AI actions transform LLMs from information processors to active participants in enterprise systems and enable completely new and practical use-cases for generative AI.

Generative AI primarily focuses on retrieving and understanding information, aiding in tasks like coding, marketing, and sales. However, its potential extends beyond just generating and analyzing data. The future of AI lies in its ability to interact with real-world systems. These applications are often called AI agents. While we’re familiar with AI applications like ChatGPT, which can generate images or perform web search, the next step is to integrate AI into enterprise systems, enabling it to not only instruct but to complete tasks on behalf of users.

AI actions: Bridging the automation gap in enterprise systems

To safely and effectively bring AI into enterprise environments, we propose a new paradigm: “AI actions.” AI actions consist of capabilities, workflows, and automation that can be securely exposed to AI agents. These actions are defined by their inputs and outputs, described in natural language, enabling AI apps to comprehend and utilize them. Actions are pre-defined code snippets executed in a controlled environment, equipped with access rights, audit trails, and necessary controls.

While actions share similarities with tools and function calling in frameworks like Langchain, Llamaindex, and Microsoft Copilot Studio, they provide additional benefits. Actions can enforce strict user approval before execution and operate within a dedicated runtime separate from the AI app, ensuring all required controls, authentication, and access rights are in place. Actions can range from simple tasks like fetching data over an API to complex operations such as creating a new purchase order in SAP. They can be synchronous or asynchronous, stateful or stateless.

The benefits of an open source approach to intelligent automation

Using pre-defined AI actions instead of generating code on the fly has a number of benefits, including:

• Ensuring the accuracy of outcomes
• Understanding and strictly enforcing the desired behavior
• Reusing actions across multiple AI apps
• Enforcing human validation for actions that have consequence
• Managing authentication to 3rd party systems
• Achieving fast and scalable performance
• Providing observability to actions

This approach also addresses the needs of enterprise decision-makers who emphasize the importance of robust AI integration into their systems. As part of our deep commitment to innovation and collaboration, Sema4.ai made our AI Action framework open source, inviting developers to contribute and expand its capabilities. Find the repository and quick start guides here. You will notice that the AI actions framework is still today named under Robocorp, which we acquired early 2024.

AI actions represent a significant leap in AI’s evolution, offering a controlled and efficient way to integrate AI into enterprise systems. This new paradigm paves the way for AI to move from an advisor to an actor in the business world, harnessing its full potential in practical and impactful ways.

Next up, intelligent AI agents and actions

While AI actions are an important part of the new enterprise AI agent stack, our work extends much broader. We are combining the best of open source and soon releasing the easiest way to build AI agents, right at your desktop.