# Inference

> Inference is the process of running a trained AI model to produce an output, such as generating a reply, classifying a ticket or extracting fields from a document. Training builds the model once; inference happens every time the model is used. When you call a model provider's API, you are paying for inference, usually billed by the number of input and output tokens processed.

## Why it matters for your business

For most businesses, inference is the ongoing operating cost of AI, separate from the one-time cost of building the system. It scales with volume and with the size of the model and prompts, so a workflow that is cheap in testing can become expensive at production volume. Inference choices also affect speed and data handling: a larger model may be more accurate but slower, and running inference on your own hardware changes both the cost structure and who sees the data. Estimate inference cost per transaction before committing to a design.

## Example

A law firm runs every incoming client inquiry through a model to classify matter type and urgency, which is one short inference call per email. When the team adds a second call that drafts a full reply for each inquiry, the monthly inference bill rises sharply, because drafting produces far more output tokens than classifying. They limit drafting to the most common matter types.

Related terms: https://www.lumeor.studio/glossary/token/, https://www.lumeor.studio/glossary/latency/, https://www.lumeor.studio/glossary/model-routing/, https://www.lumeor.studio/glossary/self-hosting/, https://www.lumeor.studio/glossary/llm-large-language-model/

---

Source: https://www.lumeor.studio/glossary/inference/