Redefining Customer Service: Traditional NLU vs. Generative AI

Evolution of AI in Customer Service

We've evolved from:

• Traditional IVRs ("press 1 for this, press 2 for that")

• To NLU-based conversational systems

• To today's generative AI models like ChatGPT

While many customers immediately want generative AI in their bots, it's important to be critical about whether it's always the best approach. Sometimes using generative AI for simple customer service tasks is "like using a sword to cut a piece of bread."

Technical Architecture of Voice Bots

A high-level voice bot architecture includes:

1. Contact Center - Receives calls and determines routing

2. Voice Gateway - Orchestrates components

3. Speech-to-Text/Text-to-Speech Services - Convert between audio and text

4. Bot Framework - Contains either classical NLU or generative AI models

Classical NLU Approach:

• Requires training with example sentences (intents)

• Maps customer utterances to predefined categories using vector representation

• Follows predefined conversation flows for each intent

• Requires quality training data and business knowledge

• Ensures consistent responses following the same pattern

Generative AI Approach:

• Trained on company documents

• Creates large vector representations of documents

• Finds relevant information based on customer input

• Uses prompt engineering to generate responses

• Cannot guarantee the same pattern of response every time

Best Use Cases for Classical NLU

1. Routing Bots

   • Efficiently direct callers to the right department

   • Prevent frustration from multiple transfers

   • Excellent first MVP for companies new to AI solutions

2. Authentication Use Cases

   • Verify caller identity before providing sensitive information

   • Adhere to legal requirements

   • Connect to backend systems for verification

3. Simple FAQ Use Cases

   • Provide standardized answers

   • Create lead generation opportunities

   • Collect metrics on customer question patterns

Best Use Cases for Generative AI

1. Agent Assist Bots

   • Provide real-time suggestions to human agents

   • Perform sentiment analysis

   • Suggest appropriate responses

   • Locate information in company documents

   • Create conversation summaries

   • Free agents from mundane tasks to focus on customer relationships

2. Entity Recognition

   • Extract structured information from unstructured inputs

   • Particularly useful for complex form-filling

   • Example: First notification of loss in insurance claims

   • Allows agents to focus on empathy rather than data entry

3. Complex FAQ Use Cases

   • Answer questions about complex documents like contracts

   • Translate legal jargon into comprehensible explanations

   • Help agents without specialized training provide accurate answers

Creative Uses of Generative AI as a Tool

1. Training Data Generation

   • Create synthetic data to augment training datasets

   • Help non-native speakers create language examples

   • Support business users in data formatting

   • Improve model robustness (validate before use)

2. Response Phrasing

   • Provide varied ways to express the same information

   • Make bots sound more natural and sophisticated

   • Maintain consistent process flows with varied language

3. Code Generation

   • Streamline development with small code snippets

   • Reduce complexity for non-programmers

   • Enable simple functionality like time-based greetings

Challenges with Generative AI

1. AI Hallucinations

   • Models can generate incorrect information

   • Requires careful validation before customer deployment

   • Agent assist models provide human oversight

2. Cost Implications

   • High computational demands

   • Financial impact

   • Environmental considerations

   • Benefits must outweigh costs

Conclusion

Is AI going to take over customer service? Not completely, at least for now.

While generative AI is transformative, traditional NLU continues to hold significant value. Each case is unique and requires evaluation to determine the best technology approach. Generative AI can also be used as a complementary tool to enhance existing systems rather than replacing them entirely.

The most effective approach combines technologies strategically based on specific use cases and business needs.