Core Concepts

What Is Artificial Intelligence?

In its broadest sense, artificial intelligence (AI) is software designed to perform tasks that normally require human intelligence. These tasks include:

• Learning from data and examples
• Recognizing patterns
• Making decisions based on incomplete information
• Understanding language
• Generating new content (text, images, etc.)

AI is not new. Recommendation algorithms on Netflix are AI. Spam filters are AI. Your support ticketing system's auto-assign feature is AI. These are all successful, mature technologies that you probably use without thinking about them.

What Is Generative AI?

Generative AI is a specific type of AI that creates new content in response to a prompt or request. It doesn't just filter, classify, or match existing data—it generates new text, images, code, or other outputs that didn't exist before.

The generative AI tools now appearing in customer support (like ChatGPT, Claude, or built-in "AI assist" features) work by learning patterns from vast amounts of text during training. Think of it this way:

Training phase (happens once, not repeatedly):

• The AI system is shown billions of sentences and documents
• It learns statistical patterns: which words tend to follow other words, which topics relate to each other, common structures in how humans write
• It doesn't "memorize" the documents—it learns the probability patterns underlying human language
• This process is expensive and complex, and happens once (not continuously)

Generation phase (happens every time you use it):

• You ask the AI a question or give it a prompt
• The AI doesn't search a database or retrieve existing documents
• Instead, it predicts the most likely next word, then the word after that, building a response token by token (word by word)
• It makes probabilistic choices at each step, so identical prompts can produce slightly different outputs
• This happens very fast, but it's fundamentally prediction, not retrieval

The Three Core Limitations of Generative AI

Understanding these three limitations is essential for using AI safely in customer support:

1. AI Does Not Have True Understanding

Generative AI is very good at pattern matching and prediction, but it does not actually understand what it's saying. Consider this:

If you ask an AI tool: "What's the capital of France?", it generates "Paris" because the training data contains millions of patterns where "capital of France" is followed by "Paris." The AI is not looking up a fact in a knowledge base. It's predicting the most likely next word.

This works fine for well-established facts. But it breaks down when:

• The customer's situation is ambiguous or context-dependent
• Policy changes recently (after the AI's training data)
• The question is about something niche or unusual
• The customer's problem requires reasoning about cause and effect

The AI has no way to verify its own output against reality. It can sound confident about something that's wrong.

Example: A customer asks about a refund policy that changed last month. The AI generates a confident response based on the old policy from its training data. It sounds authoritative. But it's wrong. The customer trusts it, and your company has a problem.

2. AI Cannot See, Understand, or Respect Policies, Guardrails, or Accountability

AI is trained on patterns in data. If your customer support documentation, internal guides, and previous responses are all available to the AI during training, the AI learns general patterns about how you handle issues. But it doesn't have a "policy reader" that ensures every response aligns with current policies.

More fundamentally: When an AI makes a mistake, who is accountable?

In customer support, accountability matters. If an agent violates a policy, there's a paper trail, and the agent (and their manager) bear responsibility. If an AI-generated response is wrong:

• Did the AI "know" it was wrong?
• Did the system operator understand what the AI was doing?
• Did the company set appropriate guardrails?
• Did a human review the output?

Accountability is human. AI cannot bear accountability. This is why human review is not optional—it's a structural requirement.

3. AI Cannot Make True Judgment Calls

Customer support often requires judgment:

• "Should I escalate this to a specialist?"
• "Is the customer's tone indicating frustration or just directness?"
• "Does this situation require a policy exception?"
• "Is the customer likely to be satisfied with this answer?"

AI can predict patterns in these scenarios. It can recognize surface-level emotional language. But it cannot make genuine judgment about context, appropriateness, or nuance. It cannot carry accountability for a judgment call.

Example: A long-term customer with a history of positive interactions is asking for an exception to a policy. An AI tool might generate: "I understand, but our policy doesn't allow this." That's technically accurate, but it misses the relationship context and the opportunity for human judgment. A human agent, reviewing the AI draft, would likely edit it to something warmer and more relationship-aware.

AI Assistance vs. AI Automation

This distinction is critical for how you'll work with AI in Level 2:

AI Assistance means:

• AI generates a draft, suggestion, or starting point
• A human reviews it thoroughly
• The human decides whether to use it, edit it, or discard it
• The human bears responsibility for the output
• The work flow is: Human -> AI -> Human -> Customer

AI Automation means:

• An AI system generates a response or makes a decision with little or no human review
• The output goes directly to the customer or directly implements an action
• The company relies on AI judgment without human oversight
• The work flow is: AI -> Customer (no human in between)

In customer support, AI assistance is often appropriate. AI automation is usually risky.

Why? Because customer support requires:

• Accountability (which is human)
• Judgment about context and nuance (which AI cannot truly make)
• Recognition of edge cases and exceptions (which AI frequently misses)
• Empathy and relationship awareness (which AI simulates but doesn't possess)

How Generative AI Actually Works: A Practical Mental Model

Here's a simplified model that captures the essentials without diving into deep mathematics:

1. Training: The AI system is trained on vast amounts of text (billions of sentences). During training, it learns statistical patterns. Given the word "customer," the next word is statistically likely to be "service," "support," "feedback," etc. This happens for all positions in all sentences—a massive pattern-matching model.
2. Prompt Processing: You give the AI a prompt: "Summarize this support ticket." The AI processes this as a pattern-matching problem: "What typically follows a request to summarize a ticket?"
3. Token Generation: The AI predicts the most likely next word (token). Then the next. Then the next. Building up the response one word at a time. Each prediction is probabilistic—there's a small chance of picking a less-likely word, which is why you can get different outputs from the same prompt.
4. Constraints and Biases: The AI's responses are shaped by:

• What was most common in its training data
• Built-in guardrails (some AI systems are trained to refuse certain types of requests)
• Your specific prompt and context

1. Delivery: The full response is delivered to you. It sounds fluent and confident, whether or not the underlying facts are accurate.

The "Hallucination" Problem

When AI generates something that sounds plausible but is actually false or made up, we call it hallucination. This is one of the most important concepts to understand.

Why does it happen?

• The AI is predicting the most likely next word, not retrieving facts from a database
• If there's a pattern in the training data where something could be true, the AI might generate it
• The AI has no internal fact-checker or verification mechanism
• The AI doesn't "know" when it's inventing vs. recalling

What does it look like?

• A specific but false product feature ("This subscription includes priority support," when it doesn't)
• A fake policy ("Our refund window is 60 days," when it's actually 30)
• A fake example or precedent ("Customers with issue XYZ typically get store credit," when this never happens)
• A confident but incorrect explanation of cause and effect

In customer support, hallucination is a real risk. A well-meaning agent uses AI to draft a response, misses the hallucinated detail, and sends it to a customer. The customer now has incorrect information. They might make a decision based on it. They might contact you again. Your company's trustworthiness takes a hit.

This is why human review is essential, and why you need to know what to look for.

What AI Can Reliably Do

It's important not to dismiss AI entirely. There are things it genuinely excels at:

Summarization: Pulling out key points from a long ticket or conversation. (But the summary can still be inaccurate, so review is needed.)

Drafting: Generating an initial response based on a ticket. You'll review and edit it, but starting from an AI draft can be faster than starting from scratch.

Pattern Matching: Identifying which of your customers' issues are similar, or spotting when a ticket matches a known pattern.

Tone and Sentiment Detection: Recognizing when a customer sounds frustrated, satisfied, or confused. (But the AI can still misread tone in context.)

Retrieval Assistance: Helping you search through knowledge bases, policies, or internal documentation by understanding what you're looking for. (But the AI can still retrieve the wrong information.)

Brainstorming: Generating multiple response options to consider, or thinking through edge cases. (But you'll need to evaluate each option.)

What makes these "reliable" is not that AI never fails—it does. It's that when AI works well in these domains, the gaps can be caught by human review. And when AI fails, the failure mode is usually "incomplete" or "needs editing," not "catastrophically wrong."