code platform. I ran this workflow 26 hours before the session, taking PharmaCrest's perspective.

The workflow structure:

A compiler agent synthesizes everything into a structured preparation document. The workflow took about three minutes to run.

The output was comprehensive: PharmaCrest's interests ranked by priority, their BATNA with explicit and implicit alternatives, their reservation price with justification. Same for McKinsey. ZOPA calculation showing overlap between both parties' acceptable ranges. Creative options for expanding the pie—performance-based pricing, AI collaboration initiatives, knowledge transfer programs.

The behavioral analysis showed David Morrison is relationship-focused, values long-term partnerships, and is likely to respond well to collaborative framing. This gave Sarah Chen specific tactical guidance.

The quality here is much more consistent than voice prompting because the agents are specialized and follow structured workflows. Each agent has specific instructions and constraints, so the output is reliable and comprehensive.

Select an upcoming negotiation and experiment with both approaches:

Voice Prompting: Use the five-component framework (role, context, instructions, criteria, output) with ChatGPT or Microsoft Copilot. Ask for a quality assessment of the output.

Compare Outputs: Note where voice prompting excels (speed, flexibility) and where it falls short (consistency, depth of analysis).

Identify Your Needs: Would your negotiation preparation benefit more from quick, flexible analysis or from consistent, comprehensive workflows?

The goal is to understand which approach fits your preparation needs and when you might need to combine both.