How Adept AI Built Actionable Models by Redefining Interfaces

TL;DR

• Challenge: Large language models could write code and text but could not operate software or take actions on behalf of users.
• Solution: Adept built a multimodal model designed specifically to interact with graphical user interfaces.
• Results: They secured massive funding and enterprise interest before pivoting their core tech.
• Investment/Strategy: Betting early on agentic workflows and computer vision for GUI navigation.

The Problem

Before Adept AI entered the scene, large language models were confined to chat boxes. You could ask a model to write a Python script, summarize a document, or draft an email, but you could not tell it to open Salesforce, find a lead, and update their contact information. Users were forced to copy and paste outputs across dozens of tabs. This friction kept AI as a passive assistant rather than an active agent.

Enterprise workflows are complex and graphical. Tools like Jira, Salesforce, and Figma do not easily map to simple text commands. Founders and developers recognized that the next frontier of AI was not just generating content, but actually doing work.

The Execution & GTM Strategy

The Technical Product Moat

Adept AI realized that APIs were not enough. Not every software product has a clean API, and human workflows span multiple applications. Instead of relying on backend integrations, Adept trained models to see the screen and use the mouse and keyboard like a human. This approach created a massive moat. By building ACT-1, they demonstrated a model that could interpret visual interfaces directly.

The Results & Takeaways

• Raised $350 million in Series B funding at a valuation over $1 billion.
• Brought enterprise attention to agentic AI workflows.
• Demonstrated the viability of UI-based interaction models.

What a small startup can take from them: Stop waiting for perfect APIs. If you want to automate workflows, build tools that interact with the software users already use. Meet the user where they are, even if it means parsing complex visual interfaces.

Frequently Asked Questions