AIConversationTypes
title: AI Conversation Types type: technique created: 2026-05-21 last_updated: 2026-05-21 related: ["radar/techniques/StructuredPromptDrivenDevelopment", "radar/techniques/LLMTDDLoop"] sources: ["https://www.nngroup.com/articles/AI-conversation-types/"] radar_quadrant: Techniques radar_ring: Assess radar_position: inner
A taxonomy of six distinct interaction patterns for generative AI, published by Nielsen Norman Group in 2023, providing a framework for designing LLM-powered product features and evaluating prompt strategies.
| Type | Description | Example |
|---|---|---|
| Retrieval | Extract specific facts from a body of content | "What does this contract say about termination?" |
| Creation | Generate new content from a prompt or template | "Write a product description for this item" |
| Summarisation | Condense longer content into a shorter form | "Summarise this meeting transcript" |
| Ideation | Brainstorm options, alternatives, or approaches | "What are 10 ways to reduce churn?" |
| Classification | Categorise or label content by defined criteria | "Is this support ticket a billing issue or a bug?" |
| Transformation | Convert content from one format or style to another | "Translate this JSON schema into a Markdown table" |
Each type has distinct requirements for prompt structure, output validation, and failure modes:
- Retrieval needs grounding in source content; hallucination risk is high without RAG or citations.
- Creation benefits from few-shot examples and explicit constraints (tone, length, format).
- Summarisation requires faithfulness metrics; key information loss is the primary failure mode.
- Ideation works best with broad prompts and iterative refinement; over-constraining reduces output diversity.
- Classification is the most automatable type; output is bounded and verifiable against a defined label set.
- Transformation requires structural validation of the output format, not just content review.
Knowing which type a feature implements shapes the entire design: prompt engineering approach, human review requirements, confidence thresholds, and failure handling.
The NNG taxonomy is a design tool, not just a classification exercise. When building an LLM-powered feature, identifying the conversation type first clarifies:
- Whether the task is automatable end-to-end or requires human review in the loop.
- Which prompt patterns apply (chain-of-thought for ideation, structured output for classification).
- What success looks like and how to measure it.
A feature mixing multiple types (e.g., retrieval + summarisation for a document QA system) should treat each stage separately, with appropriate validation at each boundary.
AI Conversation Types sits in the Assess ring of the Techniques quadrant, at inner position. First studied via the Nielsen Norman Group article (2023-11-25). The taxonomy is immediately applicable as a design tool for any LLM feature work — no tooling required, only the discipline of identifying the conversation type before writing prompts. Inner position reflects direct relevance to any project using Claude or other LLMs and strong complementarity with radar/techniques/StructuredPromptDrivenDevelopment, which provides the REASONS canvas for structuring prompts once the type is identified.
