Summary: TechCrunch reported on May 20, 2026 that IrisGo raised a $2.8 million seed round led by Andrew Ng’s AI Fund. Its Iris desktop agent watches workflows, learns from demonstration, and acts across apps, with most inference on the PC’s NPU or GPU. The company has an Acer preload deal, while cloud processing is described as requiring explicit user authorization and end-to-end encryption.
TechCrunch turned its spotlight on IrisGo on May 20, 2026: a startup whose $2.8 million seed round was led by Andrew Ng’s AI Fund, and whose ambition is the AI desktop buddy you never knew you needed. The software goes by Iris — Siri spelled backward — observing what happens on your desktop, picking up the workflows you repeat every day, then carrying them out with little to no prompting.
Everything hangs on the promise “demo once, and it knows.” In a hands-on demo, it watched how to order a latte from Philz Coffee — payment details entered, checkout completed — and from that point on Iris could run the whole sequence on its own. Co-founder Jeffrey Lai, who worked as an engineer at Apple and helped build the Chinese-language version of Siri, doesn’t sugarcoat the target market: “Our target audience is knowledge workers — white-collar companies.”
Learning Through Observation, Then Taking Over
Under the hood the engine moves through three stages: Watch sweeps up context across windows, apps, and systems; Learn converts observed behavior into workflows and decision logic with no manual setup required; Act runs flows that are event-driven, cross-app, and self-optimizing. Demonstration learning aside, the product arrives with a built-in skills library spanning email drafting, invoice processing, report building, and document summarization, and the official site catalogs roughly 20 importable community skills across categories such as Email Triage, Lead Scoring, and Contract Summary. A coding assistant that stands shoulder to shoulder with OpenAI’s Codex and Anthropic’s Claude Code comes as part of the package too. More telling than the latte is the company’s own demo workflow: a Morning Brief Builder that digests Hacker News, summarizes a podcast episode, checks stock watchlists on Yahoo Finance, adds groceries on Instacart, and drafts a Gmail summary — one event-driven flow reaching across five applications.
That demonstrate-don’t-prompt arrangement runs parallel to Thinking Machines’ research on interaction models, which trains live interaction ability directly into the model itself — and both point the same way: the interface to AI is moving from typing prompts to demonstrating behavior.
Local Inference With Privacy Built Into the Design
The pitch is “proactive” AI — needs anticipated before you ever voice them. Pulling that off means dwelling beside your desktop for long stretches, which turns privacy architecture into a core product decision rather than a footnote. The bulk of inference happens locally, drawing on your PC’s NPU and GPU; heavier tasks take a hybrid cloud path, though the company is explicit that cloud processing “only occurs when explicitly authorized by the user and uses end-to-end encryption.” The Windows beta can be downloaded today, the Mac version carries a coming soon label on the official site, and by TechCrunch’s account the betas for both platforms have launched.
The Seed Round and the AI PC Play
Andrew Ng’s AI Fund anchored the $2.8 million seed, and NVIDIA Inception, Google for Startups, Pegatron Ventures, and Intel also appear on the supporter list; the tie to Ng traces back to a shared Carnegie Mellon University network. Lai built the company together with co-founder Yichen Chu, who holds the COO seat, under Irixion Corp. Distribution is the more interesting move: a signed deal with Acer puts the app on new devices as a preload, and IrisGo hopes to close comparable agreements with other laptop makers. AI PCs need a killer app before the NPU justifies itself, and that open slot is exactly where IrisGo is placing its bet — system-level digital butler infrastructure for the AI PC era rather than one more chat window.
Rivals and the Hazards Ahead
IrisGo has company in the “watches you work” desktop-agent lane, but it laces three pieces together: learning by demonstration, on-device inference, and OEM preload. The hazards are just as legible. An agent that reads and buys in your name makes mistakes with real price tags — a payment flow learned wrong costs far more than a chat reply gone wrong. Trust has to be earned step by step: begin with low-risk skills like summarizing and drafting, then move toward high-risk actions like paying and ordering, and how permissions and confirmations are designed will decide whether products like this survive. The platform question hangs over it all: Microsoft, Apple, and Google each ship system-level assistants of their own, and a preloaded third-party agent lives or dies by how much surface the OS vendors leave it. The privacy claims, meanwhile, cannot stand on a website statement forever; independent scrutiny is on its way.