# NLU or LLM? And which model tier? (cheat sheet)

> Two quick decision guides on one page.

Source: https://conversationalaidesign.com/cheatsheets/nlu-vs-llm-and-model-tier.html. Author: Nelson Abacherli Kawakami. Reviewed October 27, 2026.

## NLU or LLM
| Question | Points to NLU or rules | Points to LLM |
|---|---|---|
| Range of requests | A few known tasks | Open-ended |
| Exact wording or audit | Yes | No |
| Languages | One or two | Many |
| Labeled data | Plenty | None |
| Cost or latency pressure | Extreme | Normal |
| Topic change | Rare | Frequent |

## Usual answer
- Hybrid: rules or NLU for precise steps, an LLM for the long tail
- A small classifier in front: confident known intents go deterministic, the rest go to the model
- Always keep a confidence threshold and a handoff

## Model tiers
| Tier | Use for |
|---|---|
| Flagship | Hard turns, planning, agents, judging |
| Mid-tier | A sensible default |
| Small and fast | Routing, classification, voice |
| Open-weight | Data control, very high volume |
| On-device | Offline and edge |

## Always
- Pick the cheapest model that passes your tests
- Measure cost per resolved conversation
- Pin versions and keep a tested fallback

Related modules: From intents to LLMs (https://conversationalaidesign.com/modules/02-from-intents-to-llms.html), The LLM landscape: choosing a model (https://conversationalaidesign.com/modules/04-llm-landscape.html)