The Customer Interaction Landscape

Lecture URL: https://skillsclinic.org/support/recognizing-ai-limitations-in-customer-interactions.php

Recognizing AI Limitations in Customer Interactions

L1.5.1—Limitations and Awareness

Level 1: Awareness

Welcome to lesson L1.5.1: Recognizing AI Limitations in Customer Interactions. This lesson teaches you to identify situations where AI is likely to fail or produce unreliable outputs in customer-facing scenarios.

AI can help in many situations. But there are specific situations where AI produces hallucinations, misses critical context, or makes inappropriate suggestions. Learning to recognize these situations is essential to using AI safely.

This isn't about being anti-AI. It's about understanding AI's real boundaries and making decisions about where to use it and where to keep humans in control.

The Customer Interaction Landscape

Customer support interactions vary along several dimensions that matter for AI safety:

Risk level: From low (routine question) to high (financial dispute, crisis situation)

Emotional intensity: From neutral (technical question) to highly emotional (angry customer, distressed customer)

Context dependency: From standalone question (what's your return policy?) to situation-specific (what should this customer be offered given their history?)

Judgment requirement: From clear right answer to genuinely ambiguous situation with tradeoffs

AI performs well in low-risk, neutral, standalone situations with clear answers. AI performs poorly in high-risk, emotional, context-dependent situations requiring judgment.

Building Your Limitation Awareness

You don't learn AI limitations from reading about them. You learn by experiencing them.

Here's how to build your awareness systematically:

Document failures: When AI fails, note it. What was the situation? What did AI get wrong? Why do you think it failed?

Identify patterns: Over time, patterns emerge. Certain types of situations are consistently problematic. Other types are reliably good.

Share with colleagues: Others have experienced failures you haven't. Their experiences teach you about limitations you might not discover alone.

Review periodically: Monthly, review the limitations you've identified. Are there new patterns? Are there areas where AI is more reliable than you thought?

This systematic approach to learning limitations is more valuable than any list I can provide.

Situation 1: Hallucination in Policy or Product Questions

The risk: AI generates false information about your product, policy, or service.

Customer: "Can I switch my plan mid-month?"

AI draft: "Absolutely! We allow unlimited plan switches with no penalty. You can change anytime."

Reality: Your policy is plans can be changed on monthly anniversary date only. AI has hallucinated a policy that doesn't exist.

Why it happens: AI was trained on lots of text about flexible plans from companies that do allow mid-month switching. It doesn't know your specific policy.

Safe approach: For any policy or product questions, verify AI's answer against your knowledge base before sending.

Situation 2: Missing Context About Customer Relationships

The risk: AI doesn't know about the customer's history, VIP status, past issues, or relationship strain.

Customer: "I'd like a refund. This product doesn't work for me."

AI draft: "We don't offer refunds on this product. Your purchase is final."

Context AI didn't know: This is a 10-year customer. They've never asked for a refund before. They've had three issues with this product line. You've decided to make exceptions for valuable long-term customers.

Why it happens: AI only knows what's in the message you shared. It doesn't have access to customer history or relationship context.

Safe approach: For customers with history or relationship significance, read their history before using AI. Include relevant context in the prompt: "This is a 10-year customer who's had issues with this product before."

Situation 3: Emotional Situations Requiring Genuine Empathy

The risk: AI generates response that sounds empathetic but lacks genuine understanding.

Customer: "My kid is sick and your product isn't helping. I've spent $200 and now I'm stuck with something that doesn't work."

AI draft: "I'm sorry to hear you're experiencing difficulties. We strive for quality and value. Please let me know how I can assist."

The problem: This sounds like a standard corporate response. The customer is clearly stressed and needs genuine understanding, not template empathy.

Why it happens: AI has learned the structure of empathetic responses but doesn't genuinely understand distress or urgency.

Safe approach: For emotionally charged situations, write the response yourself. AI can help with structure or suggestions, but the emotional core should come from you.

Situation 4: Situations Requiring Judgment Calls

The risk: AI generates response that follows standard logic but misses the judgment call involved.

Customer: "I'm asking for an exception to your 30-day return policy. I just realized this product doesn't meet my needs, and I'm 32 days out. Can you help?"

AI draft: "Our return policy is 30 days. Since you're beyond that window, I'm unable to process a return."

The judgment call: Do you make an exception? Factors:

• What's the customer's history?
• How much over the deadline are they?
• Is there a good-faith reason they're asking?
• What will retaining them be worth vs. losing $300 in merchandise?

AI doesn't know how to weigh these factors because it doesn't understand your business priorities, customer relationship value, or risk tolerance.

Why it happens: AI applies rules consistently. It's not equipped to make judgment calls about when rules should be bent.

Safe approach: For situations requiring judgment calls, use AI to draft options but make the decision yourself based on your knowledge of priorities and customer value.

Situation 5: Crises or Sensitive Topics

The risk: AI responds to a serious situation without recognizing its gravity.

Customer: "I don't know how to feed my kids this month. Your product was going to help me earn extra income, but it's not working. What do I do?"

AI draft: "I'd be happy to help troubleshoot the product. Can you describe what's not working?"

The problem: This is a crisis situation. The customer is facing food insecurity. A troubleshooting response misses the core issue.

Why it happens: AI focuses on the technical problem (product not working) and misses the human context (customer is in crisis).

Safe approach: For situations suggesting crisis or serious hardship, pause and involve a person who can assess the situation and decide on appropriate response or escalation.

Situation 6: Legal or Compliance-Sensitive Issues

The risk: AI generates response that sounds reasonable but has unintended legal implications.

Customer: "Your product caused me injury. I want compensation."

AI draft: "We're very sorry about your experience. We'll do whatever it takes to make this right. Please send us documentation and we'll cover all costs."

The problem: This response could be interpreted as accepting legal liability. It might create problems for your legal department.

Why it happens: AI is trained on friendly, solution-oriented customer service language. It doesn't understand legal implications of certain phrasings.

Safe approach: For any legal or compliance-sensitive situation, have a human review before sending. Consider escalating to legal if appropriate.

Situations Where AI Commonly Fails

The risk: AI generates response that assumes information that isn't actually there.

Customer: "Hi, I need help with my subscription."

AI draft: "I see you're having trouble with automatic renewal. Let me cancel that for you."

The problem: The customer never said what kind of help they need. AI assumed a problem and generated a solution that might be wrong.

Why it happens: AI pattern-matches to common scenarios. When information is sparse, AI fills in details that seem likely but might be completely wrong.

Safe approach: Before using AI, ensure you have enough information about what the customer actually needs. If information is sparse, ask the customer for clarification before using AI.

Learning Limitations in Context

Important: AI limitations depend on context. The same AI might be highly reliable in one situation and unreliable in another.

Example: An AI might be excellent at summarizing customer complaints but terrible at predicting escalation risk. Same AI, different contexts, different reliability levels.

This means you need to learn limitations not just broadly ("AI is bad at X") but contextually ("AI is bad at X when..."

Examples of contextual limitations:

• AI is good at drafting routine responses but poor at tone-matching for angry customers
• AI is reliable with recent product information but unreliable with discontinued products
• AI is good at identifying technical issues but poor at diagnosing customer satisfaction problems
• AI is excellent at generating multiple options but poor at choosing which option is best

Understanding contextual limitations is more sophisticated than learning general rules.

Anti-Pattern 1: Trusting AI Without Verification in High-Risk Situations

You're dealing with a refund dispute. AI drafts a response explaining your refund policy. You send it without checking the policy.

Reality: Policy has changed since AI's training data. Customer receives outdated information. You create a problem instead of solving one.

Anti-Pattern 2: Using Generic AI Responses for Specific Situations

AI has trained on thousands of responses. It can generate what sounds like a reasonable response to almost any situation.

But "sounds reasonable" isn't good enough for your specific customers in your specific situation.

Anti-Pattern 3: Assuming AI Understands Your Customer

You mention the customer's name in your prompt, thinking AI will look up their history and context.

AI doesn't look things up. It works only with information you explicitly provide.

Anti-Patterns: AI Limitation Failures

You're busy. You have 40 tickets. An AI response is probably fine. You don't have time to verify.

Pressure leads you to send unverified AI output in situations where verification matters.

Practice Prompts

Prompt 1: Risk Assessment

Assess five recent customer interactions for AI risk. For each, decide: Is this AI-safe (low risk, clear answer)? Or AI-risky (high risk, needs judgment, emotional, context-dependent)?

Prompt 2: The Hallucination Scenario

A customer asks about a feature you're not sure your product has. Would you use AI to answer? Why or why not?

Prompt 3: The Relationship Factor

How would you modify your approach to using AI when responding to a high-value customer vs. a new customer?

Prompt 4: The Emotional Situation

Describe an emotionally charged interaction you've handled. How would you use AI in that situation? What would AI be good at? What would you need to do yourself?

Situations Where AI Commonly Fails

One. AI fails in high-risk situations, emotionally charged interactions, and situations requiring context or judgment.

Two. AI commonly hallucinates about policy and product features you haven't explicitly provided.

Three. AI doesn't understand customer relationships or history unless you explicitly include it.

Four. AI can sound empathetic without genuine understanding of customer distress.

Five. AI applies rules consistently but can't make judgment calls about when exceptions are appropriate.

Six. Crisis situations, legal issues, and sparse information are red flags for AI limitations.

Seven. Always verify policy and product information AI generates before sending to customers.

Eight. Learn AI limitations contextually, not just as broad rules. Same AI is reliable in some contexts and limited in others.

Nine. Build awareness of limitations through documentation, pattern recognition, and learning from colleagues.

Ten. The most valuable skill is recognizing when NOT to use AI.

Glossary

Hallucination: When AI generates false information confidently.

Context: Information about customer history, relationships, and background relevant to the interaction.

Judgment Call: A decision involving multiple factors and tradeoffs, where reasonable people might decide differently.

Red Flag: A sign indicating a situation where AI is likely to fail.

Verification: Checking that AI output is accurate before sending to customer.

Risk Level: How much damage results if AI's output is wrong.

Reflection Exercise

Reflect on a recent interaction where you used AI. Looking back, were there any limitations that you didn't notice at the time? What information was AI missing? How might that have created problems?

Closing Remarks

Recognizing AI limitations isn't pessimism. It's realism. AI is powerful in specific contexts and limited in others. Your job is learning where those boundaries are and making deliberate decisions about when to use AI and when to rely on human judgment.

In our next lesson, L1.5.2, we'll discuss responsible data handling—protecting customer information when using AI.

Level 1: Awareness | Limitations and Awareness | Lesson 1.5.1

A SkillsClinic initiative.

Key Takeaways

Ten. The most valuable skill is recognizing when NOT to use AI.

Glossary

Hallucination: When AI generates false information confidently.

Risk Level: How much damage results if AI's output is wrong.

Reflection Exercise

Closing Remarks

Level 1: Awareness | Limitations and Awareness | Lesson 1.5.1

A SkillsClinic initiative.