AI Leadership Insights: Integrating Web Applications

Mike Vizard: Hello, and welcome to Techstrong.ai. We’re here with Pascal Weinberg who’s the CEO of Bardeen.ai. And we’re talking about how to better integrate all our web applications because, well, right now it’s a lot of work and our workflows are something of a mess. Pascal, welcome to show.

Pascal Weinberg: Thank you so much for having me. I’m super excited to be here today.

Mike Vizard: So every organization I know these days has some number of web apps. And they’re usually built to automate some process somewhere, but then they’re all siloed. And nothing seems to work together. And I’ve yet to meet a process that didn’t mind expanding at least three or four applications. And the integration is always somewhat, shall we say, tenuous at best. So how can this all get better? And what role might AI play in it?

Pascal Weinberg: Yeah, it’s a great point. So I think like this problem of having like 50 tabs open in the browser, and like copy pasting data between different apps and services, and as you mentioned, they’re all kind of like, doing a great job at like a certain part of what your daily workflows are like, but nothing really solves the whole thing end to end. So we, as the users, end up being kind of the, you know, monkey’s almost copy pasting data between them. And what we’re trying to do with Bardeen is essentially allow you in a very easy and intuitive way to bridge those gaps and build automations that work in context across your different workflows. So examples can be, you know, when you’re doing your research for the podcast, you might want to like look at LinkedIn, and CrunchBase, and news from a certain company or person that you’re talking to, and then pull all that data into some project management system like Trello, or something like that. So traditionally, you would end up by copy pasting a lot of information between different tabs and a lot of copy pasting, context switching, and so on. And with Bardeen, you could build an automation that does that for you with a click of a button, and you’d just give the name or the context of the person that you’re looking for. And then we do the hard work for you all on the browser, on the edge with privacy first, by design, and we leverage, like a new generation of AI technology to help make it very easy to build those automations. But traditionally, you would need to build it in like no-code builders. With us, at this point, just describe what you want to do, and it interprets that for you. So that’s kind of like a high level; happy to jump into more details wherever it makes sense.

Mike Vizard: So it has a natural language interface that connects to some sort of generative AI platform. Did you guys build that yourself? And how does it know what my environment looks like?

Pascal Weinberg: Yeah, that’s a great question. So we built – the Bardeen comes in a web extension, so it plugs right into your browser, which is today where most of us spend, you know, 80 plus percent of our time and workflows. And like that web extension, and all the interface and connections and connectivity between those apps and the abstractions layers are all things that we built ourselves. The actual underlying language model, we utilize some of the advancements that have been made in the last, you know, six months of this nascent technologies by companies like open AI and others, where we basically plug into their language models to do some of the work and we didn’t have to kind of like pre-process and post-process, the model’s output. I think that’s something that with a lot of like AI companies these days, you’ll see that, you know, some of these two ways that goes like what some of them try to build their own language models, but get too prohibitively expensive, and it’s a very fast moving field. So it almost makes no sense to invest there unless you can invest like hundreds of millions of dollars, versus the approach that we’re taking where we say that like, okay, this technology is going to, quote commoditize, in some sense, where there’s an app, there’s like open AI out there that is building these language models, and we build on top of those, we have kind of a lift and shift approach, where as when new models come out like GPT-5, at some point might come out, we then basically just take our platform, move it from GPT to GPT-5, and then like benefit from that advancement. So that’s kind of the approach we took. So yeah, everything around it, we had to build ourselves.

Mike Vizard: And over time, it seems to me at least, that there’s gonna be multiple large language models that people build for different use cases, but I need something that stitches all that together, right? Because they all have API’s, but, you know, not every one of them is best suited for every use case, especially if it’s a general purpose, large language model.

Pascal Weinberg: Yeah, exactly. It’s a great point. I think that you know, you’ll see that with large language models, and you also see that with like database technologies, you know, today you have 10 different task management words, you know, from Notion to Trello and more that also have like a specific purpose. And like we in general run into this problem where we have many, many fit-for-purpose tools. But it’s hard to kind of like orchestrate them and like, pull them all together into like a coherent workflow that makes it easy for the end user to use. So that’s kind of like the abstraction layer that we’re building with Bardeen, where you can integrate with your SaaS tools like, you know, Notion, Google Sheets, you name it. But you can also integrate with those AI technologies like a large language model from open AI, various different models that we have available there. Or OCR technologies or text to speech technologies, all these things become building blocks in workflows that you as an end user want to build an orchestrator. So that’s kind of like the way we think about it is this like abstraction layer that really allows the end user to make it easy for them to build workflows for their various use cases. So yeah.

Mike Vizard: We have been trying to create this category of folks called citizen developers with mixed success; we give them low-code tools. And we hope for the best and, generally speaking, they build apps that don’t scale, they’re not particularly pretty and they’re generally insecure. Um, but other than that, it’s great. So if we use your approach, can I actually get to the promise of citizen developers because they actually have the knowledge of how the workflow works, but they don’t have to be as deep into the how to code model as they might otherwise be today?

Pascal Weinberg: Yeah, I mean, I think like, there’s been a lot of amazing progress in the no-code field. As you mentioned, in general, we are much more, we’re not as focused on like the building an app approach, you know. We wouldn’t want you to publish a website with Bardeen. We’re very much focused on helping you automate the workflows that you know, best – like your own workflows. But everyone understands what they do in their day to day and they know the ins and outs of what they need to have done. And we try to now make it like, as easy as possible. Now, when we first launched the platform a bit more than a year ago, the language technology wasn’t quite there yet, where we actually tried to have something like what we’re launching now, but the models were just not strong enough. So it didn’t work. Most of the time, we decided not to go with that. And instead launched with like, an easy to use no-code builder, but you still have to kind of like dissect your workflow, like, you end up with this. And you can see all the imagery on our website, right? You kind of like end up with this blank sheet of paper. And as a user, I now have to understand that like, oh, if I want to, for example, for my sales workflow, get LinkedIn data into my CRM system, I have to first extract the data from LinkedIn. Then I have to maybe modify it in some way, and then add it to my CRM system of sorts. And that itself is almost like programmatic thinking, which we actually saw that a lot of users struggle sometimes with, like dissecting the workflows, and they might get the orders wrong, and so on. So that kind of like, pitfalls, as you say, like traditional no-code development still has. What we’re trying to do now with this approach is take all that away, and like leave that to our model to decide on how to structure the workflow where you have to only describe kind of the actual workflow that you’re doing the same way that you would describe it to like your assistant, right? Like, the analogy we’re trying to build here is that virtually anyone should have, you know, like a mini assistant that you can talk to him to say, like, hey, you know, Michael, I want to get the current page as a PDF. And I might want to share that via Google Drive with the participants of the current meeting. But the instructions I could give to my assistant in the same way, I can describe this exact same workflow into Bardeen. And it then figures out like, okay, these are the specific steps that I need to do in that workflow; build automation for me, let me preview it, because I think that’s a very important step here to make it like trustworthy and reliable for the end user. So I need to be able to like actually see what the model is going to do before it does it. And then like, you know, when it gets everything – just hit the button, and the magic happens, and that’s really kind of the approach we’re taking.

Mike Vizard: Yeah. In your experience how unique are the different workflows that people are creating? Because I sometimes wonder if we’re reinventing the same wheel in different companies, and maybe it’s not really a differentiated workflow? So maybe we just need to figure out, hey, this workflow already exists over here, and you can use it and away you go.

Pascal Weinberg: Yeah, it’s a great question. So I think like two comments on that one is – turns out, a lot of workflows are structurally the same, but kind of unique. So we might, you know, both of us might when we do our research, before interviews, we might look at the same data sources, but you might store the data into like a Google sheet and might store it, right? So it’s not exactly the same workflow. It’s kind of the same skeleton, but it has different modalities or destinations or something like that, right? So that’s something that we found. And then now you have the complexity explosion of all the you know, hundreds of SaaS apps that are out there, and all the different combinations of those. So that’s something that makes this challenging where there’s like, kind of like rebuilding out of the box, all possible combinations of those workloads becomes very challenging. In fact, we tried to do that with our current approach. We ship the product with roughly 700, almost 800, pre-built automations. So at this point, that are exactly those, you know, from us, from our users, from our usage community, what we identified as the most common patterns in various situations – but then you still have this very long tail of like, you know, we might have, you know, pre-built automations for the exam, for that you might have. But with a different tool, it doesn’t drop for you; you still have to go into the build and like edit it. And that’s exactly the part where now we want to make make that last mile, so to speak, very easy. And that model that we have which is, in fact, trained on all those pre-built automations from us, and also from the user community to try to kind of leverage that collective knowledge to make it easier for people to do that. So I think like, yes, there is some kind of distribution, where like, the skeletons of the workflows are similar, but you have a very, very long tail of very unique combinations and very unique kind of instances of workloads. And we found that it’s very critical to be able to cover those, like long tail distributions, to make the platform actually useful for people. And that’s what we’re trying to do now with bringing like the usage barrier down, introducing a simple language interface, and let people just actually describe what they want to do and take it from there.

Mike Vizard: As you kind of think this through a little bit further, will we ever get to the point where maybe AI will surface workflows that are inefficient and some suggestions for optimizing them? How smart can smart get?

Pascal Weinberg: Amazing questions that kind of like, if you think about, kind of like bringing automation to everyone, like all end users, there’s a few different problems you have to solve, right? You have to know that what you’re doing is automatable, which, you know, most of the people in the audience probably kind of like technically minded folks that – like we kind of like tend to see repeating patterns and tend to say that, like, oh, this is something I can automate. Then like once you automated it, once you identify that it’s automatable, you need to like build or find the actual automation. That’s kind of the part we talked about before, I know what I want to automate now after actually building it, and then you have to be able to use it, right? It’s kind of like this three step problem. And the way we approach that is we kind of go back once or we say like, “Okay, let us be build all the common workflows, and then just make it easy for people to use.” So you bring it right to in the context where they are making it about extension, make it super seamless to use, make it you know, very accessible. Once you have an automation, then the next step of what we launched last year was the visual builder where it has to be now easy for people to build their unique workflows, once they know they have something. And it’s also kind of in the direction of where we’re launching now with magic box, the language interface is like, once I know that I have something that’s automatable, I can easily build it. But the last step that is something that we’re actively working on, but we’re not quite ready to release it yet, is what we call smart suggestions. And it’s essentially identifying that what I do is automatable. And the idea here is that if you’re already in the browser, we already see what the user does. And like, we can do that in a privacy preserving way. There’s no data sent to the server or anything, but we kind of identify with patterns of like, oh, the users, like copy pasting data from LinkedIn to Google Sheet. And like, you might be doing that for like, whatever we could think – building out a candidate list. And in that, in that moment, when you do that, like, you know, two or three times, then we want to kind of like pop up right at the moment, and say, like, “Hey, you know, Michael, like, here’s something that you can automate. And like, we already pre built that automation for you.” So we kind of like take you all the way down and you click the button, verify, again, very important step to build trust, and make it reliable; like verify that what we understood you want to automate is correct. And then just run it with a click of a button. In fact, we have that like, already, like rolled out a very small percentage of users that get like a video of this. And we’re starting to kind of drain the instances of like identifying repeating workflows and suggesting it to users. So yeah, that’s something that I think is a big, big, big opportunity here in automation space. And I always think it’s a really cool story, because the founders of Honey – you might know, like, Honey, the Chrome extension, which helps you save money when you’re shopping – I think it’s almost like the most brilliant user experience from an end user point of view that you can get, because you install the Chrome extension, and you might completely forget that they even exist, and then you are on a checkout page, and like when they can actually apply a coupon code for you that pops up in your face and go like, “Hey, here’s something; we can save you whatever 10% on this purchase, just click here. We’ll apply the coupon code for you.” And I think what you want to get to with automation is a similar user experience, but obviously, instead of saving you money, saving your time, which ends up being also, you know, money for most people. And it’s kind of this very proactive automation approach. So that’s where a bit of this inspiration came from. But yeah, I think it’s a very good idea. And it’s a very hard problem to solve. But I think like with the modern technology, and the way we approach to the abstraction layers, we build pretty close to being able to release that to the public. And we have some like demos and use cases ready already.

Mike Vizard: Do you think we need to revisit a lot of the so called digital business transformation initiatives that we’ve done? Because seems to me, they largely consist of somebody taking a paper based process and shoving it on a mobile device, and we’re still entering data and copy and pasting stuff. So you know, if we look back in time will we say, “Hey, you know, that first generation of digital transformation was maybe cute, but hardly the point?”

Pascal Weinberg: I mean, it’s a very philosophical question. I think there’s always different iterations to any progress we have, you know, in anything humanity. I think it is a great first step to bring the paper based process into digital and maybe just have the same, you know, workflow, whether it might be inefficient or not in the digital domain, and then think about, like, how you can take it to the next step. I think very rarely we’ve seen that people go all the way from, you know, like, having a very boring like, paper based repetitive process to something that’s, you know is fast and efficient in the digital domain, right? Like expecting people to make that leap, I think is a tall ask. So I totally expect it to be like an evolution over time. Having said that, it will become a big competitive advantage for companies, people who are ahead of the spike. Ultimately, most businesses succeed because they’re more efficient in some way, shape or form than their competitors, more efficient in production, more efficient and sales more efficient at scale, etc. And being able to like be an early adopter of a technology, I think, will become a big competitive advantage. So we will see that the winners of tomorrow are going to be those companies that take a very kind of automation first approach in their digital transformation processes. And maybe they rethink, as you suggested, like the whole process entirely and think about, like, “Oh, what are the steps that we can automate, end to end? Or maybe what are the steps, we can even completely get rid of that we might need in old school paper domain that we might not even need in the digital domain?” And like we hope to play well, with that – by the end to make it easier for companies and individuals to do it.

Mike Vizard: Alright, folks, you heard it here. Maybe in a world where almost anything is going to be possible, the biggest issue is going to be our lack of imagination. So we’ll figure it out from there, Pascal, thanks for being on the show.

Pascal Weinberg: Thank you so much. All right. And thank you for watching the latest episode of Techstrong.ai. You can find this episode and others on our website along with show notes. We will see you all next time.