Key takeaways from startup founder Div Garg on agentic AI include the following:
Trust is the biggest barrier.
Security, permissions and control remain critical.
Cost limits deployment at scale.
Automation is often partial today.
Biggest near-term value is in mundane task automation.
Strong governance and access controls are required.
AI is increasingly influencing decision-making.
Agentic AI is evolving rapidly, but confidence in how these systems will behave in real-world settings remains limited.
Div Garg, CEO and founder of AI research lab AGI Inc., argues that while AI agents are already automating parts of enterprise workflows, most people are still not ready to fully trust them with sensitive or high-impact tasks. Issues like security, permissions and unpredictable behavior remain major blockers, especially when agents have access to tools such as email systems, CRM systems or financial accounts.
Garg also points to cost and deployment constraints as additional hurdles, noting that today's systems can still be expensive to run at scale and often only partially automate complex workflows. Even so, he expects agentic AI to offer meaningful value in enterprise automation, particularly across sales, recruiting and administrative processes where repetitive work is common.
In the following interview, Garg shares why he left opportunities at major tech firms to build a startup, what he learned from working at Apple, Google and Nvidia, and where he believes AI agents will have their biggest effect in the enterprise.
Editor's note: The following transcript was edited for length and clarity.
Why did you turn down a near seven-figure offer from OpenAI to build your own company?
Div Garg: I wanted to innovate myself. If you join a big company, you're a part of the machine, but you're not directing or changing things yourself. I wanted to build a product and get it out in the market -- a breakthrough in how people do things. That has always excited me.
There are so many applications out there, such as OpenClaw, that I felt I would be wasting my talent if I were to just spend my time training models at a frontier lab. There's a lot of learning in entrepreneurship. You're not just responsible for the tech. You're also figuring out the nuances of building an organization, building a great product and selling to users. It's a once-in-a-lifetime opportunity, and I didn't want to miss it.
You've worked with Apple, Google and Nvidia. What did those environments teach you about building AI?
Garg: I worked in many AI teams at these places, and it was great to see the scale and how people think internally. In a big company, there are many resources, such as GPUs, and it's very helpful to see examples of what excellence looks like. I'm inspired by those efforts.
When I was working on the Apple Cloud project, I was working with world-class experts. The world-renowned AI scientist Ian Goodfellow was my boss. It was incredible to learn from the best of the best in the field. These experts showed me their best practices -- how they train models, how they evaluate, benchmark and deploy AI in the real world.
There's a lot of like stuff we must think about. For example, how do you build trust? Why would a human trust the system, and why would someone not freak out if this car starts driving itself? I also learned a lot about how to make AI part of a user's everyday life. That's important when building something novel.
Is there anything you felt like big tech got wrong?
Garg: Things move slowly in big tech because they must manage a lot of internal politics and brand things before launching something. Companies like Google must spend three months working with lawyers before a new product launch.
Startups can move much faster and have a larger risk appetite. If you're a small company, you're not on the map, and people don't know you. Even if you launch a product that's in beta -- maybe it's 90% working but 10% of things you still need to fix -- that's totally fine, and you can then move quickly compared to a big company.
Many vendors are calling their products agentic. What distinguishes a true autonomous agent from simple workflow automation or a chatbot?
Garg: An agent is something that can intelligently make decisions in its environment, and that environment can vary. It could be a coding space, or maybe something that's taking action on my PC. However, it needs to have intelligent decision-making, where it's not purely scripted but can choose the best option based on the user's preferences.
One of the biggest things missing right now is trust.
What technical hurdles must be solved before agents become reliably useful in real-world settings?
Garg: One of the biggest things missing right now is trust. Trust is about whether the agent will do what it's told to do. This is missing in many places. For example, if you're using OpenClaw, there are so many security and privacy issues -- it might do something wrong, and you have almost zero control. How you fix that is a big thing we want to solve. Can I trust this AI agent to do exactly the right thing, especially if it has access to bank accounts and passwords? And how do we do that well?
The second aspect is cost. If you're using any kind of agent, even Claude Cowork or Claude Code, it's really expensive. You might end up burning $1,000 a day, and that doesn't make sense for a normal user. If we want to scale agents to every single person on Earth, at some point the cost of AI must be as close to zero as possible, and we are excited about this.
A lot of what we're doing is on device. For instance, how can we make the AI item on your devices fully local, running on your phones and PCs? This is a pathway to let us build a future of super intelligence running on the edge where there's no cost for inference, and anyone can go and use this.
What do you think will make people willing to let software act on their behalf?
Garg: The convenience. We all have so many boring, mundane tasks to do every day. If an AI copilot or sidekick can do all those things for you, it changes your life. It lets you be in the moment, focused on things you're passionate about.
What governance controls do autonomous AI systems need if enterprises are going to adopt them at scale?
Garg: You need a lot of access permissions. What apps can the AI access and use? How much sensitive data does it have access to? Does it have access to emails and any sort of private walls or passwords? Can it use your credit card for transactions?
We've been trying to build fine-grained permission layers. For instance, can a user choose which apps their AI can use? Maybe it can use the Uber app and DoorDash app, but not my banking app. And can I control this over time?
Where do you think agentic AI will first create meaningful business value?
Garg: Enterprise automation. Companies already spend heavily on workflow automation and robotic process automation tools, and many of those processes could eventually be handled by AI agents instead. There's also a lot of repetitive administrative work across sales, recruiting, hiring pipelines and CRM systems like Salesforce. A lot of those tasks can be automated by agents.
A lot of engineering is automated at this point -- no one is writing code themselves.
Is agentic AI is going to be replacing knowledge workers on a large scale?
Garg: We are already starting to see some of that happen. Coding agents have really taken off, and everyone is using them. A lot of engineering is automated at this point -- no one is writing code themselves. A lot of systems are being automated. There are so many AI native platforms for everything, and you don't have to manually fill in data and find information. All those tedious steps, which used to be painful and required an army of humans, are now being automated.
If a CIO is evaluating agentic AI today for their organization, what questions should they ask before selecting a vendor or tool?
Garg: CIOs should first make sure the AI tool actually solves a real business problem. Every organization has different workflows, systems and operational needs, so they need to evaluate whether the product can adapt to their environment effectively.
They also need to be realistic about the current limitations of agentic AI. In some cases, only 30% of a workflow can be automated, while the other 70% still requires significant human labor. Understanding what technology can and cannot reliably do is critical before investing in it.
How much decision-making do you expect organizations to hand over to AI systems?
Garg: Over 95%. It's already starting to happen. The best AI chatbots are very intelligent and capable of powerful reasoning. Most smart leaders are using AI for decision-making because it can uncover blind spots your team might not have even considered. This gives your organization so much power to solve problems. A lot of decision-making in complex technical environments will be done by AI.
Tim Murphy is an award-winning reporter covering IT strategy for Informa TechTarget.
