There’s no shortage of startups trying to create enterprise AI assistants. What’s less common is an AI assistant that can actually execute tasks across several of your work apps at once. That’s the promise of Narada AI, a startup building an AI assistant based on new research out of UC Berkeley.
Narada has been operating in stealth mode for two years, and made its public debut onstage today as part of the Startup Battlefield 20 at TechCrunch Disrupt 2024.
Two of its co-founders, Kurt Keutzer and Amir Gholami, co-authored a paper earlier this year on “LLM Compilers,” which are AI systems that perform multiple functions simultaneously. Their startup is largely based on this open source technique, and believes it’s a key differentiator from the general-purpose AI chatbots out there.
The startup’s co-founder and CEO, Dave Park, says his team used this as a basis to build a custom AI model that can use productivity tools. Park, a Stanford computer science PhD who spent 24 years working in enterprise sales, believes the LLM Compiler and Narada’s ability to use websites without APIs is the company’s “secret sauce” to winning the enterprise agent race.
The idea sounds promising, but how does the agent actually work? In practice, I found the assistant was successfully able to execute a few different tasks using generative AI through various work apps, ultimately saving me a few seconds or minutes at various parts of my day.
The assistant sits in a separate chat window in your browser, and can draft emails, make calendar invites, take meeting notes, and search the web on your behalf. The company says 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 another point, I prompted the AI assistant to find a highly rated Japanese restaurant in my neighborhood in San Francisco, and book a calendar invite to dinner with a friend at a time that worked with my schedule. It found a restaurant, created the calendar invite, and drafted an email to my friend with the information.
So how is the agent doing all this?
To use your email and calendar, the agent is partially using APIs to access these programs through a developer-facing back end. However, Park says their AI agent is also clicking, scrolling, and typing through the front end of websites (that’s how it’s popping open email drafts in Gmail, for example). This front end agent, which they call Web Redemption, should allow Narada to use enterprise applications without APIs, such as HubSpot.
Gholami, the startup’s CTO, says the agent is working like a Roomba, creating an internal map to understand new websites or applications. Once a user tells Narada it would like to use a new application, the agent supposedly maps it out so it can understand how to use it. That’s the idea the founders laid out to me.
But Narada is far from the only startup trying to create an AI agent that can use websites through a front end. It’s similar to the idea behind Anthropic’s computer use or Rabbit’s LAM. However, these agents are hard to implement in practice, and take a lot of maintenance to keep them running. If webpages update their layout, it can break the agent.
The main difference for Narada’s agent is that it’s solely focused on enterprise applications, instead of a general purpose agent for any website. (When I tried to use Narada for LinkedIn or Facebook, I was met with an error message, although there is a demo on the company’s website where an engineer is able to use the tool with LinkedIn.)
As for the LLM Compiler, folks around the industry already seem to be implementing the open source method. Gholami tells TechCrunch that LangChain and LlamaIndex already have integrations with the LLM Compiler. But Narada’s tool is unique from these tools because it’s 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 really. 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 made me slightly uncomfortable was how much access I had to give 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 have to trust not just the tech, but the company itself — that Narada won’t abuse your data, or your company’s data. That said, the company promises to not train its AI models on any customer data.
So far, Narada says it’s raised a few million dollars from a few advisers it has brought on, but the CEO says they’re now looking to raise more from traditional VCs.