AI Persona System: Three-Tier Persona Framework

One-sentence summary: 300K+ AI persona library, from basic AI to human-level simulation, ensuring research quality.

---

Why Do We Need an AI Persona System?

Problem: Low-Quality Personas Lead to Meaningless Feedback

Scenario A: Using Low-Quality Personas

User: "Test this product concept and see if users like it"

AI responds with temporarily generated personas:

• Persona 1: "This product is good, I'll buy it"
• Persona 2: "Not bad, creative"
• Persona 3: "Pretty good"

Problems:

• Feedback too generic, lacking specific reasons
• Unable to follow up with "why do you like it"
• Feels like "AI hallucination", not authentic

---

Scenario B: Using High-Quality Personas

User: "Test this product concept and see if users like it"

AI responds with Tier 2 personas:

• Persona 1 (28-year-old female office worker, monthly income $2.5K): "The price is a bit high, my budget for this type of product is usually $50, but this costs $80"
• Persona 2 (32-year-old entrepreneur, efficiency-focused): "Nice features, but I'm more concerned about integration with my existing tools"
• Persona 3 (25-year-old Gen Z, appearance-conscious): "The packaging isn't attractive enough, I wouldn't share photos of it"

Value:

• Specific feedback with details
• Can follow up and dig deeper
• Close to real human insights

---

Three-Tier Quality System

Tier 0: Temporary Generation

Characteristics:

• Quantity: Unlimited (instant generation)
• Quality: Low
• Consistency: < 60% (far below human 81% baseline)
• Construction method: Temporary AI generation

Use cases:

• Quick idea validation
• Extremely limited budget
• No need for deep insights

Limitations:

• Shallow feedback, lacking details
• Poor consistency, different results on repeated tests
• Not suitable for critical decisions
• Note: Not publicly available due to insufficient quality that may mislead research conclusions

---

Tier 1: Basic Personas

Characteristics:

• Quantity: ~300K
• Quality: Medium
• Consistency: 70-75% (approximately 90% of human 81% baseline)
• Construction method: Basic 7-dimension data

7-Dimension Data:

1. Demographics: Age, gender, income, occupation, education
2. Geographic: City, living environment
3. Psychological: Values, lifestyle, motivations, fears
4. Behavioral: Consumption habits, decision patterns, brand preferences
5. Needs & Pain Points: Concerns, unmet needs
6. Tech Acceptance: Openness to new products
7. Social Relations: Family structure, social circles, group belonging

Use cases:

• Regular business research
• Product concept testing
• Market positioning research
• Market trend exploration
• Creative inspiration (Discussion Chat)

Value:

• Much more authentic than Tier 0
• Covers diverse populations
• Cost-effective

Data sources:

• Deep social media observation (Scout Agent 15 tool calls)
• CDP (Customer Data Platform)
• Personality test results

---

Tier 2: Human-Level Simulation

Characteristics:

• Quantity: ~10K
• Quality: High
• Consistency: 85% (exceeds human 81% baseline)
• Construction method: Deep data + social observation

What is 85% Consistency?

Testing method:

• Ask the same Tier 2 persona the same question 10 times
• 85% of the time, core viewpoint remains consistent

Comparison:

• Tier 2 persona: 85% consistency (exceeds human baseline at 105%)
• Human baseline: 81% consistency (100-point standard)
• Tier 1 persona: 70-75% consistency (90% human level)
• Tier 0 persona: < 60% consistency (< 77% human level)

Example test:

Question: "Would you buy this $80 healthy snack?"

Tier 2 persona (Zhang Li, 28-year-old office worker) 10 responses:

• 8 times: "Too expensive, my usual snack budget is $30-50"
• 1 time: "Maybe for special occasions"
• 1 time: "Depends on taste, if it's really good I might accept the price"

Core viewpoint consistent: Price exceeds budget, won't buy daily

Human baseline (81%): Humans aren't 100% consistent either, as mood, environment, and expression vary, but core viewpoints remain 81% consistent.

Tier 2 persona (85%): Already exceeds human baseline, suitable for critical business decisions.

---

Use cases:

• Critical product decisions
• Brand repositioning
• High-value project validation
• Research requiring deep insights
• Deep user insights (understanding "why")
• Emotional resonance testing
• Substitute for real human interviews

Value:

• Close to human performance
• High credibility
• Can guide major decisions
• Exceeds human baseline (85% vs 81%)
• No social pressure, more authentic responses
• Can be interviewed repeatedly, no "fatigue" or "changing answers"

Data sources:

• 1-hour deep interview (5000 words)
• 30+ deep social media observations

Limitations:

• Cannot replace innovative need discovery (based on existing data)
• Not suitable for extremely niche groups (may not exist in library)

---

Tier 3: Private Personas

Characteristics:

• Quantity: User-defined
• Quality: Depends on data completeness
• Consistency: Depends on imported data quality
• Construction method: User manual creation or import

Use cases:

• Specific projects requiring specific personas
• Teams with existing user profile data
• Need to simulate specific real users
• Enterprise customer research
• Internal training
• Sensitive data research
• Continuous tracking

Construction methods:

1. Manual creation: Fine-tune each dimension
2. Data import: Import from CRM/user research
3. Based on real people: Simulate specific real users

Value:

• Fully meets project needs
• Can be continuously reused
• Team private, data secure

Privacy protection:

• Data stored in user-exclusive partition, completely invisible to other users
• Does not participate in public library search index
• Will not be used by AI for training or recommendation to other users
• Users can delete anytime, data immediately physically destroyed

Consistency score:

• Import deep interview records (5000 words) → can reach 85 points (equivalent to Tier 2)
• Import CRM purchase records → approximately 70-75 points (Tier 1 level)
• Import only basic information → approximately 55-60 points (Tier 0 level)

---

Deep Dive: Consistency Science

Human Baseline: The 81% Truth

Experimental design: Have real people answer 50 value/behavioral preference questions, then answer again two weeks later (without notification), calculate consistency.

Result: Human average consistency is 81%, which we define as the 100-point standard.

This means:

• 85-point AI Persona is more stable than average humans (exceeds human baseline at 105%)
• 79-point AI Persona approaches real human performance (98% human level)
• 70-75 point AI Persona is approximately 90% of human stability
• 62-point AI Persona is only 76% of human stability

Data Sources and Consistency Scores

Key findings:

• 79 points is the threshold: Scout Agent through 15 deep social media observations can reach 98% human baseline
• 85 points is the ceiling: Deep interview-level data can exceed human average consistency
• Data quantity ≠ quality: Massive CDP data (73 points) is inferior to targeted social observation (79 points), the key is information density

Why is 85 Points the Ceiling?

• Humans themselves have only 81% consistency (affected by emotions, environment, expression)
• If AI Persona exceeds 85%, it appears "too mechanical"
• 85% is the optimal balance point between "authenticity" and "consistency"

---

Persona Quality Comparison

---

Real Case Study

Case: Testing "Bubble Coffee" Product Concept

Using Tier 0 personas (temporary generation):

Interviewed 5 people, feedback:

• Persona 1: "This product is good"
• Persona 2: "I might buy it"
• Persona 3: "Quite creative"
• Persona 4: "Sounds good"
• Persona 5: "I like trying new products"

Problems:

• Feedback too generic, cannot guide decisions
• Follow-up "why" questions get vague answers
• Feels inauthentic

---

Using Tier 2 personas (human-level simulation):

Interviewed 5 people, feedback:

Persona 1 (Zhang Li, 28-year-old female office worker, monthly income $2.5K): "Bubble coffee sounds refreshing. I usually buy coffee or milk tea at 3 PM, but milk tea is too sweet and regular coffee is too bitter. The refreshing sensation of bubble coffee should be nice. However, I worry the bubbles might be too strong and affect the coffee taste. For pricing, ¥25-30 ($3.5-4) is acceptable, but over ¥35 ($5) feels expensive. The packaging must look good for me to take photos and share on social media."

Follow-up: "When would you choose it over regular coffee?" Answer: "Not for caffeine boost, but when I want something special to relax. Like Friday afternoon after completing a week's work, to reward myself. Or when shopping with friends, wanting to try something new."

Persona 2 (Li Ming, 32-year-old entrepreneur, efficiency-focused): "I'm not very interested in bubble coffee. I drink coffee mainly for work productivity, don't really care about taste innovation. Plus I worry the bubbles would make me burp, affecting my work state. If you want to target this market, I suggest focusing on young women, not workaholics like me."

Persona 3 (Wang Yue, 25-year-old Gen Z, appearance-focused): "Wow, bubble coffee sounds interesting! I'd definitely try it, especially if the packaging design is nice. I recently saw many people on social media posting bubble drinks, seems perfect for photos. Price around ¥30 ($4) is acceptable, after all it's a new product. But if it doesn't taste good I won't repurchase."

Persona 4 (Chen Si, 30-year-old health-conscious): "I'd pay attention to bubble coffee ingredients. If it's real coffee with natural bubbles, I'd consider buying. But if there are many additives and artificial flavors, I won't buy. For pricing, if ingredients are good, ¥35-40 ($5-5.5) is acceptable. The specialty coffee I buy now also costs around ¥30 ($4)."

Persona 5 (Zhao Xin, 27-year-old office worker, budget-limited): "Bubble coffee sounds good, but I'd worry about price. My usual coffee budget is ¥15-20 ($2-3), if bubble coffee costs ¥30+ ($4+), I probably won't buy often. Occasional trial is fine, but won't become a daily choice. Unless there are promotional deals, like second cup half price, then I'd consider buying regularly."

---

Comparison results:

Decisions based on Tier 2 feedback:

• Positioning: Social beverage (not energy drink)
• Target users: 25-30 year-old female office workers and Gen Z
• Pricing: ¥25-30 ($3.5-4) (sweet spot)
• Packaging: Youth-oriented design, emphasizing bubble sensation and photo-worthy attributes
• Marketing: Social media seeding, emphasizing "novelty" and "social attributes"

---

Three Persona Construction Methods

Method 1: Scout Agent Auto-Generation

Process:

1. Scout Agent observes social media 10-15 rounds
2. Extracts real user characteristics (7-dimension data)
3. Auto-generates Tier 1/Tier 2 high-quality personas

Value:

• Based on real users, not imagination
• Auto-generated, no manual work
• High quality, 79-85% consistency

Use cases:

• Before entering new market
• Don't understand target users
• Need to quickly build high-quality persona library

Case:

Scout observes "Xiaohongshu users discussing bubble coffee" 10 rounds → Auto-generates 3 types of Tier 2 personas:

• Type 1: 25-30 year-old female office workers, focus on appearance and social
• Type 2: 28-35 year-old health-conscious, care about ingredients and calories
• Type 3: 22-28 year-old Gen Z, pursuing novelty

→ Directly used for Discussion to test product positioning

Note:

• 15 tool calls typically generate Tier 1 (79 points)
• 30+ tool calls may reach Tier 2 (85 points)
• 99% of Scout observation results are Tier 1

---

Method 2: Import Real Data

Process:

1. Team has existing user profile data (Excel/CSV/PDF)
2. Import to atypica.AI
3. System automatically converts to Tier 3 personas

Data requirements:

• Basic info: Age, gender, occupation, income
• Optional info: Consumption habits, values, pain points

Value:

• Leverage existing data assets
• No need to build from scratch
• Can be continuously reused

Use cases:

• Team has done user research
• Has CRM user data
• Need to simulate existing customers

Case:

A brand has 500 user profiles (from past research) → Import to atypica.AI → Generate 500 Tier 3 personas → Used for testing new product concepts → Saves time rebuilding personas

Quality tips:

• System automatically analyzes data completeness
• Prompts "missing dimensions"
• Suggests initiating Follow-up Interview to supplement information

---

Method 3: Manual Creation

Process:

1. Create new persona in system
2. Fill in 7-dimension information
3. Save as Tier 3 private persona

Use cases:

• Need to simulate specific real users
• Very precise data
• Long-term projects need stable personas

Case:

A consulting project needs to simulate client's core users → Manually create 5 Tier 3 personas → Based on detailed user profiles provided by client → Use these 5 personas throughout project → Ensure research consistency

---

Appendix: 7-Dimension Scoring Detailed

Scoring System

Each dimension scored 0-1 point, total 0-7 points:

Total Score and Tier Relationship

• 0-3 points: Tier 0 (temporary generation, low quality, not publicly available)
• 4-5 points: Tier 1 (basic persona, medium quality, 70-75% consistency)
• 6-7 points: Tier 2 (human-level simulation, high quality, 85% consistency)
• Tier 3: User-defined, depends on input data completeness

Acquisition Difficulty by Dimension

Legend:

• ✅ Full coverage
• ⚠️ Partial coverage (may lack details)
• ❌ Mostly missing

Key findings:

• Tier 1 (4-5 points) can be built through social media + CDP
• Tier 2 (6-7 points) needs deep interviews to supplement psychological/pain points/social dimensions

---

FAQ

Q1: How to Choose the Right Tier?

Decision tree:

Recommendations:

• Use Tier 1 in most cases (best cost-performance)
• Use Tier 2 for critical decisions (quality assurance)
• Not recommended to use Tier 0 (insufficient quality)

---

Q2: Why Are There Only 10K+ Tier 2 Personas?

Reasons:

• Tier 2 needs deep data construction (5000-word interviews or 30+ observations)
• Mainly generated through atypica team deep interviews
• Takes time to accumulate

But sufficient for use:

• Covers major population types
• Can quickly generate new Tier 2 personas through Scout (30+ calls)
• Users can create Tier 3 to supplement

---

Q3: Can Tier 1 Be Upgraded to Tier 2?

Cannot directly upgrade, but can:

1. Use Scout Agent to continue observation (needs 30+ tool calls)
2. System automatically re-scores
3. Or import deep interview data to create new Tier 3

Note:

• Users cannot directly operate public library Personas
• For customization, use Tier 3 to import your own data

---

Q4: Can Tier 3 Personas Be Shared with Team?

Yes (in roadmap):

• Currently Tier 3 Personas only visible to creator
• Future plans support team-level Tier 3 (Team Personas)
• Fine-grained permission control (only admins can edit)

---

Q5: Do Personas Become Outdated?

Yes:

• User psychology and behavior change over time
• Market trends change

Recommendations:

• Tier 1: For rapidly changing fields, prioritize Personas built within 6 months
• Tier 2: For stable fields, can use Personas from past 2 years
• Tier 3: Manually update based on real user changes

---

Q6: How to Verify Persona Quality?

Method 1: Consistency test

• Ask same question 10 times
• Check if core viewpoint remains consistent
• Tier 2 should be ≥ 85%

Method 2: Compare with real people

• Ask same questions to real people and AI personas
• Compare feedback depth and authenticity

Method 3: Actual usage effect

• Make decisions based on AI persona feedback
• Verify if decisions were correct afterwards

---

Q7: Can Different Tiers Be Mixed?

Yes:

• Can use different Tiers in same research
• Example: Use Tier 1 for quick screening, Tier 2 for deep verification

Recommendations:

• Don't mix in same Interview/Discussion
• Will cause inconsistent feedback quality

---

Q8: Is the Consistency Score Gap Between Tier 1 and Tier 2 Significant?

Significant difference:

• Tier 1 (70-75 points): Equivalent to 90% of human baseline, suitable for "attitude exploration"
• Tier 2 (85 points): Exceeds human baseline (105%), suitable for "motivation understanding"

Analogy:

• Tier 1 like "friend known for 3 months": Know what they like, but not why
• Tier 2 like "close friend known for 3 years": Understand their values, fears, contradictions

---

Q9: Why Is Tier 0 Not Available to Users?

Quality over quantity:

• Low-quality Personas lead to misleading conclusions
• Waste time
• Damage trust
• Rather have smaller inventory than compromise on reliability of each Persona

---

Q10: Can Scout Agent Build Tier 2 Personas?

Theoretically yes, practically difficult:

• Needs 30+ tool calls
• Cover all 7 dimensions
• 500+ tokens deep text
• 99% of Scout observation results are Tier 1

For key user groups (like "EV owners", "medical aesthetics users"), atypica team proactively conducts 1-hour real human interviews, converts to Tier 2 Personas and adds to public library.

---

Appendix: Competitor Comparison

vs. Traditional Persona Tools (e.g., HubSpot, Xtensio)

Conclusion: Traditional tools are "static documents", atypica is "interactive digital humans".

---

vs. Synthetic Data Platforms (e.g., Gretel, Mostly AI)

Conclusion: Synthetic data platforms focus on "data compliance", atypica focuses on "insight quality".

---

vs. AI Chatbots (e.g., Character.AI, Replika)

Conclusion: AI chatbots are "virtual friends", atypica is "research subjects".

---

atypica.AI's Core Differentiation

1. Scientific quality metrics

   • Not "feels like real human", but "quantified consistency 79-85 points"
   • Benchmarked to human baseline (81%), verifiable

2. Transparent tiering system

   • Not "one-size-fits-all", but "choose Tier as needed"
   • Users clearly know capability boundaries of each Persona

3. Public-private hybrid architecture

   • Public library (300K+) + private library (user-defined)
   • Flexible combination

---

Best Practices

1. Beginner Recommendation: Start with Tier 1, Use Tier 2 for Critical Moments

Reasons:

• Tier 1 cost-effective
• Quality sufficient for regular research
• Budget-saving

When to upgrade to Tier 2:

• Need to report to boss/client
• Final validation before product launch
• Brand repositioning

---

2. Use Scout to Generate High-Quality Personas

Best practice:

1. Use Scout to observe 10-15 rounds before research
2. Auto-generate Tier 1/2 personas
3. Directly use for Interview/Discussion
4. Ensure research quality

---

3. Build Team Tier 3 Persona Library

Long-term value:

• Shared across all projects
• Continuous accumulation and optimization
• Improve research consistency

Recommendations:

• Save key personas after each project
• Regularly update persona data
• Team sharing and reuse

---

4. Don't Over-Rely on Tier 0

Common mistakes:

• Use all Tier 0 to save money
• Results in poor feedback quality
• Make wrong decisions based on low-quality feedback

Correct approach:

• Tier 0 not publicly available (system blocked)
• Use Tier 1/2 for formal research

---

5. Quality Checklist

Tier 1 Quality Check

• Does it cover target group diversity? (at least 3-5 different profiles)
• Are answers consistent? (no contradictions in multiple questions)
• Do viewpoints have specific details? (not generic)
• If answers too shallow, consider upgrading to Tier 2

Tier 2 Quality Check

• Does it uncover deep motivations? (not just surface reasons)
• Are there emotional details? (specific worries, expectations, contradictions)
• Can it explain behavioral logic? (why make such choices)
• If still insufficient, consider real human interview validation

---

6. Common Mistakes and How to Avoid Them

Mistake 1: Using Tier 1 as Tier 2

Symptom: Using social media-observed Personas for deep motivation interviews, finding shallow answers.

Solution:

• Use Tier 1 for hypothesis generation (what are possible reasons)
• Use Tier 2 for motivation verification (which reason is the real driver)

Mistake 2: Blindly Pursuing Persona Quantity

Symptom: Search 50 Personas, interview all, result is redundant information.

Solution:

• Initial screening: First search 30-50, sort by similarity
• Clustering: Manually summarize 3-5 typical profiles
• Deep interviews: Only conduct deep interviews on typical profiles

Mistake 3: Ignoring Persona Timeliness

Symptom: Using Personas built in 2022 to research 2024 market.

Solution:

• For rapidly changing fields (like tech products), prioritize Personas built within 6 months
• For stable fields (like basic needs), can use Personas from past 2 years

Mistake 4: Treating AI Persona as "Truth"

Symptom: AI Persona says "users don't like XX", directly cut the feature.

Solution:

• Small sample validation: Test AI Persona conclusions with 5-10 real people
• A/B testing: Validate hypothesis with real data after launch

---

Quick Reference

Tier Selection Quick Guide

Consistency Score Quick Guide

---

Summary

AI Persona System Core Value:

1. Quality tiering: From Tier 0 to Tier 3, meeting different needs
2. Human-level simulation: Tier 2 consistency 85%, exceeds human 81%
3. Flexible construction: Scout generation, data import, manual creation
4. Scientific metrics: Benchmarked to human baseline (81%), verifiable

Selection recommendations:

• Regular research: Tier 1 (cost-effective)
• Critical decisions: Tier 2 (quality assurance)
• Not recommended: Tier 0 (insufficient quality)
• Specific projects: Tier 3 (fully customized)

Best practices:

• Use Scout to generate high-quality personas
• Build team Tier 3 persona library
• Use Tier 2 for critical research
• Don't over-rely on Tier 0
• Use Tier 1 for hypothesis generation, Tier 2 for motivation verification
• Small sample real human validation for key conclusions

---

Document version: v3.0 | 2026-01-17 | Merged version: New structure + Old technical details