liveBook · Manning

Hybrid approach

As explained in Effective Conversational AI

A hybrid approach combines generative AI capabilities with human-builder control in the creation or improvement of conversational AI systems.

When a conversational AI repeatedly misunderstands users, improving its intent data manually offers control but demands time and specialized skill. A hybrid approach uses generative AI as a partner while keeping the human builder responsible for the final result. This reduces effort, expands available test data, and limits hallucination risk. section 7.0

Use the approach to balance complementary strengths:

  • Have generative AI suggest vocabulary, utterances, or intent examples that a builder reviews and selects. section 7.0
  • Retain purpose-built conversational AI when speed, low cost, and controlled responses matter. section 7.1
  • Combine retrieval-augmented generation with contextual analysis when accurate information alone does not resolve the user’s deeper need. section 9.3.2

Password-reset support

An IT help-desk assistant needs a strong password-reset intent but has no production utterances. Instead of relying only on “forgot password” and “cannot login,” the builder asks generative AI to imagine broader user phrasing, then reviews and incorporates suitable examples. The builder—not the model—decides what enters the training data. section 7.0

Table 7.1 Comparing and contrasting traditional natural language processing (NLP) in conversational AI and generative AI on the classification task
Feature
Traditional NLP
Generative AI
Model
Purpose-built for excellence at only one task: classification
Generalized model good at many tasks
Runtime speed
Fast
Slow
Runtime cost
Low
High
Accuracy
Mostly accurate (trained by you on small data)
Mostly accurate (pretrained on huge data)
Scalability
Manageable for up to 100 intents; very difficult afterward
Generalizes very well via RAG pattern
Controllability
Strictly controlled by humans; requires extensive testing
Prone to hallucinate when given full control; hallucinations are hard to detect automatically

Compare purpose-built conversational AI and generative AI across speed, cost, accuracy, scalability, and controllability before choosing where each belongs. section 7.1

How Effective Conversational AI covers it

How deep it goes
Explains the trade-offs and illustrates human review, but stops short of specifying a complete implementation architecture.
Examples
Uses an IT help-desk password-reset intent and contrasts factual retrieval with context-aware assistance for confused users.
What it assumes
Assumes familiarity with conversational AI intents and introduces the balance between manual training, generative AI, and RAG.
What you can do after
Apply generative AI to expand conversational data while reviewing outputs and preserving controlled responses.
How it connects
Connects generative AI with intent training, retrieval-augmented generation, traditional NLP, and contextual analysis.
Cover of Effective Conversational AI
liveBookmanning.com
Read Effective Conversational AI on liveBook Get the book at manning.com

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Effective Conversational AI. The section links point to the parts of the book it draws on.