> For the complete documentation index, see [llms.txt](https://jamiewen00.gitbook.io/ai-engineering-handbook/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://jamiewen00.gitbook.io/ai-engineering-handbook/building-ai-products/ai-product.md).

# AI Product

Partial Autonomy Applications

## The Autonomy Slider

Andrej Karpathy created the <mark style="background-color:purple;">autonomy slider</mark>. It classifies AI products by their level of autonomy.

<figure><img src="/files/w9aqmYsjJo3ApkgOCyAp" alt="A slider showing the range of AI product autonomy, from AI-enhanced to AI-first."><figcaption></figcaption></figure>

* **AI-Enhanced Products**: Add LLM features to an existing product. The core workflow stays the same. Examples: Miro, Jira, Slack, Google Docs.
* **AI-First Products**: Design the UX around the LLM or agent. Examples: ChatGPT, Claude, Devin.

<figure><img src="/files/BTjKkiLrI6XbHpLKoTPt" alt="Diagram comparing AI-enhanced and AI-first product concepts."><figcaption></figcaption></figure>

## AI Value Ratio

Another angle to look at AI products is from the perspective of the **AI Value Ratio**: the ratio of human input to (valuable) AI output. This concept was introduced by [swyx](https://www.swyx.io/about) in the [AI Engineer World's Fair 2025 Keynote](https://www.youtube.com/live/z4zXicOAF28?si=U5N51b3TPC1geFqU\&t=1928). It is similar to the concept we use in product design, least amount of effort to achieve the most value.

* 1:0.5 - Copilot: suggesting code as you type
* 1:1 - Chatbot: returning equivalent value to your prompt
* 1:10 - Reasoning models, workflows: multiply your input through workflows
* 1:10000 - Deep Research, NotebookLM: generate comprehensive insights from simple queries

> 💭: A chatbot demands a paragraph but returns one line. Worse ratio than traditional products. Why would users use it?

## AI as the Engine

AI can act as an engine, adding LLM power to an existing product. Here, AI is a key component, not the whole product.

Simon Wardan's YOW! 2025 talk explores this idea. [View slides](https://australian.software/YOW2025/).

<figure><img src="/files/die8WTaXTr5LPyRXbQfc" alt="Visual representation of AI as the engine of a product."><figcaption></figcaption></figure>

## AI as an Assistant

This pattern deeply integrates LLM features into a traditional UI. It creates tight interaction between the UI and the LLM.

Examples include code assistants that are tightly integrated into the development environment.

<figure><img src="/files/k9Fiia861kjAO1h9g0Tm" alt="Lex.page, an example of AI as an assistant in a word processor."><figcaption></figcaption></figure>

<figure><img src="/files/iRtDkMHKhJNWDzsLPbMN" alt="Cursor, an example of AI as an assistant in an IDE."><figcaption></figcaption></figure>

## AI as an Agent

In this model, the UI is built around the agent. See this Manus replay for an example:

{% embed url="<https://manus.im/share/QfmleZqV5JDSk5uPYijvt2?replay=1>" %}