Bardeen's automating all of your repetitive tasks away 🎛

CV Deep Dive

Today, we’re talking with Pascal Weinberger and Artem Harutyunyan, co-founders of Bardeen AI.

Bardeen is an AI-powered workflow automation platform, purpose-built for enterprise teams to automate repetitive tasks including sales, marketing, and data analysis. Founded in 2020 by Pascal and Artem, the startup’s mission is to simplify automations and help teams increase impact and save time on their everyday workflows - and today, the startup counts itself as one of a select group of ‘AI agent’ startups looking to leverage LLMs and generative AI to help businesses become more efficient. 

Today, Bardeen has thousands of companies using its platform for automating workflows, including companies such as Deel and others. Currently available via a Chrome extension, Bardeen also touts its differentiated approach to the AI agents space, with a heavy focus on the combination of deterministic and AI-driven automation. In this conversation, Pascal and Artem walk us through the founding premise of Bardeen, their unique approach to AI automation, and Bardeen’s position within the AI agent landscape. 

Let’s dive in ⚡️

Read time: 8 mins

Our Chat with Pascal and Artem 💬

Pascal and Artem - welcome to Cerebral Valley. First off, give us a little bit about your backgrounds and what led you to co-found Bardeen? 

Hey there! My name is Pascal, and I’m the CEO and co-founder of Bardeen. My background is in computer vision - I've worked with multiple teams, ran my own company in the computer vision space, and also led a large AI team for Telefonica. 

My name is Artem, my background is in systems and infrastructure engineering and my last key role was with Mesosphere (later known as D2iQ) where I led engineering and product teams. 

The interesting part of Artem and I’s story is that we were both essentially trying to automate ourselves out of our jobs. We used many different automation software solutions, but none of them quite met our needs. So, we both independently came to the idea that we needed something akin to a "mini-me"—that was actually the initial project title we each independently thought of. We pitched it to a common friend of ours, Flo Liebert, who was then the outgoing CEO of Mesosphere and later became our first seed investor. 

It was quite surprising—almost scary—how we were pitching each other’s ideas, finishing each other's sentences during our very first discussions. After a few months of exploration and getting to know each other better, we decided to build this company together and ultimately create an automation platform that could be used by everyone. If you look at existing automation tools, they’re typically designed for big companies with large consulting teams, or they are fairly technical and you have to know how to program to actually use them. They are all designed for “operations people”, but there’s a fairly limited choice for tools for those who do the actual job of selling, marketing, recruiting, managing and so on

What we’re aiming to do with Bardeen is build an automation tool that anyone can use, especially those who are in the trenches doing the work day-to-day. Our goal is to bring automation to the 90% of the market that should be automating today but isn't.

How would you describe Bardeen to the uninitiated developer or business owner? 

Bardeen is like using generative AI to discover, build and run automations for you directly in your browser. To unpack that a little bit, we use generative AI to make it super easy for the end user to build automations. You can create automations using natural language by describing what you need automated, and we build the automation for you. We chose to build it into the browser because we see the browser as the runtime of the future, especially considering that most people spend 90% of their time in the browser with 50+ tabs open.

You basically spend a lot of time moving and transforming data between different platforms as part of your workflows, whether it’s in sales, recruiting, project management, or even researching potential companies to interview next. Many of these workflows are highly repetitive - you’re transferring data between multiple browser tabs. That’s ultimately the kind of workflows we’re aiming to simplify with Bardeen.

Who are your users today? Who’s finding the most value in what you’re building with Bardeen?

Our target audience is frontline knowledge workers, or more colloquially, people who "actually do the work"—with an emphasis on "actually." The reason for this is because many existing automation tools are designed for operations roles, like sales or marketing. However, the people directly involved in the work, like SDRs, BDRs, recruiters, and marketers, aren’t being served properly. We're seeing a lot of traction in areas like lead generation, outreach, tracking, and project management, where these workflows —sales, marketing, recruiting—tend to have a very similar shape. 

Let’s say you have a source of data, which could be platforms like LinkedIn or Apollo, then you move this data into some sort of tracking system—this could be a spreadsheet, Notion, or a CRM, and from there, you do filtering and outreach. This pattern applies across many verticals, and we’re seeing a lot of traction with it, especially among salespeople, real estate professionals, recruiters, and marketers.

Our challenge is building the product and also changing behavior. It’s as much an education problem as helping people understand that they can and should automate. We invest heavily in content—written guides and videos—to help people learn this skill.

Could you give us an idea of the positive impact that using Bardeen has on performance and latency of individual workflows, relative to other AI-powered automation platforms? 

The beauty of having a deterministic runtime is that you don’t have a language model involved most of the time at runtime. We handle the inference and interpret the user’s intent—which are high latency, high cost problems to solve—and we do that at what we think of as build time, which is when the automation is being created. Now, once you’ve done that, every time you run the automation it may not even need an agent model. Instead, it just executes as a series of API calls and DOM interactions that have been hard-coded. 

This means that when you execute these automations, there's virtually no latency other than the API latencies that you would also get if you were to manually interact with it as a user - you don't even need to wait for the UI to get rendered. As you know, APIs are very much optimized for load latency - they're very optimized systems, so the latency will always be shorter. Here, you have no load times - you don't need to wait for a language model to reason or structure with it. You also safeguard yourself from non-determinism, hallucinations and other issues inherent to LLMs. 

So, in terms of both latency, reliability and cost, Bardeen is orders of magnitude faster than other approaches at runtime. 

This is different at build time. In build time we utilize the full power of LLMs to make it easy for users to build and test automations around their workflows. This is where AI is extremely useful because it unlocks a lot of use cases that weren’t possible before. It not only helps to discover what can and should be automation, but also to go from an intent formulated in English to a working automation agent playbook.  

Walk us through the evolution of Bardeen from the pre-LLM days, all the way to now. How has the advent of generative AI and agents changed the trajectory of Bardeen? 

It’s pretty interesting because when we first started, we were already talking about how people would be able to use automation almost like a search engine. You’d just describe in natural language what you want to automate, and we’d figure it out from there. Back then, there was a lot of hesitation and questions around whether it’d be possible to build something like this. In our minds, we were already thinking about how we would launch a product, gather a ton of data, train our own models, and build this all from scratch. We laid the foundation by building a DSL for automatons specifically designed with a purpose of the code being easily generatable with language models.  Then, GPT-3.5 came out and (since it was trained on a lot of code) right from the get go, we started seeing some amazing results for the semantic parsing problem that was essential for solving our product puzzle. 

Today, you can do a lot with deterministic actions, via API calls and DOM manipulation that we’ve built with our integrations, but there's always a long tail of actions that people will want to take, that have no API. That's ultimately where this next generation of AI agents comes in and fills that gap for the long tail. Everything you can do via an API, you should still do because that's the most reliable modality - but for the things you can’t do via an API, you now can use a Web UI to accomplish your goals. This is the approach that we use to build our AI agents that are very versatile and reliable.  

Our Agent approach enables a long tail of much more valuable, multi-step workflows. One thing that hasn’t changed is our view that there needs to be some kind of a deterministic substrate powering the  agent, copilot or automation. We still believe that just shoving DOM down the LLM throat is not going to cut it.

Artem - you’ve been public with your thoughts about where today’s AI agent frameworks are falling short, and how Bardeen is thinking about this differently. Could you walk us through some of the main takeaways? 

Absolutely. Firstly, I’d say the web UIs that we have today are designed for humans, and humans get very easily overwhelmed. So, the best UIs, the Apples, Linears and Notions of this world, start simple and clean - however, ‘simple’ means that the actual functionality is hidden behind menus and layers. On top of that, UIs are also very fluid - they change all the time. RIght off the bat, an AI agent is at a huge disadvantage - the person on the other end is building the application for something else, and doing something that kind of hurts what you're trying to achieve. 

Secondly, the web is horrible for agents as a medium because agents are not distinguishable from various bots, and normally those are associated with malicious activities. On the web, you have bot protections, and web apps - unbeknownst to themselves - are often sabotaging what you're trying to do. The other thing is that the underlying structure of the web application, the DOM, is bloated - it’s a huge tree of information designed for the browser to render the page in the most efficient way, but it's not optimized for an agent interacting with it. Lastly, for a task to even be worth automating, it has to be something more than what you spend 5 minutes every other month doing. It has to be valuable and frequent enough.

Our approach is that we believe there has to be determinism wherever possible - meaning we have to take this AI uncertainty out of the picture whenever we can. If I want to fill 100 rows in a spreadsheet, I shouldn't be trying to do this with AI. There are API calls for that, but there is a long tail for which there are no APIs. This part of building automations is very useful because a human can describe it, and those are the right places to have the AI. 

So, our approach is radically different from anything that we've seen so far in that we have a very clear separation in our mind between building an automation and running an automation. We believe that when running an automation, you have to make it as deterministic and as less likely to fail as possible. That's why we rely heavily on proven APIs out there, and use AI for everything else.

Multimodal is a huge area of huge excitement for those building in AI. How does the incorporation of multimodal AI feature into Bardeen’s roadmap? 

Multimodal is very interesting to us. We recently published a paper called "WILBUR" about that, and we outlined our approach and benchmarks. It turns out that if you do the agents correctly, then your text output can be on par with multimodal performance, which is an order of magnitude cheaper and faster. Obviously, models evolve - it's a very dynamic field, and so it remains to be seen how it progresses. We're also doing research to try to push the field forward.

In terms of productization, I’d say the people who are actually doing the job are waiting on the productivity gains from multimodal and AI to become fully realized - and honestly, it feels like some of them are running out of patience. For us, this is a year of execution, not flashy Twitter demos and PR. We're heads down building stuff and talking to our customers every day, making sure we actually help them drive value out of this. People are now expecting real results - for example, a car dealer from South Africa who has 20 dealerships to run, and serious business problems that we're solving for them. 

So for us, it's about making AI more accessible and robust, and basically unlocking the tech for more people. That's what we're focused on - making the product simpler, better, faster, cheaper, and getting more users. We have tens of thousands of people who use Bardeen weekly, and a lot of them are paying us. Ultimately, we want to work on automation end to end, where you don't even need to think about what you want to automate, and you just have a genie sitting on your shoulder watching your every move and proactively suggesting information to you. That's something that we're super excited about, and we're very well-positioned to execute on this.

Lastly, share a little bit about your team. What is the culture you’re trying to build, and what do you look for in prospective new team members? 

We're about 30 people total. We're very much engineering-heavy at the moment - as we've already talked about, we're tackling a super hard technical problem. This is reflected in our culture; we're a team of problem solvers and people who love to challenge themselves until they break through tough barriers. We’re aiming to build a culture where everyone is extremely intellectually curious and has high persistence. 

The problems we’re solving are very hard, both on the go-to-market side but also on the technical and research side. We're looking for people who are super intellectually curious and eager to push the market, push the boundaries in research, and elevate everything to the next level. We operate as a fully remote team with a heavy bias towards San Francisco. We're hiring for both go-to-market and research engineering roles. If anyone reading this is interested, feel free to reach out.

Additional Resources:

WILBUR (AI Agent):

• WILBUR paper - https://arxiv.org/html/2404.05902v1

• WILBUR Summary, overview and explanation - https://www.bardeen.ai/posts/wilbur-adaptive-in-context-learning-for-robust-and-accurate-web-agents

• https://x.com/_akhaliq/status/1778636890616307876

Blog Links

• Agent Blog post https://www.bardeen.ai/posts/the-right-way-for-organizations-to-leverage-ai-agents

• AI Agents for Sales - https://www.bardeen.ai/posts/ai-web-agents-for-sales

• Another Agent blog https://www.bardeen.ai/posts/are-ai-web-agents-a-gimmick

Resources

• Guide to help salespeople learn how to automate work - https://www.bardeen.ai/resources/field-guide-for-sdrs 

Tweets:

• https://x.com/hartem/status/1783563839612785086

• https://x.com/hartem/status/1783563839612785086

• https://x.com/bardeenai/status/1781019971587317765

• https://x.com/_akhaliq/status/1778636890616307876

• https://x.com/hartem/status/1749459419954602056

Conclusion

To stay up to date on the latest with Bardeen, follow them on X and learn more about them at Bardeen.

Read our past few Deep Dives below:

• Baseten is pushing the boundaries of AI Inference ⚡️

• Martian's interpretable alternative to the Transformer 🔌

• Our chat with Groq's Chief Evangelist, Mark Heaps

• MultiOn is building software with a brain 🧠

• Galileo AI's groundbreaking prompt-to-UI tool ✨

If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email ([email protected]) or DM us on Twitter or LinkedIn.

Join Slack | All Events | Jobs