The Frame AI Acquisition

Foreign George, welcome. Thanks for being here. Thanks, Kyle. And hi everybody. I'm really excited to be here. Kyle told me about this. I think I was, I think I'd been at HubSpot for two weeks when I first heard about this group. But it was, it was a great example for me of why it was exciting to join HubSpot. Because, you know this, there's this great community co investment. Right. I think everybody I've met at HubSpot Cares a lot about hearing from the community. We have folks like Dana helping organize these communities from the integration side and then all of you as admins. And so I'm super excited to be part of that. I warned Kyle that I'm not here to sort of announce any product features. So if something even sounds like that, it's probably a misunderstanding. Count on the roadmap for that. But I am very much here for questions and feature requests and so on. And so I have a few slides to go through to kind of share what Frame AI was and what it's becoming now and sort of what that might mean. But I hope we'll have most of the time to chat and ask your tough ones. I guess Kyle also already warned you that I can get pretty nerdy. My background's in machine learning and AI. I love what I do and feel free to tell me to cool it down in the chat if things get too spicy from a technical perspective. But I'm going to share my screen here. Can you all see my. Kyle, are you seeing this? Yep. All right, great. I'll start just from saying, even prior to Frame AI, this really motivating question for myself and a lot of folks who work in AI is just how amazingly good humans are at working in a world with no instructions, no configuration, nothing. So I love this photograph because I have kids. I have a two year old right now and I can tell this little girl on the right, she is learning, nobody had to tell her exactly how that she should watch this kid or how she should watch him or give a thousand examples. She is learning by watching and observing what's going on in the world around her. And she's going to be better at buttoning buttons because she's cribbing notes from this fellow who's maybe a year older. Right. And humans are just incredible at this. We communicate with each other in open ended ways. We watch what each other are doing, we learn from experience and we bring that knowledge into our work. And so for a lot of SMBs, that's what your employees are doing, especially In a small company, everybody wears lots of hats and they learn from each other. And it is really kind of ironic that software is the opposite of this. Historically, our SaaS software. I think you all as an admin group probably feel this the most. Historically, everything that the computer can do needs to be extremely manually configured and input. And even in previous generations of machine learning, lots of training, et cetera, you have to really handhold and set up a system to be able to do work in a much more brittle way than a human can learn to do it very naturally. And so there's flip sides to this. I mean, the advantage of these structured systems is that when we spend time configuring computer systems really explicitly, we have control of them, we know what they're going to do and we have scalability to them. Right? They can kind of amplify what we're capable of. But what if we could have both of those things? What if we could have sort of flexibility, learning from experience and also have some degree of control and scalability? And that's, that's really what our goal was at Frame AI. It's to help use all of the unstructured information that businesses generate to produce, to bring that into people's workflows and processes and basically make their businesses run more effectively. So when we were an independent company, we built an architecture, we actually weren't facing SMBs. Mainly that's part of my adventure at HubSpot is learning what it means to bring these capabilities to a much broader group of businesses. We were facing enterprises and saying, hey, we're going to take your email threads, your sales calls, your support tickets, your chat conversations. We're going to, in much the same way that humans working with those customer facing channels, learn about your business and learn how to, how to conduct operations, whether it's sales or service, etc. We're going to set up AI to get that information out of those data sources and into operational processes that help you do more work. So some examples on the left, these are a bunch of different sort of enterprise sized companies that you might recognize that we worked with. On the right, you have a bunch of the capabilities that we helped build for those teams. This frame was a platform where we could bring unstructured data like emails and calls and so on into our system, analyze it with AI and push it out to drive these different features that included things like assessing sentiment, escalating service calls that needed help, figuring out what was driving the main costs in your operations QA and coaching for customers. A lot of features that you've probably seen in different point solution platforms. We were doing that across data sources that enterprises had for the very complex workflows that they had built across multiple systems. That's the origin story of Frame. It's this kind of building a platform to use unstructured data to help customer facing teams be more efficient through a wide variety of features. And the way that we did this. And by the way, I'm losing Kyle, I am. I don't have the chat when I'm looking at my slides here, so that's okay. They are not behaving themselves. Why are these getting a slowdown? Poindexter? No, no, no, you're doing great. They're. They're just. You showed a screenshot of Salesforce and they are all just Talking about how HubSpot is so much better than Salesforce. Oh, certainly, certainly didn't mean to cast any aspersions there, but yeah. So the way that we did this if depending. So I imagine some of you in this group have heard about a technique called Retrieval augmented generation or rag. So when you use a lot of AI features that you use in HubSpot or in other platforms work something like this, you do something in the interface and then behind the scenes it goes out to a library of information and it says, hey, what do I know that's relevant to this user's query? Let me get that information out of your CRM, out of whatever. I'm going to bring it in, I'm going to rewrite the thing you asked as a prompt, send that over to an AI and then give you back the answer. A lot of features that you might use in an interactive basis when you're using, if you're not using ChatGPT or something directly, you might on an interactive basis be plugging into a tool that then calls these APIs under the hood or has its own LLM that it reaches out to you under the hood and you wait for an answer. I bring this all up because it's not actually what we do. This is Retrieval Augmented generation. It works really well if you've got users coming in with queries that they want to respond to. What we did that was a little bit different is what we call Stream Trigger Augmented Generation or STAG as opposed to rag. This was this different idea that instead of waiting for a user to query something, we were going to keep our own catalog of things users care about in advance. We were going to know, hey, product users care about how different features are being used and how their Sentiment is accumulating. Service managers care about things like whether their staff are being polite. Essentially we have this list of queries that we know are important all the time and that the people who care about them, they really care about them in almost any kind of communication that comes through, not just in the tickets that they can read. So we would take all of these different streaming communication sources and pipe their communication into this fixed set of queries. And when we found something interesting, and I'll say something about how we decide something is interesting, we could trigger alerts and push out information that then could kind of trigger a workflow downstream. So the big difference between this and. You don't have to understand this diagram to hopefully to get anything out of my rest of my talk. But the big difference is that we were doing proactive AI. How can we automatically analyze data that's coming into your system, communications, documents, et cetera, for the things we know you care about as a business, and then push that information actively to where you need to go, maybe as an alert, but maybe also as just preparing that information so that when you, maybe the next time you go and you look at that account, we've already prepared this entire list of the new concerns that got raised for that particular account or the next time you go to look at your dashboard for your, for your service, for your customer service operations, you can see that there was a big spike triggered by a new topic that you never heard of before because all your customers were having login issues or something along those lines. So we were doing that kind of proactive analysis. And this last, and this is slide on here, I think. But the really cool thing is you can chain together these different queries to build up really capable features. So this is again, we're not announcing any features, not describing them. This is actually kind of a schematic of a system we built for a pre acquisition customer. But if you know in advance some of the things people care about, you cannot just call one LLM query, but actually chain them together to extract information that you care about on an ongoing basis. So for example, as calls come in, we can use what people are talking about on the call to build a profile of your business and say, oh, these are the products you sell, this is the overall differentiation that you have, et cetera. And we can use that business profile to power more processing downstream. We can say what are customers common objections that they are bringing up in relation to this business? What are the sales policies that are being, that we can observe taking place on your calls? And if once we've extracted that background information, we can push it forward so that we can make better answers to questions like, what's my next best action on this one account? Right. If instead of just saying something generic that any salesperson at any company could do, can we bring forward the context that we've learned about your company to make a much more targeted version of that same next best action? Again, I'm flying through these. We can dig into details here, but the type of work that we did, it's not an either or. Between the proactive data analysis that we were providing or the reactive AI that you might be used to, the two of them actually work really well together because we can at Frame, build up a set of information about your business, about your deals, about your customers that provide really valuable context. And then when you come in to use a feature that's specific to what you're trying to do in HubSpot, maybe you're in Sales Hub trying to prepare for your next meeting, or maybe you're using Copilot to ask about all of your deals in Florida. Instead of those tools starting from scratch with no context about your business, they have this context that we've built up proactively that can plug into those tools. This became the vision of what we could do at HubSpot. We basically are working in the background as a proactive processing layer to analyze communication and bring up this context that's specific to your business and your operations and your customers. That then becomes part of the background information for other AI features. So you're never going to see a feature in HubSpot that just says, oh, this was Frame. This is what the. This is the feature they bought Frame to build instead. What we hope you'll see, and I'll get into some more details of this, is that across all of HubSpot, lots of AI features benefit from having more context about your business. And you can really tell the difference between, hey, by doing this in HubSpot, in our AI, I have all of this context built in that's helping the AI give a better answer as opposed to if I just ask a generic chatbot in some other experience, hopefully a lot of you saw a lot of these things. A lot of AI features come out in Spotlight a couple of weeks ago. So I think there were over 200 features announced. Probably about 80 to 90 of those involved some degree of AI. One of the first things we were doing under the hood for those releases is actually using knowledge about a company to improve transcription of phone calls. Right. Like historically, when you look at transcripts and you feed a raw phone call into a transcription engine, that engine doesn't know what your business does, what it sells, and as a result, it mistranscribes things that are brand names and the people's names who are involved in the call, et cetera. And we can make a massive improvement to that by bringing context into the transcription AI. And that has effects downstream when you try to use that transcript to build other features. So that's pretty much as far as we've gotten since the acquisition in terms of integrating deeply into the AI features at HubSpot. But when it comes to inbound, you're going to see a lot more exciting stuff.