Chatbot Conversation Writing

Chatbot Conversation Writing

Overview

A chatbot is a conversation, not an interface. Every turn the bot takes should advance the user’s goal or surface a clear next move. The work breaks into three layers:

1. Persona — who the bot is, what it sounds like, what it refuses to do.
2. Turn design — the individual response, plus the choices it offers.
3. Recovery design — what happens when the bot doesn’t understand, when the user is angry, or when the conversation has to leave the bot.

Core Concepts

1. Cooper-Reeves conversational design

The four Gricean maxims:

• Quantity — say as much as is needed, not more
• Quality — only say true things
• Relation — stay on the user’s topic
• Manner — be clear, brief, and orderly

2. Persona consistency

A minimum persona doc includes:

• Name (or explicit no-name policy)
• Role (what the bot is for — support? sales? in-product help?)
• Voice traits (3–5 adjectives — e.g., “warm, precise, never cute”)
• Refuses to do (e.g., diagnose medical issues, take payment, promise pricing)
• Vocabulary (3–5 words it uses; 3–5 it doesn’t)
• Disclosure boilerplate (the exact “I’m an AI assistant” line)
• Escalation trigger phrases

3. The “no, I’m not human” disclosure rule (2024–2026 norms)

As of 2026, plain-and-unambiguous AI disclosure is no longer optional:

• California SB 243 (effective Jan 1, 2026) requires disclosure, plus reminders every three hours for minor users
• EU AI Act requires upfront disclosure for any AI system interacting with natural persons
• FTC treats undisclosed AI as potentially deceptive

Writing implications:

• Disclose in the first turn, not buried in a tooltip
• Use direct language: “I’m an AI assistant” — not “I’m an enhanced automation experience”
• If asked “are you a human?” answer plainly: “No, I’m an AI assistant.”

4. Turn-taking conventions

A conversational turn has three jobs:

1. Acknowledge what the user just said
2. Resolve the user’s intent
3. Hand the turn back with a clear next move

Length convention: 1–3 short sentences per turn for support bots, 1–5 sentences for in-product help.

5. The graceful-confusion pattern

When the bot doesn’t understand, three things must happen:

1. Admit it cleanly. “I’m not sure what you mean by ‘reset’.”
2. Offer a small, finite menu. Two or three concrete interpretations.
3. Provide a path out. Always include “talk to a human” no later than the second failed understanding.

6. Fallback hierarchies

7. Brand-voice-in-bot transfer

Three transfer rules:

• Drop the edge in failure modes. Humor in a working flow is fine. Humor when the user is locked out is not.
• Keep the rhythm, simplify the vocabulary.
• Reuse hero phrases sparingly. One signature phrase per conversation.

8. Escalation-to-human prose

Three elements must be present:

1. Confirm the handoff is happening. “I’m connecting you to a person now.”
2. Preserve context. “I’ll share what we’ve discussed so you won’t have to repeat yourself.”
3. Set the wait expectation. “Average wait is about 4 minutes.”

Anti-Patterns

• The “I’m an enhanced automation experience” dodge. Coy disclosure reads as deceptive.
• The infinite “I didn’t understand, please rephrase” loop. Two failures is the ceiling; escalate.
• The hidden escalation path. Surface the escalation in the conversation itself.
• The promise-the-bot-can’t-keep. “I’ll refund you” when refund authority is human-only.

References

• Erika Hall, Conversational Design (A Book Apart, 2018).
• California SB 243 (Companion Chatbot Disclosure Act), effective Jan 1, 2026.
• EU AI Act Article 50 (transparency obligations for AI systems).
• Byron Reeves and Clifford Nass, The Media Equation (CSLI / Cambridge, 1996).