Adept AI

The Promise

Adept AI emerged in April 2022 with a vision that captured the imagination: AI that could actually do things, not just talk about them. Founded by David Luan (former VP of Engineering at OpenAI), the company promised to build “AI teammates” that could operate software the way humans do—clicking buttons, filling forms, navigating interfaces, accomplishing tasks.

The pitch distinguished Adept from the chatbot crowd. While ChatGPT and its competitors generated text, Adept’s “Action Models” would generate actions. Tell it to “book me a flight to New York next Tuesday,” and the AI would actually navigate Expedia, select flights, and complete the booking. It was the promise of a true AI assistant—one that could do your work, not just discuss it.

The team was stellar. Besides Luan, co-founders included researchers from Google Brain and DeepMind. These were the people who had built transformer architectures and trained frontier models. If anyone could make AI agents work, it was this group.

The Rise

Investors believed. Adept raised $65 million in a Series A led by Greylock, with participation from Addition and a16z. Then came a $350 million Series B at a valuation reportedly exceeding $1 billion. Total funding reached $415 million. For a company with no shipped product, the numbers were staggering—a testament to the team’s pedigree and the allure of the agent vision.

Adept released demos that looked like magic. Videos showed AI navigating websites, operating enterprise software, and completing multi-step workflows. The “ACT-1” model could use a web browser like a human, understanding interfaces and taking appropriate actions.

The company positioned itself squarely at the enterprise market. The pitch to CIOs was irresistible: automate the tedious work that consumed knowledge workers’ days. Every enterprise had thousands of employees doing repetitive tasks that Adept’s AI could theoretically handle.

The Fall

But the demos didn’t become products. The gap between impressive videos and reliable, shippable software proved vast. AI agents that could navigate arbitrary websites made for great demos but failed unpredictably in production. Websites change layouts. Edge cases multiply. Error handling becomes exponentially complex.

The enterprise sales cycle was brutal. Companies wanted reliability guarantees that Adept couldn’t provide. The competitive landscape shifted as OpenAI and others announced their own agent capabilities. The window for Adept to establish market leadership was closing while the product remained stuck in demo mode.

The acqui-hire came in June 2024. Amazon hired David Luan to lead its “AGI Autonomy” team, along with most of Adept’s key employees. The deal was structured as a licensing agreement plus talent acquisition, with Amazon making investors whole using funds paid as licensing fees.

Adept nominally continues to exist, pivoting to focus on building agents for a narrower set of use cases. But the original vision—the universal AI teammate that could operate any software—died when the founders walked out the door to Amazon.

Warning Signs

• Demo-to-production gap: Impressive demos don’t guarantee shippable products, especially with agent systems
• Reliability requirements: Enterprise customers need 99%+ reliability; agent systems operating on arbitrary interfaces can’t provide it
• Competitive timing: The window between innovation and commoditization in AI is measured in months, not years
• Founder dependency: The company’s value was overwhelmingly in the team, making acqui-hire the likely exit
• AGI positioning: Pitching “AI teammates” implied near-AGI capabilities that couldn’t be delivered on VC timelines

Epitaph

🪦 Promised to act, but only acted as acqui-hire bait