How We Built It: Tabs - Fireside Chat with Harrison Chase & Deepak Bapat

My name is Harrison. I'm the co-founder, CEO of LangChain. We're going to be doing a fireside chat with Deepak Bapat from Tabs. I figured we could start with some quick intros, and then we're just going to have a fun conversation. We were talking before this about coding agents, ambient agents, UX things, lots of fun stuff. So we'll see where it takes us. But by way of introduction, my name is Harrison, co-founder, CEO of LangChain. We build developer tools to make it as easy as possible to build different types of LLM applications. So we started with LangChain. An open-source package. And now we also build LangGraph, which is orchestration for agents, LangSmith, which is testing and observability for LLM applications, and then LangGraph Platform, which is deployment for agents. And so that's a quick background on myself. Do you want to give an intro of yourself and what you're building at Tabs?

Yeah, for sure. Thanks everyone for having me. Deepak Bapat, CTO and co-founder of Tabs. Tabs is a vertical AI company in the finance space, Specifically, we focus on finance for B2B sellers, the revenue side specifically. So historically, when you think about the general ledger, there are a lot of companies focused on payroll, there's a lot of companies focused on expense, and then really the last, I would say, area of innovation has been in revenue. And so we are a vertical AI company focused on revenue intelligence for B2B companies. And yeah, happy to be here and talk more about what we do and how we've been thinking through our agentic experiences. And I'm incredibly excited because I think, especially in New York City, I don't get out to San Francisco very often. And so to have someone from San Francisco who's in the weeds every single day here is pretty awesome. So thanks for coming.

Should be fun.

Yeah.

What does revenue intelligence mean?

Yeah, it's a great question. So I think when you look at most B2B companies, the revenue team or most back office teams tend to be fairly lean. So you can think of a $50 million revenue B2B company has probably one VP of finance and maybe someone in the Philippines that's doing all of the billing. Really, revenue intelligence specifically is around how do we get the most money into your bank account as a B2B company as quickly as possible, as well as be able to report against it as effectively as possible. So when it all boils down, it's really just about saving time, collecting faster, and be able to report accurately against it.

And so where does AI and where do agents come into this? And maybe along the way, like how did Tabs get started? Started?

Yeah, so I can start with the first one. So I think Tabs started— we started around December of 2022. And the whole focus of the company originally, actually, if you go way back, was myself and Ali, my CEO and co-founder, we came from a company called Latch, which was a smart access product in New York City. And he was the chief operating officer there. And so The original idea came out of this need to bill specifically for hardware software devices. And in December of 2022, I think when that ChatGPT moment happened, for the first 3 months of our existence, I didn't write a single line of code. So I was brought on, I didn't write a single line of code myself and the rest of the founding team. We really just took a lot of CFO interviews and we realized the problem that we're solving for is a lot more ubiquitous. And so what we fundamentally do is we take, in a sales-led motion, we will take your contracts, extract all of the key information from those contracts, and run all of your revenue off of it. And so we fundamentally started and we pivoted into this space at the same time as the advent of ChatGPT because I think we realized very early on that At the end of the day, document information extraction, information retrieval was going to become a commoditized product. What we needed to do was focus on a very specific vertical to use that against and wedge in as quickly as possible. And that's kind of how we ended up here. What I'm actually— and we kind of talked about this before— what I'm actually more excited about than the information extraction piece of this is And Tavily did— the company does a great job of talking about the knowledge graph and this whole concept of Simba with the knowledge graph and organizational graph. What we are focused on is what we call the commercial graph. So if you look at— I mean, everyone here knows Rippling. They parrot this idea of the employee graph, right? So you can connect everything about an employee into your HRS system. I'm sure if you talk to Ramp, they would posit about a vendor graph. And what we posit about is a commercial graph. So the idea that we can, if we just hydrate this graph with all of the possible information around the merchant and customer relationship, we can run fully intelligent workflows off of it. And so where we've been really spending our time is on the agentic workflows that exist after the contract has been processed, invoices have been generated, how do we use that information to get the most money into your bank account and then report against it?

What I think is awesome about that, and I think mirrors a lot of the companies that are having success in spaces, is very vertical. It's solving a specific problem. You're building a data model that's like, it's the commercial graph. It's a very specific type of data model. And I think that is a, when everyone's talking about moats, that's a good moat, that's a great moat to have. Both the kind of data itself, but also the workflows that you build up around it. So I'm curious, as you start to build out these workflows, which I'm assuming are getting more and more agentic, how do you go about building those workflows?

Yeah, it's a great question. So I think for the first, so I think you kind of have to, and Ali, my co-founder, always says this, you have to know where the puck is and you have to meet the customer at the puck, but then know where it's going and be able to trend them there. The way we've actually built these workflows is interesting. It's been traditional software engineering for a year and a half, and now we've been layering on a lot of the, what we call guided AI experiences. They're in application. It's a lot of insight-driven, but the operator is still responsible for completing it.

Like a copilot?

Yeah, exactly. And then the other, the other alternative to that is much more agentic. So it is a proactive approach where the agent is doing something or asking you if it can do something. And so we started with the operational software. We've built up a bit of a guided experience. But really what we're trying to do is kind of leapfrog that into a fully agentic experience. And I think I mean, we talk about this a lot internally, which is if you go back 5, 10 years before the advent of this agentic era, you ended up in a world where there were a ton of great guided software experiences. So people would always talk about, we're a marketing platform that can drive insights for marketers, right? Or we are a sales platform that can drive insights amongst the data in sales on top of a system of record. And what we're trying to do, I think a lot of people in this room are trying to do, is almost vault over that step directly into agents, and we're trying to do the same.

What does the end state look like?

Yeah, we had this conversation earlier, which was— and I would actually pause it given our conversation. The honest answer, and I'm going to be very honest, is I think it's going to be an operational piece of software and a fully headless agentic experience, where the way I put it to my team, many of whom thankfully have come to support me, so I appreciate that. The way I put it to them is—

None of my team came.

That's not true, we got some managers. Yeah.

Two.

The way I put it is I want to build beautiful operational software that no one ever has to go into. And so fundamentally, I think the end state is, You've got a very lean team. You've got a CFO. You've got probably a VP of finance, very strategic individuals. You may, for a large enough company, have a small group of day-to-day operators. But mostly it's what we've talked about, ambient agents. And I really like how you put this. So for people who don't know, LangChain, I think in January, released this blog basically introducing ambient agents. And I think— A lot of us have been thinking about it, but none of us— I haven't seen another articulation of it in such a nice way, which is— and then I know you gave that talk. Everyone should check it out on YouTube, where Harrison gave a talk at Sequoia specifically about ambient agents and the agent inbox. My personal belief is that ambient agents are going to be the future. Again, one of the things maybe a year ago that we were really thinking about and trying to figure out is a lot of finance teams will go hire outsourced tactical operators. What does it look like? And all of their communication happens over Slack or Teams or whatever. What does it look like for us to bring that entire experience into Tabs where you are onboarding an agent similar like you'd onboard a billing specialist out of offshore team and the only way you're communicating with them is through a Slack or Teams experience. And I believe that the end state is you've got this lean internal finance team and a bunch of agents that are running ambiently in the background handling all of the tasks necessary. I think that's stage 1. And then stage 2 is actually moving up the chain to actually provide more insight and be able to generate things like board decks and interesting reports for the higher orders of what we would call skilled labor, although I do believe that all of it is in some ways skilled, and you find that out when you start building these things, in such a way that we end up in a world where you've got, again, you've got people focused on the strategic and everything else is taken care of. Care.

One of the things that I think has been interesting in the past few weeks is, like, in the past 2 weeks we were talking about this. I think there's been, like, 4 or 5 different coding agents that have launched as kind of, like, the background async long-running thing. So, like, OpenAI's Codex in the cloud, Google's Jules, CursorLot, which is kind of, like, their version of that. And then I think Anthropic today, I think, had a version in integration with GitHub where it kind of, like, runs in the background. I don't know if I'd call them ambient agents, because you still kick them off. So they're not triggered by events. But they're very much like background agents. And maybe that's a better word. And I think probably we'll have background agents before we have ambient agents, because the human— I think a key part of this is still having the human in the loop in some capacity. And so having a human giving orders and kicking it off, as opposed to just listening to events in the background, seems a little bit more manageable in the short term.

I think that's probably true in— in some cases, but I think specifically there are modes of communication that I think about, things like email, things like Slack, where really almost every workflow is kicked off either through a mode of communication or through some scheduled process. So you either have every 9:00 AM I'm coming in and I'm doing a series of tasks, or it's, hey, X, can you do something? Specifically with with what you're talking about, I think the weird— outside of the cursor thing where it's in your IDE, I think a lot of it comes out of the fact that it's disconnected from those modes of communication. So for us specifically, one thing that we're really working on right now is being able to— for a finance professional, being able to monitor all email communication between you and your customer. And being able to kick off what I would consider, yeah, it's a background agent, but it's also an ambient agent that is listening to that and is able to drive action. Whereas if the Codex in the cloud, for instance, is something that is kicked off from someone saying to someone else or from someone having a thought. And I think that's fundamentally the shift that takes you from background to ambient.

Like you're saying, like there'll be some email communication from your customer and rather than you saying, hey Adrian, go process this, it will just do it automatically and that's the trigger, that's the event.

[Speaker] Yeah, and actually if you were to go into Tabs today, we actually don't have, I think I mentioned this to you, we don't have a chat experience at all for our Agentic experience. Everything is done ambiently. And part of the reason for that is like when we started Tabs, one of the things that drew me to the idea of it is like, I love the idea of bringing the way developers work. I mean, maybe this is somewhat of a cocky thing, but I think developers are really tuned into efficiency, and we are fairly efficient. Good ones are fairly efficient on how they work. And so bringing developer processes to other areas of an organization. And so things like running sprints, Kanban, how do you do that in the finance stack, specifically in revenue? And it became a question of, Let's say you have two invoices and one of them's a $10,000 invoice that might get paid and a $100 invoice that is definitely going to get paid if you reach out. Which one do you prioritize? And so everything became a— I got really excited about the idea of prioritizing tasks in tabs. And then after listening to your talk, I realized, OK, maybe this isn't a prioritization problem. Maybe it's just a scale problem. What if we just had— endless number of parallel threads that could run and do all of these. It didn't matter. We don't have to prioritize anything. Really, the only places where there are gates then are in moments of ambiguity or in dependencies on each other. So in finance, there's a concept of month-end close. You can't really do your month-end close until you've figured out what to do with all of your invoices and you've checked all your customers. There are dependencies there. And once those are cleared, you can do the next step, or in the ambiguous case. But outside of that, you should be able to do any number of tactical things with these agents.

So what does the interface for these agents look like? Because you mentioned that they kind of just exist in Slack, and so that presumes kind of like there's no interface. But I think there's also some aspect of if I want to see what an agent has done, Is that also in Slack? Where's the audit trail? Are you building an interface for that? I think I've also seen some email assistants where you just add an email address to the thing, and that's how you trigger the agent. There's nothing new, it just kind of lives in email. And so I don't know if you have thoughts on this, but what is the interface? And then I think another really cool version of this that I've seen some people say on Twitter, especially recently with these more ambient ones, It's like playing a video game. Like you just have, like, you just send your little agents over here to go code, and your other little agents over here to go market or something like that. And so it's more like StarCraft or something like that. It's like a futuristic version. Do you have any thoughts on that?

Yeah, it's a great question. So I'll talk about the first one. When it comes to— and by the way, I was terrible at StarCraft, and I used all the cheat codes in Warcraft II. But fundamentally, when I think about the user experience and the user interface for these ambient agents, I think it's two places. I think it is in the modes of communication. So for instance, oftentimes I will— if I'm writing an email to a really important customer, I will oftentimes have it checked by someone. And so I'm not doing that in some sort of portal. I will go and send the Slack message to someone saying, hey, does this look right? Or I'll put it in an email and say, this is what I plan on sending. Does this look right? I think the same thing is going to be true for some things. So for us, for instance, one of our agents that talks over email will send a draft to what we call our merchants, which are our customers. We call them our merchants. We'll send a message to one of our merchants over Slack saying, we received this email. This is what we'd like to— send out, does it look okay? And they can say yes or no or whatever, just like you would in a normal experience with a person. But I think in terms of an audit trail, and again, at the end of the day, a lot of what we do is pattern matching and derivations of what already exists. I think Linear, we were talking about, just released an agentic experience where you can tag agents on Linear tickets. And you can have integrations with agents and you can give them linear tickets. I think specifically for what we do in the finance stack, I think being able to audit things off of like, for instance, tickets or audit things off of a location that also manages for humans the prioritization of tasks, that's where I see the UI being built up. In some ways, we're moving away from a traditional read-only audit log to a more collaborative environment like in Slack or in an interface where you can leave comments and you can talk with these agents or talk with each other.

So are you guys going to build a project manager, a Kanban board type thing?

Yeah, and this is what I mean by what really excited me about Tabs. 2.5 years ago as well, it's building almost this developer experience into the finance stack. I think that is probably where we're going. We're already building an agent inbox and we can actually leverage that to do a lot of the task management as well as leaving comments. That's part of the grand scheme of how we're going to interact with these agents, but I think that is a really important thing.

A lot of this is like human-to-agent communication.

Yeah.

You talk also about kind of like going up the stack, which to me maybe suggests there's a manager agent that's managing its subagents and stuff like that. Does that also happen in these interfaces that are designed for human-to-agent things? I guess you could say the pro would be then if I want to go in and look at an audit log, it's really easy because I can see the manager saying this. But like, you know, these are communications that we built For humans, agents are different in some ways.

Yeah.

Does it need its own— will there be something different, or will they just communicate on Slack to each other?

Yeah, it's a great question. I don't think it's going to be necessarily through Slack, although I might be— I'm happy to be wrong about that in the long run. I do think that specifically for our user base and our industry, which is finance, and one of the things specifically about— accountants and finance people is like, the honest answer is when the numbers are wrong, they are left to blame. So what they really want is they want a place where they can go, they can click a button, and they can say, oh, like yes or no, like explicitly. I do think even for the manager agents, like if we start to wrap everything into this task or workflow model where it is, there is a UI component as well as a backend workflow or backend agent component, At least for us, that makes the most sense for our industry and our vertical. And so that is where we are at today.

Talking about where we're going.

Yeah.

How do we get there? What's the hard part of building these agents right now? What are you guys like? Yeah, where are you hitting roadblocks? What's the hard part?

Yeah, I think there's a couple of places. One is, I appreciate the conversations that just happened right before this presentation. I think one of the things is you can hydrate your knowledge graph as much as you want, but it's really about knowing what is the most important stuff to pull. For instance, for us, knowing customer behaviors as part of the knowledge graph, how do you know what is important and what is not, what is relevant and what is not to a specific specific task, I think is actually one of the hardest challenges we've had. And we are spending a lot of time thinking through how do you— information retrieval at the end of the day is so important to any sort of agent that is going to run independently of human in the loop and for any agent that is doing something that requires minimal human intervention. Human in the loop is fine, but what you want to avoid is a situation where Every single time the human is having to correct the same thing. And so it's like, how do you start to store that memory? How do you start to add that memory to your knowledge graph so that next time around the agent knows what to do next?

And do you feel like the blocker there is figuring out how to add things to the knowledge graph or how to retrieve from the knowledge graph or both?

I think it's— I think it's exactly—

I say knowledge graph, but like maybe you don't necessarily have to.

Yeah, of course. I think it's— I think For us, I think it's a combination of structuring because there is a ton of unstructured data in a B2B enterprise. So we need to gather everything from sentiment analysis of the customer to usage data to what's in the contract, to what has been changed, to customer communication. There's just so much data that you could possibly hydrate the graph with. Not to mention across different merchants, being able to aggregate that data to see trends and gather insights off of that. That is a huge one. So I think structuring and pruning is important, and then the retrieval piece is also incredibly important.

And I imagine probably a lot of that is probably domain-specific. You mentioned things that I would never have thought of.

Yeah, I think that's right. I think every vertical AI company is going to experience their own— I think everyone's going to— the hydration logic, once you start hydrating it enough, it is inevitably going to become an information retrieval problem. And then beyond that, I think there is oftentimes a problem of remembering what the last thing you did is. Exactly what the last presentation was about so that you don't have to continue to ask over and over again. You shouldn't have to read— like after I've drafted 3 emails to 3 different customers, I should know generally what my CFO wants me to write and I shouldn't have to go back to that.

And I think that's like, I think that's an underrated point of like the background or ambient agents as well as like background or ambient doesn't mean like fully autonomous. Like I think the human in the loop is important not only for getting a better result and for, and for trust, But that's because that's how you learn. If you don't have these interactions, you don't learn. And so I think designing the right interaction pattern, maybe it's Slack, maybe you let them modify the draft, like in a draft as you see in an email, and then you just track the changes that are made, and then you reflect on that. But whatever that step is, I think is actually really important for a bunch of reasons, and memory is one of them.

Yeah, that's exactly right. And I think, are you doing memory right now?

Are you doing this learning, or is that more on the, future roadmap?

Yeah, I think it's generally on the future roadmap. Right now, what we are focused on is getting the core functionality in a spot where, look, when we think about general building of these ambient agents, I think honestly what customers want today is to see that you are trending in the right direction, especially for a lot of our customers. At this, like, what we are focused on is twofold. One, hydrating that graph with everything we possibly can. I'm okay having data in there that doesn't even matter in the long run. And then two, starting to build out how we start to retrieve the first version of the information, because that is the hardest problem we have, is getting the relevant information into the agent. And then oftentimes what we can do rather than doing like Memory is just putting that back into the graph and then trying to figure out how to make that the relevant information that we grab for the next run.

Do you think memory will be a moat that companies can kind of build up over time?

Yeah, 100%. I mean, I think it's like your best employees, right? Like at the end of the day. And I think if a company can figure out how to enterprise over enterprise, how to manage that, I think that is a huge moat. Every enterprise that we work with has a different way in which they do business. And being able to do it on an enterprise-by-enterprise basis is important.

So I think I agree with this, but then ChatGPT added memory. And I feel like it hasn't, at least for me, it hasn't improved the experience for me. That much. And I think I've seen a few people on Twitter saying the same. I don't know. I don't know if you disagree and it has improved the experience for you or if there's something like maybe that's too general purpose. And so general memories, I do think general memory is hard to get right. But maybe like, yeah, this specific workflow is more tangible.

I think that's it, right? It's like if you're in a vertical, if you're a vertical AI company, like you have very specific guardrails around what you care about. Whereas I mean, I don't use projects in ChatGPT, and I just continue to write in the same thing, every, like my stream of consciousness. So it doesn't help anything.

I think we're nearing the end-ish of time, but we probably have time for a few questions. I don't really know what the time schedule is like, so we can make it up. But if there are questions, we can maybe take a few for—