Of course, there are numerous startups trying to create an enterprise AI assistant. The rarer is an AI assistant that really can execute tasks across multiple work apps at the same time. Here is the promise of Narada AI: a new startup building an AI assistant based on fresh research at UC Berkeley.

Narada has been stealthy for two years now, and today it made its public debut on stage, as part of the Startup Battlefield 20 at TechCrunch Disrupt 2024.

Its co-founders, Kurt Keutzer and Amir Gholami, earlier this year wrote a paper on "LLM Compilers," an AI system that performs many different functions at the same time. Their startup is very much based on that open source technique, and thinks it's a major point of differentiation from general-purpose AI chatbots.

According to the co-founder and CEO Dave Park of the startup, his team used this as a basis to build a custom AI model that can use productivity tools. A Stanford computer science PhD with 24 years in enterprise sales, Park believes LLM Compiler and Narada's ability to use websites without APIs is the company's "secret sauce" to winning the enterprise agent race.

Sounds promising, but how does the agent actually work? In practice, I found that the assistant was able to successfully execute a few different tasks using generative AI through various work apps and, ultimately, saved me a few seconds or minutes at various parts of my day.

The assistant sits in a separate chat window in your browser, can draft emails, make calendar invites, take meeting notes, and search the web on your behalf. According to the company, its assistant can also navigate enterprise applications, such as finding an invoice in SAP, taking notes on a video call, or analyzing information from Salesforce's many apps.

I asked the AI assistant to draft a friendly-sounding email declining an invitation I had received. In seconds, a drafted email appeared in my Gmail with the correct recipient (even though I didn't tell it the person's email, it found the right one), subject, and body all filled out with my signature at the bottom. All I had to do was review it and click send.

At one juncture, I prompted the AI assistant to discover a well-reviewed Japanese restaurant in my neighborhood in San Francisco, and schedule a calendar invite to dinner with a friend at a time that would fit with my schedule. It found a restaurant, constructed the calendar invite, and composed an email to my friend with the details.

So what's the agent doing it all?

For its part, an agent that uses your email and calendar is partially clicking through a developer-facing back end using APIs to reach these programs. But Park's AI agent is also, in many ways, just clicking, scrolling, and typing through the front-end of websites: that's why it is popping open the draft email in Gmail. This front-end agent which they have called Web Redemption should, in theory enable Narada to use any enterprise applications without APIs for example HubSpot.

Gholami, the CTO of the company, claims that the agent is working like a Roomba because it creates an internal map when it comes to understanding new websites or applications. In case a user wants to use a new application, the agent is supposed to create a map for it and will understand how to go about using it. Those are the ideas the founders told me about.

But Narada is far from the only startup that is trying to build an AI agent to use websites through a frontend. Much more in the vein of Anthropic's computer use or Rabbit's LAM. In practice, though, these agents are a pain to implement, and it hurts to maintain them if they do not work seamlessly in case the structure of the web pages being updated breaks the agent.

The key benefit for Narada's agent is that it is really focused on enterprise applications alone, rather than a kind of general-purpose agent useful for any website. It is worth noting that whereas I was unable to log in to LinkedIn or even Facebook using Narada itself, there is a demonstration on the company's home page where an engineer actually uses the tool with his LinkedIn account.

As for the LLM Compiler, people seem to be already implementing that open source method. LangChain and LlamaIndex already have integrations, according to Gholami, with the LLM Compiler. But Narada's tool is unique from these tools in that it is focused on the enterprise – the startup already has a Fortune 500 company using its agent but wouldn't disclose which one.

So is this a replacement for a real-life assistant? Not quite. However, the tool at times felt like using a shortcut for mundane tasks, which is more than I can say for a lot of AI tools today.

One thing that really made me uncomfortable was how much access I was granting to this AI assistant. Narada can read all of my emails, it can see my entire calendar, and it knows my full contact lists.

Like any "smart assistant" or helper app like this, you need to trust not only the tech but also the company itself, that Narada won't abuse your data, or your company's data. The company promised not to train its AI models on any customer data.

So far, Narada has raised a few million dollars from a few advisers it has brought on, according to the company's CEO, but it is now looking to raise more from traditional VCs.