IrisGo closed a $2.8M seed round led by Andrew Ng's AI Fund to build a desktop AI agent that learns user workflows once and then automates them without further prompting. Nvidia and Google also backed the round, which closed earlier this year. The company is positioning the product as a proactive automation layer for knowledge workers — a system that anticipates repetitive tasks and handles them autonomously.
The core mechanic is watch-once-repeat-forever: a user performs a task on their desktop, Iris records the steps, and the agent then replicates the workflow on command or automatically. Co-founder Jeffrey Lai, a former Apple engineer who built the Chinese language version of Siri, demonstrated the system by teaching Iris to place a coffee order at Philz Coffee. Iris recorded the steps — selecting a latte, filling out payment details, clicking purchase — and then repeated the order independently when asked.
The agent ships with a built-in skills library covering email drafting, invoice processing, report building, and document summarization. It also includes a coding assistant similar in function to OpenAI's Codex or Anthropic's Claude Code. The system observes desktop behavior and adds new workflows to its action-item list as it learns.
Key facts
- 01IrisGo closed a $2.8M seed round led by Andrew Ng's AI Fund, with backing from Nvidia and Google.
- 02The desktop agent learns workflows by watching the user perform a task once, then automates it without further prompting.
- 03Jeffrey Lai co-founded IrisGo after working at Apple on the Chinese language version of Siri.
- 04IrisGo recently launched beta apps on macOS and Windows and struck a preinstall deal with Acer.
- 05The system processes data on-device for privacy, with cloud processing only when explicitly authorized using end-to-end encryption.
IrisGo processes much of its data on-device rather than in the cloud, a design choice aimed at stronger privacy protections. The architecture is hybrid — larger or more complex tasks route to the cloud — but the company says cloud processing only occurs when explicitly authorized by the user and uses end-to-end encryption. That on-device bias is a competitive differentiator in a market where most AI agents rely heavily on cloud compute.
Lai secured the AI Fund's lead by leveraging a shared connection with Ng through Carnegie Mellon University, where both are alumni. Ng co-founded Google Brain, one of the formative deep learning research teams, and his backing has lent credibility to IrisGo's pitch. The startup recently launched beta versions of its macOS and Windows apps.
IrisGo is also pursuing preinstall deals with laptop manufacturers. The company struck a deal with Acer to bundle the agent on new devices, and Lai said the goal is to replicate that arrangement with other OEMs. Preinstalls would give IrisGo distribution at scale without relying on user acquisition spend.
The product arrives as the AI industry shifts toward proactive agents — systems that anticipate needs and act before the user articulates a request. IrisGo's bet is that white-collar workers spend significant time on repetitive tasks that could be automated if the agent could learn the pattern once and run it indefinitely. The question is whether the watch-once mechanic proves robust enough for complex multi-step workflows across different applications.
Ng's backing and the Acer deal give IrisGo credibility, but the market for desktop AI agents is already crowded. On-device processing is a selling point for privacy-conscious enterprises, but hybrid architectures still introduce cloud dependencies that may limit adoption in regulated industries. The company's success will hinge on whether its automation holds up across the messy reality of desktop workflows — browser tabs, file systems, internal tools — without breaking or requiring constant user correction.
IrisGo is betting that knowledge work is still too manual even with frontier models in the loop. If the agent can deliver on its promise of background automation, it could carve out a niche. If the watch-once mechanic proves fragile, it risks becoming another AI demo that works in a controlled environment but falters in production.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
