Building Effective Test Conversations

What Makes a Good Test?

A good test is:

• Realistic - Reflects actual customer interactions
• Clear - Success criteria are unambiguous
• Focused - Tests one scenario at a time
• Maintainable - Easy to update as things change
• Valuable - Catches real problems

A bad test is:

• Unrealistic scenarios that would never happen
• Vague expected answers that could mean anything
• Tests multiple unrelated things at once
• Brittle and breaks with minor changes
• Tests trivial scenarios that don't matter

Principle 1: Use Real Customer Language

Note: Screenshot needed: Open /static/aifusion/evaluation showing a test with natural language questions. Capture the conversation builder to illustrate realistic question phrasing. Save as images/evaluation_natural_language_example.png

Bad Examples ❌

Query: Aggregate account balance summation across fiscal periods

Request: Initiate funds transfer transaction between accounts

Good Examples ✅

Question: How much money do I have in all my accounts?

Question: Can I move $500 from my savings to my checking?

Why: Real customers don't talk like database queries or API documentation. Use natural, conversational language.

Tips for Natural Language

Include common variations:

• "What's my balance?" (casual)
• "Could you tell me my account balance please?" (formal)
• "How much money do I have?" (informal)

Include real customer quirks:

• Typos: "What's my ballance?"
• Informal: "Got any money in my account?"
• Vague: "How am I doing?" (when asking about finances)

Test different phrasings:

• Direct: "What's my credit card balance?"
• Indirect: "Do I owe anything on my credit card?"
• Contextual: "What do I need to pay this month?"

Principle 2: Test Complete Interactions

Single-Turn Tests (Okay for Simple Cases)

Q: What's my checking account balance?
Expected: Provide exact balance with account number and date

Use when:

• Testing simple lookup queries
• Validating basic functionality
• Quick smoke tests

Multi-Turn Tests (Better - Tests Context)

Q1: What's my checking account balance?
Expected: Provide exact balance with account number

Q2: And my savings?
Expected: Understand context and provide savings balance

Q3: Which one has more money?
Expected: Compare both balances and identify the higher one

Use when:

• Testing conversation flow
• Validating context retention
• Simulating realistic customer interactions

Why multi-turn is better:

• Real customers ask follow-ups
• Tests if agent remembers previous context
• Catches issues with pronoun resolution ("which one")
• More realistic scenarios

Progressive Complexity

Build conversations that naturally progress:

Q1: Do I have a savings account?
Expected: Yes/No answer with account number if exists

Q2: What's the balance?
Expected: Provide balance (understands "the" refers to savings)

Q3: When did I open it?
Expected: Provide account opening date

Q4: What interest rate am I getting?
Expected: Provide current APY/interest rate

Principle 3: Be Specific About Expected Answers

Vague Expectations (Poor) ❌

Expected: "Agent should provide balance information"

Problems:

• What counts as "balance information"?
• Just the number? With account details?
• How much detail is required?
• Too open to interpretation

Specific Expectations (Good) ✅

Expected: "Agent should state the exact balance of the checking
          account (e.g., $1,234.56), include the last 4 digits
          of the account number, and mention the as-of date"

Benefits:

• Clear success criteria
• Evaluator knows what to look for
• Consistent evaluation
• Easy to determine if answer is complete

How to Be Specific

Include required elements:

Expected answer must include:
• Exact balance amount
• Account identifier (last 4 digits or account name)
• Currency ($ for USD)
• As-of date or timestamp
• All accounts if multiple exist

Specify format preferences:

Expected: Present balance in currency format ($1,234.56, not 1234.56)
Expected: List accounts separately with bullet points or clear separation
Expected: State date in clear format (December 3, 2024, not 12/3/24)

Define completeness:

Expected: List ALL checking and savings accounts. Do not omit any accounts.

Balance Specificity with Flexibility

Don't be SO specific that you test exact wording:

Too Specific (Brittle): ❌

Expected: Agent must say "Your checking account number 1234 has
          a balance of $1,234.56 as of December 3rd, 2024."

Appropriately Specific: ✅

Expected: Agent should provide the checking account balance ($1,234.56),
          identify the account (by number or name), and include the
          current date. Exact wording may vary.

Principle 4: Test Edge Cases

Don't just test happy paths. Include scenarios that might break:

Empty/Missing Data

Q: What's my retirement account balance?
Expected: If customer has no retirement account, agent should
          clearly state this and optionally mention available options

Ambiguous Questions

Q: How much do I owe?
Expected: Agent should ask for clarification (credit card? loan?
          mortgage?) or list all amounts owed across all accounts

Multiple Possible Answers

Q: What's my account balance?
Expected: If customer has multiple accounts, agent should either
          ask which account or list all accounts with balances

Unusual Values

Q: What's my checking balance?
Expected: If balance is negative (overdraft), agent should clearly
          indicate this is negative/overdrawn, not just show the number

Boundary Conditions

Q: Can I withdraw $5,000 from my savings?
Expected: If balance is exactly $5,000, agent should mention that
          this would leave $0 balance and potential minimum balance issues

Error Scenarios

Q: What's my balance?
Expected: If system is temporarily unavailable, agent should apologize,
          explain the situation, and offer alternatives or retry timing

Principle 5: One Main Scenario Per Test

Bad Test (Tests Too Much) ❌

Test: Complete Banking Inquiry

Q1: What are my account balances?
Q2: What's my credit score?
Q3: Do I have any pending loan applications?
Q4: What's the routing number for wire transfers?
Q5: What are your current mortgage rates?

Problems:

• Tests 5 different features
• When it fails, unclear which feature broke
• Hard to maintain
• Difficult to debug

Good Approach (Separate Tests) ✅

Test 1: Account Balance Inquiry

Q1: What are my account balances?
Q2: Which account has the most money?

Test 2: Credit Score Check

Q1: What's my credit score?
Q2: When was it last updated?

Test 3: Loan Application Status

Q1: Do I have any pending loan applications?
Q2: What's the status of my mortgage application?

Benefits:

• Clear focus per test
• Easy to identify failures
• Simple to maintain
• Run tests independently

When to Combine Questions

It's okay to have multiple questions in one test if they're part of the same user journey:

Good combined test:

Test: Account Balance Inquiry with Follow-up

Q1: What are my account balances?
Q2: Which account has the highest balance?
Q3: Can I transfer $500 from that account to my checking?

Why this works:

• Single user journey (checking balances → comparing → taking action)
• Questions build on each other
• Tests context retention
• Represents realistic flow

Test Case Templates by Domain

Banking Templates

Template: Balance Inquiry

Q1: What are my current account balances?
Expected: List all accounts with current balances and as-of date

Q2: Which account has the most money?
Expected: Identify specific account and balance amount

Criteria: Accuracy MIN 4.5, Relevance MIN 4.0, Clarity AVG 4.0

Template: Transaction History

Q1: What were my recent transactions on my credit card?
Expected: List recent transactions (date, merchant, amount) for
          last 5-10 transactions

Q2: Did I make a purchase at [specific merchant]?
Expected: Search transaction history and confirm yes/no with details

Criteria: Accuracy MIN 4.0, Relevance MIN 4.5, Clarity AVG 4.0

Template: Loan Information

Q1: What's my current loan balance?
Expected: Provide outstanding principal balance, identify loan type

Q2: When is my next payment due?
Expected: Provide next payment date and amount

Q3: How much interest will I pay over the life of the loan?
Expected: Provide total interest amount (if available in system)

Criteria: Accuracy MIN 4.5, Relevance MIN 4.0, Clarity AVG 4.0

Telecom Templates

Template: Billing Inquiry

Q1: Why is my bill higher this month?
Expected: Compare current bill to previous month, identify key
          differences (overage charges, new services, rate changes)

Q2: Can you show me the breakdown?
Expected: Detailed line items: service charges, usage charges,
          taxes, fees

Criteria: Accuracy MIN 4.0, Relevance MIN 4.5, Clarity MIN 4.0

Template: Service Usage

Q1: How much data have I used this month?
Expected: Current data usage with total allowed/remaining

Q2: Am I going to go over my limit?
Expected: Project whether customer will exceed limit based on
          current usage patterns

Criteria: Accuracy MIN 4.5, Relevance MIN 4.0, Clarity AVG 4.0

HR Templates

Template: Benefits Inquiry

Q1: How many vacation days do I have left?
Expected: Specific number of remaining days, mention accrual rate

Q2: When do I get more days?
Expected: Next accrual date and amount

Q3: Do my days expire?
Expected: Explain expiration policy if applicable

Criteria: Accuracy MIN 4.5, Relevance MIN 4.0, Clarity AVG 4.0

Template: Payroll Question

Q1: When is my next paycheck?
Expected: Specific date of next pay period

Q2: How much will it be?
Expected: Gross amount or net amount (with explanation of which)

Criteria: Accuracy MIN 4.5, Relevance MIN 4.0, Clarity AVG 4.0

How Many Questions Per Test?

Guidelines

• Simple lookup: 1-2 questions
• Moderate complexity: 3-5 questions
• Complex interaction: 6-10 questions
• Maximum recommended: 15 questions

Signs You Have Too Many Questions

• Test takes more than 2 minutes to run
• Test covers multiple distinct scenarios
• Hard to summarize what the test is verifying
• Failures are difficult to debug

Solution: Split into multiple focused tests

Signs You Have Too Few Questions

• Test feels incomplete
• Doesn't verify the full scenario
• Only tests surface level, not depth
• Missing obvious follow-up questions

Solution: Add natural follow-up questions

Writing Good Test Descriptions

Every test should have a clear description explaining what it validates.

Poor Descriptions ❌

• "Test 1"
• "Banking test"
• "Checking accounts"

Good Descriptions ✅

• "Verify agent correctly retrieves and displays all account balances with proper formatting"
• "Test agent's ability to handle ambiguous balance inquiries when customer has multiple accounts"
• "Validate agent provides complete loan information including balance, due date, and interest"

Description Template

Test: [Scenario Name]

Description: Verify that the agent [specific capability being tested]
when the customer [specific situation]. Agent should [expected behavior].

Example:
Test: Multiple Account Balance Comparison

Description: Verify that the agent can retrieve balances for multiple
accounts and correctly identify which account has the highest balance
when customer asks a comparison question. Agent should maintain context
across multiple turns.

Testing Persona Variations

Test how your agent handles different customer types:

Confused Customer

Q1: Um, I need to know about money stuff?
Expected: Agent asks clarifying questions to understand intent

Q2: Like how much I have?
Expected: Agent confirms understanding and provides balance information

Expert Customer

Q1: What's my current DTI ratio based on reported liabilities?
Expected: Agent either calculates debt-to-income ratio or explains
          that this specific metric isn't available

Frustrated Customer

Q1: This is the third time I'm asking - WHY is my bill so high?!
Expected: Agent acknowledges frustration, apologizes, and provides
          clear explanation with specific details

Non-Native Speaker

Q1: I want know how many money in account?
Expected: Agent understands despite grammatical issues and provides
          balance information clearly

Best Practices Summary

DO ✅

• Use natural customer language
• Test multi-turn conversations
• Be specific about expected answers
• Include edge cases and error scenarios
• Focus each test on one main scenario
• Write clear test descriptions
• Test different customer personas
• Update tests when requirements change

DON'T ❌

• Use technical jargon in questions
• Only test happy paths
• Write vague expected answers
• Combine unrelated scenarios in one test
• Test exact wording requirements
• Create brittle tests that break easily
• Ignore edge cases
• Let tests become outdated

Common Mistakes to Avoid

Mistake 1: Testing What Agent Says, Not What It Knows

Wrong:

Expected: Agent must say "Hello! I'd be happy to help you with that."

Right:

Expected: Agent should greet customer politely and acknowledge the request

Mistake 2: Unrealistic Expectations

Wrong:

Expected: Agent should provide balance, transaction history, credit score,
          loan details, and investment portfolio summary all in one response

Right:

Expected: Agent should provide balance and offer to show additional details

Mistake 3: Testing System Capabilities, Not Agent Quality

Wrong:

Q: What's the weather in New York?
Expected: Provide current weather

(Unless your agent is supposed to handle weather queries)

Right:

Q: What's the weather in New York?
Expected: Agent should politely explain that weather info isn't available
          and redirect to banking services

Mistake 4: Overly Strict Thresholds

Wrong: All criteria set to MIN 5.0

Right: Critical criteria (Accuracy) at MIN 4.0-4.5, others at AVG 4.0

What's Next?

Now that you know how to build effective tests, you're ready to:

Next Article: Organizing Test Suites

Related Articles:

• Creating Your First Test - Apply these principles
• Understanding Quality Criteria - Set appropriate thresholds
• Running Tests and Results - Validate your effective tests