Jeffrey Lai, Founder and CEO of IrisGo – Interview Series

Jeffrey Lai, Founder and CEO of IrisGo, is an AI entrepreneur and product leader with deep roots in intelligent assistants and language technologies. Before launching IrisGo in 2025 and serving as a Founder in Residence at AI Fund, Lai spent more than 12 years at Apple, where he helped build and scale Siri, ultimately serving as Siri Language Engineering Manager. His work focused on advancing multilingual conversational AI, including the development of the Chinese-language version of Siri. Prior to Apple, he held engineering roles at Cisco Systems and completed an internship with the Intelligent Robotics Group at NASA Ames Research Center. With a career spanning robotics, enterprise software, and consumer AI, Lai has consistently focused on making complex technologies more accessible and useful in everyday workflows.

IrisGo is developing an AI operating system designed for the new generation of AI-powered personal computers. Rather than operating as a standalone chatbot, the platform is built to observe how users perform tasks across applications, learn those workflows, and automate them through an intelligent on-device assistant. Its core approach centers on allowing users to demonstrate a task once and then convert that process into a reusable workflow, helping reduce repetitive digital work. By leveraging the dedicated AI processing capabilities found in modern AI PCs, IrisGo aims to enable more private, responsive, and context-aware automation, positioning itself within the growing market for agentic AI systems that can actively execute tasks on behalf of users rather than simply providing recommendations.

After helping build Siri at Apple, what specific experiences convinced you that it was time to start IrisGo, and what was the key problem in human-computer interaction that you felt remained unsolved?

When we were building Siri, we were incredibly focused on natural language understanding. But Siri struggled with context, which is challenging because the way people actually work and communicate is full of implicit context. When you ask a colleague, “Can you send me that report?” they know which report. They know your shared history, your current project, what you’ve been discussing. Siri didn’t have that.

For many years, AI has been designed to respond to you. Type something in, get something back. That’s the interaction model we’ve inherited from search engines and chatbots. But most of what people actually do on a computer isn’t a question, it’s a workflow. Compile this report. Cross-reference these spreadsheets. Pull data from three different apps and format it into a document. That kind of work doesn’t fit in a prompt.

What convinced me to start IrisGo was realizing the industry wasn’t building for everyday use cases. Everyone was building smarter models, better language understanding, faster responses. But the fundamental gap, where AI can actually do the work, not just respond to requests, was still there. That’s the problem we set out to solve.

Looking back at your years working on Siri, what were the biggest lessons you learned about the limitations of voice assistants, particularly when it comes to understanding context, intent, and user behavior?

The biggest lesson I learned is that context is everything. People rarely communicate in isolated commands; their requests are shaped by what they’re working on and what they’re trying to accomplish. Early voice assistants were designed around answering individual requests, which works well for many tasks but is less effective for ongoing work that spans multiple applications and workflows. That experience taught me that understanding intent requires understanding context, and that insight became one of the foundational ideas behind IrisGo.

What we learned is that intent is often layered. Someone says one thing and means another, and getting that right requires understanding the person, not just the words. That understanding is baked into IrisGo from the beginning.

IrisGo is often described as an AI operating system rather than an AI assistant. What is the difference, and why do you believe the industry needs a new software layer for the AI PC era?

An AI assistant waits for you. You open it, you type something, it responds, and then it’s done. The interaction model is essentially the same as a search engine, where the user makes a request and the AI generates a response.

IrisGo, which is an AI operating system, works alongside everything you’re doing, across your entire system, all the time. It has access to your files, your apps, your browser, your email. It understands how you work because it’s been watching, with your permission. And crucially, it can act, not just advise.

Think about what Windows did for personal computing. Before it, you had to be technical to use a computer. Windows created a universal, intuitive layer between people and software. The AI PC era needs its own version of that layer, taking the raw power of AI models and making it truly useful to people doing their day-to-day work.

That’s what IrisGo is: the intelligence layer of the AI PC. When you ship AI PCs without that layer, you’ve just got a powerful chip sitting idle. What you need is software that knows how to put it to work on behalf of the person using it.

One of IrisGo’s core ideas is that users can teach the system a task through a single demonstration. Why do you believe “watch and learn” is a more natural model than prompting or manually building workflows?

What makes the ‘Watch and Learn’ feature so powerful is that it’s based on how we actually teach each other. When you onboard a new employee, you don’t write them a 40-step prompt. You show them how you do something once, they watch, they learn, and they take it from there. That’s intuitive. That’s natural.

The problem with prompting is that it puts the burden on the user to articulate exactly what they want, in a format the AI can understand. For simple queries, that works fine. But how about for complex, multi-step workflows involving your specific files and your specific way of doing things? It can break down fast. Most people can’t describe their workflows with enough precision to prompt an AI to do them correctly.

Showing is easier than telling. That’s not a new insight; it’s common sense. What Watch and Learn does is apply that logic to automation. You do your task once, the way you normally would. IrisGo learns the pattern. And from that point on, it handles that workflow for you.

Many AI agents today still require significant setup and supervision. What separates a truly autonomous agent from a chatbot that simply follows instructions?

The short answer is that a truly autonomous agent has its own context, makes its own judgment calls, and keeps working even when you’re not watching.

A chatbot is reactive by design. It lives and dies by the prompt. You give it an instruction, it executes, and then it stops. If something changes, if a step fails, or if it needs to adapt, it gets lost. It’ll ask you what to do next.

A genuinely autonomous agent has a goal, not just an instruction. It has context about the environment it’s operating in. It can handle exceptions. It can decide whether to proceed, retry, or flag something for human review. And importantly, it can run in the background while you work on other things.

Most of what’s being sold as ‘AI agents’ today are really sophisticated chatbots with extra steps. The gap between those products and real autonomy is enormous. What makes IrisGo different is that it starts from a model of how you work, learned directly from watching you, and uses that to make the right judgment calls, not just follow instructions.

You have spoken about the importance of persistent memory and local context. Why are these capabilities essential if AI is going to move beyond answering questions and begin acting on a user’s behalf?

Imagine hiring an assistant who, every single morning, had no memory of who you are, what you’re working on, or how you prefer things done. You’d spend half your day re-explaining context, and you’d become extremely frustrated. That’s essentially what most AI tools ask you to do today.

Persistent memory means the system knows you. It knows your projects, your workflows, your preferences, and how things have evolved over time. You don’t have to reestablish context every session. The AI carries that forward.

Local context is equally critical, because your work doesn’t live in the cloud. It lives on your computer, in files, applications, emails, browser tabs. IrisGo runs on your device, which means it has access to your entire working environment and can operate on your behalf across all of it.

These two things together — memory and local access — are what allow AI to stop being a tool you use and start being a system that works for you. We’re seeing more of the tech industry start to recognize the importance of this, including the Big Tech companies, so it’s a clear validation of our approach.

Privacy remains one of the biggest concerns surrounding AI. How do you balance the need for personalization and memory with user expectations around security and data ownership?

The assumption that privacy and personalization are in tension is the first thing to challenge. We built IrisGo so that personalization can happen on-device, not just in the cloud.

The vast majority of IrisGo’s processing happens locally, on your computer. Your workflows, your files, your context all stay with you. Cloud processing only happens when you explicitly authorize it, and when it does, it’s encrypted end-to-end. You’re in control of what goes where.

This model makes IrisGo more powerful, because it has access to the full context of your working environment. And it means users don’t have to choose between a useful AI and a secure one.

People trust their personal computers with sensitive work. Our goal is to make IrisGo worthy of that same trust. This is fundamental to how the product is designed.

IrisGo is already being shipped on AI PCs from Acer, with plans to expand to additional manufacturers. How important do you think distribution partnerships will be in determining which AI platforms ultimately reach mainstream adoption?

Distribution is absolutely critical for mainstream adoption. Just look at the history of platform competition in tech. Windows won because it was on virtually every PC. Android won because it was the most widely used operating system for mobile devices that weren’t iPhones. The companies that tried to win through pure product quality, without the distribution to back it up, mostly didn’t make it.

The AI PC market is following a similar pattern. When IrisGo is pre-installed on a device, it’s there when you open the box. You don’t have to search for it, download it, or decide to give it a chance. That’s an enormous advantage.

We’re already in 3 million Acer laptop devices that will be shipped this year, and we’re in conversations with additional tier-one PC manufacturers. The goal is to be the AI Operating System  for PCs by default. That distribution strategy is central to how we become the standard in this category.

As AI agents become capable of handling invoices, research, reporting, and other business workflows, what new governance, oversight, and trust challenges should organizations be preparing for today?

The first challenge is auditability. When an AI agent completes a workflow, such as processing an invoice or generating a report, can you see what it did, why it did it, and where the data came from? That kind of transparent audit trail is essential for compliance and for catching errors before they compound.

The second is defining boundaries. Which workflows are appropriate for autonomous execution, and which ones require human review? Companies need clear policies around this. The tendency is to let AI do more and more, incrementally, and then realize too late that high-stakes decisions are being made without sufficient oversight.

Third is trust calibration. Employees need to understand what the AI can and can’t do, and when to verify its output rather than assume it’s correct. That’s a culture and training issue as much as a technology one.

I believe that the organizations preparing for this today will have a meaningful advantage, because as AI capabilities grow, the more governance frameworks will compound.

Five years from now, do you think people will still spend most of their time interacting with AI through chat interfaces, or will AI increasingly fade into the background and proactively complete work without being asked?

The trajectory is already showing that AI will increasingly move into the background. The model we’re building toward is one where your computer understands your work and takes care of the routine pieces of it, so your attention is freed up for the things that actually require you.

Think about how autopilot works in aviation. Pilots aren’t hands-on every second of every flight. The system handles the baseline, and the human engages when judgment and expertise are needed. That’s the future of AI-assisted knowledge work. Your computer handles the repetitive, rule-based workflows. You focus on strategy, creativity, and the decisions that require real human judgment.

Five years from now, the most productive people will be the ones alongside systems that understand how the user operates. In the past, people did all the work on their computers. In the future, your computer will also do work for you.

Thank you for the great interview, readers who wish to learn more should visit IrisGo.