Dominik Facher

Head of Product, Chorus

Bootstrapping AI into B2B Products: How Chorus Kickstarted a B2B AI Product

Dominik Facher: Welcome everybody. First of all [inaudible] I think we're rolling. We're running a little bit late, but I'm, I think I'm going to get you out before lunch. And then, so we're going to spend 20, 25 minutes just talking through AI and how to get AI into a B2B product. And then after, once we have time for questions, um, so if I jump forward a little bit, um, why does AI matter? And there's certain companies and certain types of problems where AI really makes a difference and so of course happens to be one of them. Um, and I'm going to share a little bit about that and then we go into like specific examples of what we learned along the way. Um, chorus, like really briefly, what we do is we are a conversation intelligence platform and what that means is we work especially with sales teams capturing and then analyzing their conversations. So if you're running a sales team, one of the problems that you have is basically what's on here.

Dominik Facher: Your entire team is spending most of their day in conversations with the market, with prospects. And so, um, the problem around this is that most of those conversations are really a black box to you. So you imagine you're running a company, you're spending hundreds of thousands or even millions of dollars in marketing, those produce great leads. Now they hit the sales team and then now the sales team walks them through conversations and through the funnel. Um, the challenge is when you ask your VP of sales, hey, why are those conversations not closing? It's actually a really, really hard problem to answer, right? And that's where you need AI. So if the VP of sales wants to give you a good answer, what he actually has to do is sit in on every single conversation, try to figure out what's the difference between my top rep and someone else, and why do some of those conversations close and others not?

Dominik Facher: Right? So like, you actually have a complete black box and you don't know what's happening. And so when you look at the numbers, you'll see that most reps like 57% of reps, they don't make their quota. 80% of deals don't close. And the moment the conversation hits your sales team, it costs about a hundred dollars a minute to run this right? So that's like a really, really big problem. And so if you asked your VP of sales to solve that problem, well this person would have to sit on every call, which is technically impossible, especially if you have a team that's a thousand people or even more. And so that's one of those types of problems where you need AI, you need something that actually sits in, listens on every conversations and tries to figure out the patterns that help you close deals or not. And so when we, when you think about this, um, I want to walk you through a little bit of the things that we learned along the way and how we approached this problem.

Dominik Facher: Um, if you think through startups, startups are essentially just a function of theorists [inaudible] right? So you're, you're starting a company, most companies, there's two types of, there is technology risks and there's market risks. For a typical B2B SAS company today, there's no technology risk. It's just market risk. Technology is out there. You built a product. But what do you have to do is you have to try to get it into the market. And so there's established playbooks, lean start up, build, measure, learn. So you've run small experiments, you try to get in front of customers and you figure out what's most helpful to them. You're reducing market risk at that point. And that's really important. Now when you start an AI first company, that's actually somewhat different. You in addition to that, have technology risk and that fundamentally changes the economics of your business. So if you're an investor investing in an AI first company, you actually have to know that there's technology risk.

Dominik Facher: It takes time to build AI from scratch. Models take time to learn. You can't just go to a customer and then all of a sudden you have this same sort of basic product and it just works and you learn over time how to improve the user interface. You add additional features, actually have technology that you have to build. And so that changes the economics of your business. And so when we think through that, I'll just walk you through some of the things we learned along the way. Um, and what technology risks means. The first thing is it actually takes time to build. So, um, the typical SAS company takes about six months to go to market, build a product, go to market test, and then actually put a website up, um, and get the first set of customers. Um, of course it took us one and a half years before we even had an MVP.

Dominik Facher: And so you would actually say, hey, that, that's completely stupid. Like, why would you hire a big team? Why would you spend one and a half years building something when you don't even know what works or the market even wants that? And so in our case, we have to actually be convinced that there's a big enough problem that if we build it, they will buy. And that for us was, hey, if we can actually help teams understand what a good conversation looks like, why certain conversations close and others don't, that's actually really meaningful. And so we got early validation on the level of like, Hey, yes, I would definitely buy this. I'm happy to help you build this, but it takes a lot of time to build this MVP. So here's an example of how that looks like today. So this is a dashboard that you would get from chorus of course automatically captures all the meetings that your teams have.

Dominik Facher: It then understands what the topics are that you're talking about. It maps that back to your sales process because every team talks about things differently. There's different features, there's different elements. It then figures out what are some of the things that actually are correlated with winning versus not winning. What is the difference between a top rep and a bottom rep and how much time do they spend in each of those sections of the call? Sort of you can get an understanding like in this particular case, um, this is a top rep on the right and someone who is not as good on the left. And then you see they're actually spending more time talking about uncovering pain with customers. They also spend more time on specific customer stories or a specific integration like this is actually our own dashboard, like this is our own data that you're looking at here right now for all in sales team, and so what we had to build in order for like to literally get to that level of MVP, we had to, we learned that we had to build our own speech recognition because the speech recognition that you can get from the market is not good enough to actually get meaningful insights and meaningful signal.

Dominik Facher: So we had to build a system that actually learns on business conversation as opposed to the general conversation or the general speech recognition. Then we had to build a self-learning system that actually starts mapping and deconstructing your sales process into themes so that we can map out like what are the different sections you're talking about, pricing, discovery, the types of questions that you're asking. And then we had to go even one step further and figure out each customer now gets their own speech recognition because it turns out that the features you're talking about competitors, all of those things, they're actually very unique to your business. And if you don't have a speech recognition that picks up on those signals, like if you get your competitor name right or wrong, how on earth can you trust the data behind it? And so you have to build a system that actually learns what are your competitors, what are the features you're talking about and so on and so forth.

Dominik Facher: And so this is actually, this was necessary for an MVP for us in order to solve that problem. And it took us one and a half years to build that end. We're anything but done. Um, but it's good enough for an MVP. And so one thing we learned along the way is you have to actually design intentionally for patients in that period. You need a team that doesn't care about not having a product in the market that's just self motivated to work on this. You need investors that don't benchmark you two typical SAS metrics or anything like that. They need to be okay that you're spending millions of dollars not on marketing, not on sales, not on going to market, but on building a product. And that's very different. And so there's only, there's only a few investors out there that actually want this kind of have an appetite for this kind of risk.

Dominik Facher: Um, and then it's also a different launch. Your product might just be an algorithm in the background that might not be a fancy front end to it and you have to be intentional and you have to design for that patients because some of that stuff takes time. Now let's assume you went through all of that, um, technology risk, um, and you captured that. So now what's next? How do you attack the market risk? How do you actually start learning? And so here's some of the things that we, that we went through. So the first thing is if European, and you think about an AI first company, every single one of your roadmaps needs to have a swim lane about data capture. Because the one thing that's different about this type of company is that you're actually a proprietary data set. And so you have to think about product in terms of how do I build something that captures this data set at scale.

Dominik Facher: And in our case we thought about it in three different ways. The first one is you have to build a new sensor. And what I mean by sensor is you have business conversations, you have a Google calendar, so your meetings are scheduled on a calendar. Then those meetings take place like likely on a zoom meeting. And then as a sales rep, you're supposed to take those meetings, those insights and put them back into your CRM system. There's nothing out there that ever connected the business outcomes of what's happening on your calendar or your CRM system with a content that's happening in meetings. And so when we thought about chorus, we thought about what's the sensor that actually understands what's happening in your conversations and maps that back to business outcomes. That's a data set that just hasn't been there before and that's completely proprietary. And so you need to build a system that actually connects those data sources to create something new.

Dominik Facher: The second part is built around the workflow. Most of the time you will have an existing workflow that people anyways are going through. And so what do you have to do is figure out how do you take that workflow, how do you augment that and capture your data along that workflow? Because chances are they're going to do that multiple times a day. And so in our case, we figured out that the biggest workflow and the biggest problem is actually CRM note-taking itself. So you run a meeting afterwards, you're supposed to take notes from that meeting, put it back into your CRM cause it's the system of record. It turns out that there's a lot of information that happens in a meeting. A tiny fraction of that ends up on the notepad of a salesperson. And then a tiny fraction of that ends up back in your CRM and that's your system of record.

Dominik Facher: That's literally the source of truth of what you make decisions on and so on and so forth. And it's a tiny fraction and every sales person hates updating the CRM because it's just busy work. And so we just went for that process. How do you actually take notes for rep? And so we automatically connect to their meetings and afterwards we'd give them the summary. And the two use cases we started with and that got us to a couple of thousands of users was very simple. The first problem is competitors. And the second problem is next steps. So most only one in 10 competitive deals are tracked properly in a CRM that annoys sales, that annoys marketing and probably anybody else. And so we just built, if you run porous, we will automatically put your competitors back into the CRM accurately. That's it. And on top of that, we will create an activity for every meeting that you have automatically in your CRM system. And we'll put the next steps into it so that you don't have to write them out yourself. That's it. And so sales reps love that. And only by launching that product was part of a workflow. They have to do anything else except for signing up for the product. We made the CRM data Beta better. And so with that we actually had the, they allowed us to capture the data that we needed to build all the rest of the AI.

Speaker 1: And that's it.

Dominik Facher: And then the last part we learned is humanize your AI. And what I mean by that is your algorithms are probably not going to be perfect. It's very hard for someone to emotionally connect to an algorithm or something that runs in the background as opposed to like an actual product that you have in front of us that you can fall in love with or that even has a great UI. And so what we, what we learned in what we invested in was customer success early on. And so every time we deployed something, yes it was 80% of the background, but there was a person that actually really cared about the user and about the customer. We did like very personalized onboarding. We followed up with people. And so they always build a connection, not just with them, the algorithm that runs in the background, but with a person that cared about the results. And that actually helped us because we got so much feedback we could learn and so on and so forth. And that made all the difference in the world for building out all the technology along the way.

Dominik Facher: All right, so now you are at a point where you over invested in technology, you probably have less customers than a typical SAS company. You have a lot less revenue. So why is everybody kind of excited about this type of business? And the short answer is it's the metrics. And so any business, what you're looking at is like LTV to CAC. That's just the metric for any SAS business. And so the belief is that with building a different type of company, you can actually meaningfully change those metrics. In the long run. And so the one thing that you're doing differently is you're building this proprietary data set and that doesn't just apply to your product. It actually applies to the entire funnel as well. So if you think about marketing, you're organically building a data set. And that's really, really interesting for everything that has you have to do with customer acquisition.

Dominik Facher: So in our case, I'll give you an example here. This is just something that comes out of our product organically. So it turns out that during a deal, if you ask between two to five engaging questions, it's 60% more likely that you advanced the deal that's actually meaningful. And so you can use that, you can plug that into content marketing, you can use that as an asset through the entire funnel. It's organically coming from your product. So it's almost zero cost, but it meaningfully changes your customer acquisition costs and your likelihood of getting someone through the funnel. You can run almost a consultative sale that, and so you have data that is so meaningful and it's coming organically from what you build, any ways you use that.

Dominik Facher: Now let's talk about the value side. So the idea, if you build your algorithms right and if you build in flywheels, you actually get better over time, your product gets more sticky. I can give you an example. Every time someone uses chorus today, it gets to a specific meeting or a specific moment. Shares a moment, gives feedback on something. Anytime they engaged with the product, we learn why they care about a specific moment, what's the content behind it? And we use that to recommend better. Like we drive our engagement through that. And so the product gets better over time. It learns how to be a better assistant. We get more data points that we plug back into the CRM, so it gets more valuable. Chances are your business is going to be more sticky than the average SAS B2B company. And so over time you have a higher chance of upsell.

Dominik Facher: And when you look at the underlying metrics, the ACV, the stickiness, et Cetera, you're performing or hopefully performing better than a typical SAS company and running that through will actually give you better metrics in the long run. Um, so takeaways, um, designed for patients, data capture needs to be the one key thing that's on your roadmap all the time. Think about it intentionally. Um, use your data in your customer acquisition funnel and then build in those learning flywheels that actually help you make your product better with almost zero effort over time. So higher stickiness and probably higher upsell. All right, that's, that's it from, um, from our takeaways. And then I would love to have the rest of the time just for questions or areas you guys want to maybe get deeper into.

Speaker 1: Any questions? Go ahead. Yeah. How did you get your data before you had [inaudible]? You just like doing fake sales calls? Like...

Dominik Facher: Yeah, it's a great question. So we connect to the CRM and so every time someone runs through a call, we actually immediately get everything that is in the CRM. And then we also already had the transcription engine. And so what we did in the beginning is even if there's just 10 calls in the system, you could actually see how, um, that deal progresses and you could map it to a specific content. Um, in the conversations. And so we started with a couple of hypothesis, for example, competitors and next steps. We built those algorithms before we even connected with customers. So a system that automatically identifies those, um, those items. And then the moment someone runs their first call through it, we could actually map it to the CRM. So we had labeled outcome data from day one and we had features that we built before we even started engaging and that, that worked.

Speaker 1: Okay. Go ahead. All the types of [inaudible] a small [inaudible].

Dominik Facher: Yeah, yeah, that's a great question. So we have a pretty large research team, um, and are, they're basically doing exactly what guy taught before and navigating through the different types of algorithms that you use. Um, when it comes to outcomes, most of the algorithms are actually pretty basic. Um, what's more important? And we spend probably 80% of our resources on feature engineering. And what I mean by that is you need the data points first. So more data points and more data gets you to better outcomes. So I'll give you an example. We do speaker separation. That means there's a meeting and there's let's say five people in that meeting. We have algorithms that figure out who's talking at what point in time. That took us over a year to build. And all it gives you is like a timestamp and a person. And then you literally just take that time seminar person and you correlate it to the outcomes. You can say, hey, if you engage with the decision maker for more than two minutes in this conversation, you're more likely to close the deal. Extremely simple linear model afterwards because you have that data point and you have the outcome and all the work, or most of our work actually goes into feature engineering, which is how do I actually detect that person? How do I build, we call them voice fingerprints for four people. Um, and that's all the different algorithms that we saw before. Um, and that's a really hard problem to solve.

Speaker 1: Yeah, go ahead. [inaudible]

Dominik Facher: Yeah, it's a great question. So the question was like, do you go broad or do you go narrow? Especially when it comes to like business outcomes, like CRM data, which is very noisy. Um, to begin with, we always took the approach of like, making one customer really successful and then moving onto the next one. So we went intentionally deep and it turned out that we were lucky in the beginning because a couple of the models that we build, for example, around next steps or competitors, they were fairly generalizable and you could go with different customers. Um, now we're at a point where customers actually approach us and say, like, our CRM data is so messy. Can you please make it right? Can you please fill everything out? Um, and we're not quite there yet, but that's the, that's the way to go. Um, and yeah, we were very intentional on like one customer at a time.

Speaker 1: [inaudible]

Dominik Facher: Um, we don't do that. Um, we, we thought about it initially, but we immediately figured out that like, if a salesperson has to do anything in addition, it's not gonna work. And so that was, that was painful learnings. And so we built the entire workflow to capture automatically whatever data we need. Otherwise you'll, you'll don't get it to scale. And it's probably noisy as well. But yeah, so that's the only thing we use. Yeah, go ahead.

Speaker 1: [inaudible] Yeah, there's a lot of voice app API [inaudible] that area where they're [inaudible], uh, API that are, um, kind of [inaudible] again that are available.

Dominik Facher: Yeah. And even in that case, we build our own speech recognition, um, because we've learned that it needs to be better than what's currently out there in the market. And so, um, it comes back to how domain specific is your application. And so Google is doing a fantastic job. IBM is doing a fantastic job and yet this wasn't good enough for us. We could never build anything that's as good as what Google built, but what we can do is we can be very narrow in our domain. And so in our case it was speech recognition for business conversations. And then even in that, it's like sales conversations. And so we had to build our own speech recognition to be more accurate than what you get off the shelf. Um, the other core things that we use, speaker separation, um, there's no API, uh, for that. And then, um, so we built most of our tech in house is the short answer.

Speaker 1: [inaudible] One more question. If there isn't an API for that, obviously [inaudible] larger market [inaudible] many other roles that are developing technology, how did you get in, come out of like, oh, maybe you should you sale [inaudible] you get back [inaudible]

Dominik Facher: Yeah, it's a fantastic question. Um, we think about this stuff long and hard all the time. And in the beginning, um, we, we actually had a like very specific project. I'm thinking about that and we discovered it's too hard of a problem for us too to actually go abroad. Um, the only way, like with our limited resources we can actually make it work and to a viable solution was to stay very focused on that specific domain. Um, and so chances are we are more going from sales to customer success rather than from sales to even call centers or the general public.

Speaker 1: Go ahead. [inaudible]

Dominik Facher: It's a question and it's a, it's a really important one, um, design your privacy agreement to, to be able to do that. Yes, that's, that's really important. Um, we, um, obviously anatomize the data and so we've built models that go across customers. Um, and that's, that's the only way you can actually train it, but it's, it's fully anonymized and, um, compliant. But yeah, it's our models. Yeah. Yep. Um, so the customers own their own data, they can delete it and we take it out of the models if they deleted. But, um, the anonymized training data, um, stays with us. [inaudible]

Dominik Facher: The, yeah. So every piece of customer data gets taken out, but the anonymized parts that we use to train the model, they stay with us. Yeah. Go ahead and say again. What do you mean by that? Say More

Speaker 1: [inaudible] right?

Dominik Facher: Yeah, it's a, it's a great question. So if you think about what anonymized means, it's the personally identifiable data that you take out of it. And so in this case, we only take parts of a transcript, for example, that helps you build an algorithm for identifying next steps. And so what we do is we remove everything like the speaker, like the video, et Cetera, in order to train that particular algorithm. And so that one is you, there's not even an audio part there anymore. And so we take it everything out. That would be at that level.

Speaker 1: [inaudible]

Dominik Facher: That's all in our service. Yes.

Speaker 1: Yeah. [inaudible] a little problem. [inaudible] highlight action [inaudible] but I'm not sure how [inaudible]

Dominik Facher: Yeah. Um, the first thing, the biggest part is actually spend a lot of time with them. Um, that's, that's the ultimate thing. Um, and it's all about domain knowledge. So we over-invest in getting the entire data science and research team very close to customers and like really being experts in sales. Um, they actually listened to more sales conversations than probably anybody else in the, in the company. They really understand sales by now. And so, um, that's ultimately what it comes down to. And then everything that we're sort of thinking about what's coming next. Um, we worked very much hypothesis and prototype driven. Um, and so there's just a lot of back and forth around, but it's about like domain knowledge and very, very close frequent interaction. Yup. Please go ahead.

Speaker 1: Do you guys think about filming [inaudible]

Dominik Facher: Yeah.

Speaker 1: [inaudible]

Dominik Facher: Exactly. Yeah. So the question was build versus buy, especially in that, um, we are very strong built company, um, as opposed to buy and we evaluate that on every single piece of technology that we build or buy in that case. Um, and we want to own the parts that we think actually make a meaningful difference to the outcome. Um, speaker separation is just not available. There is no API or APIs properly on the market. And so by now, I think there's the first one's coming from Microsoft as far as I know. Um, but when we started this, there was actually nothing. There was a couple of research papers, but that's it. No commercial solutions. Um, sentiment analysis is a problem that's much more solved today. There's not much you have to actually have to be invent. Um, and so that's something that we'll probably more take, um, from the market and available and others. Yeah. Where you have to control the outcome and the quality, you sometimes need to own them and it takes time to do that. All right. Any other questions? Well then, thank you very much.

Speaker 1: [inaudible]

Keep me posted on Empower 2019.