Large Action Model (LAM)

A large action model is an AI system that goes beyond generating text to actually performing actions in software. While a large language model writes you instructions for how to book a flight, a large action model opens the booking website, searches for flights, selects options, and completes the purchase.

From words to deeds

Traditional LLMs are trained on text and produce text. LAMs are trained on sequences of actions — mouse clicks, keystrokes, menu selections, API calls — and produce action sequences. They learn by observing how humans interact with software and then replicate those interaction patterns.

How LAMs work

A LAM typically combines several capabilities:

• Screen understanding: The model can interpret what is displayed on a screen, identifying buttons, forms, menus, and content.
• Action planning: Given a goal like "schedule a meeting with Sarah for next Tuesday at 2pm," the model breaks it down into a sequence of interface actions.
• Execution: The model carries out those actions, either by controlling a browser, calling APIs, or simulating user input.
• Error recovery: When something unexpected happens (a pop-up appears, a page layout changes), the model adapts its plan.

Current examples

• Browser agents that can navigate websites and complete multi-step tasks.
• Computer-use models that control desktop applications.
• Mobile assistants that interact with apps on your phone.
• Robotic process automation systems enhanced with AI understanding.

LAMs vs AI agents

There is significant overlap between LAMs and AI agents. The distinction is subtle: LAMs specifically emphasise learning action patterns from human demonstrations, while AI agents is a broader term for any AI system that takes autonomous actions. In practice, most advanced AI agents incorporate LAM-like capabilities.

Challenges

• Safety: An AI that can click buttons and fill forms can also make mistakes with real consequences — sending wrong emails, making incorrect purchases, or modifying data.
• Reliability: Software interfaces change frequently, and LAMs must handle these changes gracefully.
• Trust: Users need confidence that the model will do exactly what they intend, especially for high-stakes actions.