The rise of Large Action Models (LAMs) promises to revolutionize enterprise automation, but significant challenges lie ahead. This post explores the potential and pitfalls of this emerging technology.

The Promise of Large Action Models

Large Action Models (LAMs) represent the cutting edge of AI technology. They aim to automate complex enterprise tasks by learning from user actions. Companies like Adept, H-Company, and Orby are pioneering this field, attracting over $500 million in funding in the past year alone.

Why LAMs Matter

1. Automation potential: LAMs could address the inefficiencies in legacy SaaS systems
2. Cost savings: Enterprises spend $1 trillion on software annually, with flat revenue per employee
3. Reduced manual work: LAMs aim to minimize meaningless tasks performed by humans

The Self-Driving Car Analogy

Implementing LAMs in enterprise environments is often compared to developing self-driving cars. However, the enterprise landscape presents unique challenges:

• Complex systems: Enterprises use a mix of modern and legacy technologies
• Regulatory hurdles: GDPR and the EU AI Act create additional obstacles
• Organizational resistance: IT departments and executives may resist integration

Data and IP Challenges

LAMs face significant hurdles in data acquisition and intellectual property rights:

1. Data complexity: Processing diverse data types (images, videos, process data) is expensive
2. Scale requirements: Building effective LAMs requires vast amounts of data
3. Data ownership: Multiple vendors control enterprise data, complicating access
4. Legal considerations: Determining what data can be legally used is a major challenge

The Human Factor

Human involvement remains crucial in developing and implementing LAMs:

• Non-linear thinking: Humans possess knowledge and decision-making processes that are difficult for machines to replicate
• Feedback loops: High-quality human feedback is essential for improving LAM performance
• Context-dependent labeling: Enterprise use cases require specialized knowledge for effective data labeling

Apple’s Approach: Generating vs. Understanding

Apple’s recent AI strategy offers an alternative perspective:

1. Ecosystem integration: Leveraging hardware-software synergy on mobile devices
2. App Intents: Encouraging developers to add LLM-friendly handles to their apps
3. Generating vs. understanding: Creating AI-friendly environments rather than trying to automate existing systems fully

The Future: Composable Work Blocks

An emerging approach to enterprise automation focuses on generating tailored solutions:

1. Problem-specific generation: Creating AI solutions for contained problems
2. Composability: Developing “work blocks” that can be combined for complex tasks
3. Human Validation: Ensuring trust in AI-generated outputs through human oversight

Conclusion

While Large Action Models hold immense potential for enterprise automation, the path forward is complex. By learning from approaches like Apple’s and focusing on composable, human-validated solutions, we can work towards more efficient and effective AI-driven automation in the enterprise world. What are your thoughts on the future of AI in enterprise automation? Share your insights in the comments below!

References:

1. https://www.adept.ai
2. https://www.orbyai.com
3. https://www.h-company.com
4. https://gdpr-info.eu/
5. https://www.eurocompanyformations.com/blog/the-eu-ai-act-clears-its-final-hurdle-a-landmark-for-global-ai-regulation/
6. https://www.apple.com/newsroom/2024/06/introducing-apple-intelligence-for-iphone-ipad-and-mac/