# Foundations of conversational AI

> What an AI assistant actually is, what makes conversation hard to design, and how to tell whether one is working.

Source: https://conversationalaidesign.com/modules/01-foundations.html. Author: Nelson Abacherli Kawakami. Reviewed October 27, 2026.

## What you will learn
- Tell menu bots, intent bots, LLM assistants, agents and voice assistants apart
- Explain why conversation is a design problem, not just a language problem
- Name the three jobs an assistant does: answer, act and guide
- Pick success metrics that don't reward bad behavior


## What counts as conversational AI in 2026

Conversational AI is software you talk or type to, in natural language, to get something done. The label covers very different systems, and your first design decision is which kind you are building.

| Kind | How it works | Strength | Typical failure |
|---|---|---|---|
| Menu or IVR tree | Fixed options, little or no language understanding | Predictable and cheap | Dead ends and endless menus |
| Intent-based assistant | Classifies each message into a known intent and fills slots | Reliable for narrow tasks | "Sorry, I didn't understand" |
| LLM assistant | A language model writes each reply from instructions and context | Handles messy, open language | Confident invention |
| Agent | An LLM that can call tools and take multi-step actions | Gets real work done | Wrong or unsafe actions |
| Voice assistant | Any of the above over speech | Hands-free and fast | Latency, mishearing, no screen to fall back on |

Most real products combine several of these. A good assistant might use a deterministic flow for payments, retrieval for policy questions and an agent for open-ended help, all behind one conversation.

## Why conversation is a design problem

People bring expectations from talking to other people. The philosopher Paul Grice described them as four cooperative principles, and they make a useful design checklist:

- **Quantity:** say as much as needed, no more.
- **Quality:** say only what you believe is true.
- **Relation:** be relevant to what was asked.
- **Manner:** be clear, brief and orderly.

**A failure of relevance**
> **User:** Can I bring my dog on the train?
> **Bot:** Our trains run every 15 minutes between 6 am and midnight.

Nothing in that reply is false, yet it fails the user. Language models add a new risk to the list: they can break the *quality* principle fluently, with perfect grammar and total confidence. Conversation also needs turn-taking, shared context and **repair**, the ability to notice a misunderstanding and fix it. We return to all of these in module 3.

## The three jobs of an assistant

Almost every assistant does some mix of three things, and each one has a different risk profile.

| Job | Example | Main risk | Where we cover it |
|---|---|---|---|
| **Answer** | "What is your baggage policy?" | Misinformation | RAG, module 6 |
| **Act** | "Cancel my order." | Irreversible mistakes | Tools and MCP, module 8 |
| **Guide** | "Help me choose a plan." | Friction and bad advice | Orchestration, module 9 |

Knowing which job a feature performs tells you how much safeguard it needs. A wrong answer can be corrected in the next message. A wrong cancellation may not be reversible.

## What good looks like

Choose metrics that reflect the user's outcome, not only the company's savings:

- **Resolution rate:** was the problem solved, confirmed by the user or by the absence of a repeat contact?
- **Repeat contact:** did the same person return about the same issue within a week?
- **Escalation quality:** when the assistant hands off, does the person receive context, and is it quick?
- **Satisfaction:** a simple rating plus a free-text reason.
- **Cost per *resolved* conversation:** not cost per conversation.
- **Safety incidents:** policy violations, harmful outputs, wrong actions.

> **Containment is not success**
> "The bot handled 70% of chats" sounds like a win, but those chats may include people who gave up. The Klarna case study shows a company that celebrated containment and later admitted that quality had suffered.

## How the course fits together

A modern assistant is built in layers. Each module covers one.

- Users and goals → Conversation design → Model choice → Knowledge (RAG) → Tools (MCP) → Orchestration → Evaluation → Safety

## Try it: write the transcript first

Pick a task you know well, such as returning a purchase or rescheduling an appointment. Write three conversations before you think about technology:

1. A **happy path** where everything goes smoothly.
2. A **messy path** where the user is vague, changes their mind or gives two answers at once.
3. A **failure path** where the assistant can't help and must hand off.

Keep them. You will reuse them in module 3 to write your system prompt and in module 14 as the start of your test set.


## Key takeaways
- Decide which kind of assistant you are building before you choose technology.
- People expect a cooperative partner: informative, truthful, relevant and clear.
- Answering, acting and guiding carry different risks and need different safeguards.
- Containment is not resolution. Measure whether the problem was actually solved.

## Sources
- AI Incident Database: real assistant failures: https://incidentdatabase.ai
- Case study: Klarna's AI assistant: https://conversationalaidesign.com/case-studies/klarna-ai-assistant.html