The Story
Most AI assistants today are still waiting for you. You open a chat box, you type a request, you hit enter, and the thing does something. It’s reactive by design. A startup called IrisGo wants to flip that around — and it just pulled in $2.8 million to try.
The company came out of stealth-ish mode on May 20, backed by a seed round led by Andrew Ng’s AI Fund, with Nvidia and Google also on the cap table. That’s a notable trio. Ng has spent the last couple of years arguing that the real money in AI isn’t the models themselves but the applications built on top of them, and IrisGo is exactly that kind of bet. Nvidia and Google rarely show up together on a $2.8 million seed unless something about the architecture interests them.
So what is it? IrisGo is a desktop agent — it lives on your computer, not in a browser tab — and its whole pitch is what I’d call “show, don’t tell” automation. You do a task once, with the agent watching. It learns the steps. Then it can repeat that workflow on its own, without you scripting anything or wiring up APIs. The company calls the end goal a “proactive” assistant: software that notices a repetitive chore and offers to take it off your plate, rather than sitting there until you spell everything out.
The founder is Jeffrey Lai, and his résumé is the most interesting tell here. Lai is a former Apple engineer who built the Chinese-language version of Siri. The company name, “Iris,” is literally Siri spelled backward. That’s not subtle, and it’s not meant to be. The message is: the old voice assistant model — wake word, single command, one answer — was a dead end, and the next thing watches your screen and acts.
Under the hood, IrisGo runs what the company describes as a hybrid setup. A lot of the processing happens on-device, leaning on the local NPU or GPU in modern “AI PCs.” Heavier or more complex tasks can get routed to the cloud — but, the company says, only when you explicitly authorize it, and with end-to-end encryption when it does. That on-device emphasis is doing real work in the pitch, because the obvious objection to any agent that watches your entire desktop is: where does all that screen data go?
Quick note on the privacy story: it’s not airtight yet. TechCrunch and the company’s own materials describe the on-device-first design, but at least one trade outlet covering the seed round flagged that the exact split — what stays local, what leaves the machine — hasn’t been fully spelled out. For a tool aimed at office workers handling invoices and internal documents, that detail is going to matter a lot more than the demo reel.
The product ships with a built-in skills library — IrisGo’s site lists 20-plus pre-built workflows across sales, support, finance, and HR, things like email drafting, invoice processing, report building, and document summarization. There’s also a coding assistant bundled in. The target user isn’t developers, though. It’s the knowledge worker who spends a chunk of every day on clerical busywork and has no interest in writing automation scripts. Lai frames the goal as letting humans focus on “high-level conceptual work while agentic systems take care of all the clerical work in the background.”
On availability: the agent is in beta, and there’s a small wrinkle in the coverage worth being honest about. TechCrunch reported a beta for both macOS and Windows. IrisGo’s own website, as of this writing, lists Windows availability now with Mac “coming soon.” Those don’t quite line up — the safe read is that Windows is the live platform and Mac is close behind, but treat the Mac timing as unconfirmed. Either way, IrisGo has lined up a preinstall deal with Acer, meaning the agent could ship already loaded on new PCs, and the company says it’s chasing more OEM deals like it. No public pricing yet.
The Takeaway
“Proactive” is the word doing the heavy lifting here, and it’s worth being a little skeptical of it. We’ve heard versions of this promise before — software that anticipates your needs — and most of it ended up as annoying notifications. The interesting part of IrisGo isn’t the adjective. It’s the mechanism: learning a workflow from a single demonstration, at the OS level, with no API plumbing.
That mechanism is the actual moat, if there is one. Think about how this connects to what we’ve covered before. Back in February, Anthropic acquired Vercept, and we wrote about how that signaled the arrival of “computer-using” AI agents — models that operate a desktop the way a person does, by looking at the screen and moving the cursor. IrisGo is playing in the same field, but it’s solved for a different problem. Anthropic’s approach is about a powerful general model that can be told to do anything. IrisGo’s is about removing the setup friction entirely: you don’t tell it anything, you just let it watch. For a non-technical office worker, that friction is the whole reason desktop automation never caught on. The old enterprise tool for this — RPA, robotic process automation — needed consultants and scripts and broke every time a button moved. “Demonstrate once” is a genuinely different answer to that problem.
Here’s the other thread worth pulling. We’ve spent a lot of time on this blog tracking on-device AI — Google’s Ironwood TPU, Penn’s light-based computing work, the broader push to do inference without a round trip to a data center. IrisGo is a consumer-facing example of why that race matters. An agent that watches everything on your screen is precisely the kind of application you do not want streaming to a cloud server. The on-device-first design isn’t a nice-to-have feature; it’s the thing that makes the product legally and psychologically usable in an office at all. Which is also why the unanswered question about the local-versus-cloud split is the most important thing to watch. If IrisGo can’t draw that line clearly and credibly, the Acer preinstall deal becomes a liability instead of a growth channel.
And the Acer deal itself is the quiet strategic move. IrisGo isn’t trying to win by being the best agent — it’s a $2.8 million seed-stage company, it can’t out-model Google or Anthropic. It’s trying to win by being the agent that’s already on the laptop when you take it out of the box. That’s a distribution play, not a technology play, and historically distribution plays are how the underdog actually survives this kind of fight. The model everyone uses isn’t always the best one. It’s the one that was already there.
My read: IrisGo is a smart, narrow bet on a real gap — the space between clunky enterprise RPA and chatbots that wait for instructions. The “demonstrate once” mechanism is the part worth taking seriously. The “proactive” branding is the part worth ignoring until the beta proves it. Watch two things: how clearly they answer the privacy question, and how many OEMs sign up after Acer. Those, not the demos, will tell you whether this one has legs.
Photo: Steve A Johnson / Unsplash
댓글 남기기