IrisGo Wants to Learn from Your Screen and Do the Work for You

Andrew Ng rarely backs ventures without good reason. That his investment fund has put money into IrisGo—a startup that, according to TechCrunch, positions itself as an "AI butler" for the desktop—is a signal worth paying attention to, even if the concept sounds like something we've seen before.

What sets IrisGo apart from the flood of productivity assistants is its learning mechanism: the agent observes in real time what happens on the user's screen and, through continuous observation, learns to replicate and automate those same tasks. No manual configuration, no explicit workflows. The company's co-founder sums it up this way: the system simply watches and learns.

What IrisGo Actually Does

The product operates as a persistent background process. It captures visual context from the desktop—which applications are open, what actions the user performs, in what order—and builds a personalised behaviour model. Over time, it can anticipate repetitive steps and execute them autonomously or propose their automation.

The concept isn't new in its broadest form: tools like Zapier, Make, or the automation workflows built into operating systems have occupied this space for years. The difference IrisGo attempts to exploit is that it doesn't require users to articulate what they want to automate. The system infers patterns on its own, which significantly lowers the barrier to entry for non-technical users.

It remains unclear what model underpins Iris's core, what data is stored locally versus sent to the cloud, or what privacy controls the product offers. These are questions the team will need to answer precisely before any enterprise adoption becomes realistic.

Why Andrew Ng's Backing Matters

Ng is a figure with genuine technical credibility in applied AI, not just a brand-name investor. His involvement in projects like DeepLearning.AI and AI Fund has given him visibility into which types of agents actually work in practice and which remain demonstrations. That he's decided to back IrisGo suggests that screen-observation technology has reached a level of reliability that makes the product plausible, or at least that the founding team has something concrete under the hood.

That doesn't mean the product will work as described at launch. Agents that rely on desktop vision have a long history of promises and often fragile execution: small changes in an application's interface can break entire workflows. Robustness in the face of real-world environment variability is the unsolved problem any system of this kind has to demonstrate at scale.

Who This Makes Sense for Right Now

In its current state—assuming the product works reasonably as described—IrisGo makes sense for three profiles:

• Knowledge workers with highly repetitive tasks in standard applications (spreadsheets, email, CRMs) who lack the resources or time to set up conventional automations.
• Small teams without internal technical support who need to gain efficiency without hiring someone to maintain Zapier workflows or automation scripts.
• Companies experimenting with desktop agents who want to evaluate whether passive observation is a viable mechanism before betting on more complex API-based or MCP server solutions.

For environments with strict security or compliance requirements, an agent that continuously observes the screen will generate legitimate questions that probably block adoption in the short term.

Our Take

IrisGo is an interesting experiment in the right direction: reducing configuration friction so automation reaches more people. But "learning by observation" is a technical challenge that gets complicated quickly in real-world environments, and Ng's backing doesn't by itself resolve the privacy and robustness questions that this type of agent inevitably raises.