Dear Negotiation Explorer,
Welcome to issue 27 of our NegoAI series.
Over the past two weeks, we've covered how to prompt AI for negotiation preparation (the metaprompt in issue 25) and why that approach works (the 5 elements in issue 26).
But there's a step most people skip: checking if the output is actually good.
AI sounds confident. It structures responses cleanly. It gives you what looks like solid advice. But confident delivery isn't the same as quality content.
This week: how to assess AI output before you trust it—and how to make AI improve its own work until it meets your standard.
The Problem
I've seen it repeatedly in courses and workshops:
Someone prompts AI, gets a detailed response, and moves forward. Then in the negotiation, something doesn't hold up:
A statistic was hallucinated
The advice was generic—could apply to any deal
A key stakeholder's interest was missed
The "specific talking point" wasn't actually usable
The problem isn't bad prompting. It's accepting output without verification.
And here's the thing: you can make AI verify its own work. You just need to tell it what to check for.
The 4 Factors
Quality AI output for negotiation comes down to four factors:
1. Accuracy (Weight: 30)
Can every claim be verified from the information you provided?
AI can hallucinate facts, statistics, precedents, and even quotes. It presents everything with the same confident tone—whether it's drawing from your context or making things up.
What to check:
Numbers and statistics
Claims about the counterparty
References to past interactions or history
Any "typically" or "usually" statements (often signals guessing)
2. Relevance (Weight: 30)
Is this specific to your situation? Or could this advice apply to any negotiation?
Generic advice is the most common AI failure. "Build rapport before negotiating" is true but useless. "Given their Q3 deadline pressure, open with timeline flexibility" is specific and useful.
What to check:
Does it reference your specific constraints, goals, and context?
Could a competitor get the same advice with different inputs?
Is it telling you what you already know?
3. Completeness (Weight: 25)
Are there gaps? What's missing? What didn't the AI consider that you know matters?
AI works with what you give it. It can't know what it doesn't know. But it also sometimes ignores information you did provide, or fails to connect dots that matter.
What to check:
Key stakeholders not addressed
Risks or downsides not considered
Missing alternatives or Plan B
Information you provided that wasn't used
4. Actionability (Weight: 15)
Can you actually use this?
Vague recommendations aren't actionable. "Consider the relationship dynamics" tells you nothing. "Open with acknowledgment of the delayed payment before asking for the contract extension" is something you can do.
What to check:
Are there specific next steps?
Could you act on this tomorrow?
Is the language something you'd actually say?
Two Frameworks for Assessment
I'll give you two methods—one quick, one comprehensive. Both follow the same principle: make AI evaluate its own work, then improve it.
Framework 1: Traffic Light (Quick Check)
Use when: Time pressure, lower stakes, or initial screening.
Time required: 30 seconds to 2 minutes per iteration.
The Prompt:
Evaluate your previous output using the Traffic Light method.
Base your evaluation ONLY on the context and information I have provided.
Do not assume facts or context beyond what was given.
For each section of your output, answer these 4 questions:
1. ACCURACY: Can every claim be verified from the information provided?
2. RELEVANCE: Is this specific to my situation, not generic advice?
3. COMPLETENESS: Are there gaps or missing considerations?
4. ACTIONABILITY: Can I use this directly in my negotiation?
Assign:
- GREEN: Yes to all 4
- YELLOW: 1-2 questions answered "No" or "Partially"
- RED: 3+ questions answered "No", or any critical failure on Accuracy
For each section:
1. Answer the 4 questions (Yes / No / Partially)
2. Assign Green, Yellow, or Red
3. If Yellow or Red: explain why in one sentence
After scoring, provide:
- Summary of what needs improvement (if anything is Yellow or Red)
- Specific recommendations to reach all Green
Then STOP and wait for my instruction before making changes.After AI responds, you decide:
"Apply your recommendations and re-evaluate" → AI improves and checks again
"Good enough, proceed" → You're done
Threshold: All Green to trust the output.
Framework 2: Weighted Assessment (Comprehensive)
Use when: High-stakes negotiation, important preparation, when you need confidence in the output.
Time required: 5-15 minutes per iteration.
The Prompt:
Evaluate your previous output using these 4 factors.
Base your evaluation ONLY on the context and information I have provided.
Do not assume facts or context beyond what was given.
| Factor | Weight | Score 1-5 |
|--------|--------|-----------|
| Accuracy (verifiable facts, no hallucinations) | 30 | ? |
| Relevance (specific to my situation, not generic) | 30 | ? |
| Completeness (covers what I need, no major gaps) | 25 | ? |
| Actionability (I can use this in my negotiation) | 15 | ? |
1. Score each factor 1-5
2. Calculate weighted total (max 500)
3. For any factor below 4: explain the weakness and specific recommendation
Present:
- Scores and weighted total
- Recommendations for improvement (if below 450)
Then STOP and wait for my instruction before making changes.
Target: 450+ to proceed.Free Live Session: February 12th
Build Your AI Negotiation System — No Code, No Experience
Thursday, February 12th at 6:00 PM CET | 30 minutes | Free
In 30 minutes, you'll see:
How to build your first AI negotiation assistant from zero (no coding needed)
How to prepare for a deal in less than 25 minutes
The 3 questions to ask before trusting any AI with your negotiation strategy
How to spot when AI makes things up—and structure prompts for reliable outputs
The TOAST Method: Loop Until Quality
Both frameworks follow the same loop:
┌─────────────────────────────────────────┐
│ TEST: AI generates output │
└─────────────────┬───────────────────────┘
↓
┌─────────────────────────────────────────┐
│ OBSERVE: AI evaluates its own work │
│ (Traffic Light or Weighted Assessment) │
└─────────────────┬───────────────────────┘
↓
┌─────────────────────────────────────────┐
│ ANALYZE: AI identifies weaknesses │
│ and recommends improvements │
└─────────────────┬───────────────────────┘
↓
┌─────────────────────────────────────────┐
│ [STOP] — Human reviews and decides │
│ "Improve" or "Good enough" │
└─────────────────┬───────────────────────┘
↓
┌─────────────────────────────────────────┐
│ SOLVE: AI applies improvements │
└─────────────────┬───────────────────────┘
↓
┌─────────────────────────────────────────┐
│ TEST AGAIN: AI re-evaluates │
│ Loop until threshold met │
└─────────────────────────────────────────┘Test → Observe → Analyze → Solve → Test again
The key: you stay in control at every iteration. AI doesn't automatically improve and move on. It stops, shows you the assessment, and waits for your decision.
This matters because:
You might see something AI missed
You might decide "good enough" before perfection
You might want to add context that changes the evaluation
You maintain ownership of the final output
When to Use Which
Situation | Framework | Why |
|---|---|---|
Quick prep, time pressure | Traffic Light | Fast feedback loop, sufficient for lower stakes |
Important negotiation | Weighted Assessment | Thorough evaluation, catches subtle issues |
Initial screening | Traffic Light | Quick filter before investing more time |
Final check before meeting | Weighted Assessment | Confidence that nothing was missed |
Iterating on a draft | Traffic Light first, Weighted at the end | Efficient improvement cycle |
The Complete Workflow
Putting it all together with what we've covered:
Prompt using the metaprompt structure (issue 25)
Build your prompt using the 5 elements (issue 26)
Assess output using Traffic Light or Weighted Assessment (this issue)
Iterate using TOAST until threshold met
Use the verified output in your negotiation
This Week's Exercise (15-20 minutes)
Step 1: Take output from a recent AI negotiation prep (or run the metaprompt from issue 25).
Step 2: Run the Traffic Light prompt. See what comes back Yellow or Red.
Step 3: Tell AI to apply improvements. Re-run the evaluation.
Step 4: Once all Green, run the Weighted Assessment. See if it hits 450.
Step 5: Notice the difference between your initial output and final output.
What's Next
Next week: How to give AI the context it actually needs.
We've covered prompting, the elements that matter, and output assessment. But the quality of everything depends on what you feed in.
I'll share a framework for preparing negotiation context—what to include, what to leave out, and how to structure it so AI can use it effectively.
Prompting is half the job. Verifying is the other half.