AI at ZoomInfo: Superpowering GTM teams | Ali Dasdan, CTO, ZoomInfo

Thank you for having me.

Yeah, so we help companies to sell better, if I just summarize it that way. So the personas that we work with are salespeople and marketing people, and every company, in principle, is our customer, potential customer. And we want to make sure that the sellers can not only find new customers very easily, but also retain their existing customers. So we provide them the data, insights, and signals to achieve that.

Yeah. So you're going to come to, let's say, our product. We also notify you. You don't have to come, but let's say you're looking at our user interface, and you can do searches for anything you care about. Companies, let's say you're selling chairs: which companies are in the market right now for buying chairs, right? Or if these are existing customers, definitely you can figure out whether they are researching, they are doing similar searches. Once you find out that this company is a potential customer for you, then we provide information about this company.

If it's a startup, let's say whether they got funding recently or not, you can get information about the organizational hierarchy. So who are the people who are going to be decision-makers for that? All the information about them, how do you reach them? We have the mechanisms for you to reach them using, again, AI. We can actually create even the email to send, which is personalized for that use case. So fundamentally, basically not only finding the right customers and right people, but reaching out to them and making it very easy for you actually to make the sale.

Expanded, yeah. So since we have a very strong data foundation, we have lots of information about companies and business people in these companies. As a result, you can literally use that information for multiple purposes. Also, tying back to, since I'm going to help you sell better, marketing is part of that flow, part of that funnel. So that's one of the aspects of it. And also, the fact that you have a basis like that, let's say you want to find the best people from a business context, and we have a TalentOS product that you can actually use for that.

And OperationsOS product is to help you clean, for example, your data, to remove the duplicates, to make sure that it is ready for you to go.

Yeah, data as a service, we call it DaaS, is one of those businesses. So some companies, even though they are customers of ours, they have licenses, seats that their sales teams are using our product. But sometimes, let's say they want to get a copy of that data and enrich within their own platform. Let's say they have their own first-party data. Our platform enables you to bring first-party data. We have lots of third-party data. We can actually combine this on your behalf very easily.

But at the same time, some companies prefer that. And as a result of that, you can just get a copy of this data through multiple platforms. You can go to AWS, Databricks, Snowflake. It's ready to go for you. It's in their marketplaces.

You can get it through API, you can get it as a file, or you can get it ready to go in the marketplaces for these companies, Databricks and Snowflake and GCP and AWS, and it's ready to merge for you. Basically, you can just do that processing over there. So either way, basically, my point is, our point is not to create any friction. The point is to help you make the sale, to make your sellers better salespeople. And as a result, if you need the data, we also have a capability like that.

Yeah, I mean, originally came to the U.S. for a PhD, worked in lots of small startups as well as big companies.

My PhD was mostly in graph theory and algorithms. It was for doing timing analysis for embedded real-time systems. So it had applications to chip design software, electronic design automation. That was the reason my first company was Synopsys, which is one well-known company in this space. And after that, Yahoo Web Search, did lots of big data and data science and advertising and search over there. eBay for e-commerce. And I did real-time advertising in a startup. I worked in the UK for two years in retail, but built their data platform and marketing platform.

Then Atlassian in the productivity space. I was leading one of the three business units over there, which was Confluence, Trello, Jira Work Management, and some of these products were part of my team. And then after four years over there, I moved to ZoomInfo.

It's roughly over 900 people, close to 1,000 people. So the way we are organized, maybe we're going to come back to it. Like you're saying, definitely, we have a strong platform team, and data platform is part of that. Data platform covers data engineering, data science, and the data platform by itself. There's a separate enterprise engineering team. We call it the Enterprise Productivity Team because we care about productivity a lot. And that deals with serving the internal customers: revenue team, finance team, HR team, and all that.

Definitely operations team, site reliability, security teams, IT teams. And on top of the platform, we have multiple products. These product teams you mentioned: sales product, marketing product, operations, talent. We have Chorus for conversational intelligence and a couple more.

So the way it is structured is, definitely, we have a cloud infrastructure team underneath. We are using both Google Cloud as well as Amazon Web Services. And on top of it, we have actually two data platforms. One is for all the data that customers need and our products need. And there's another one that's internally focused for our internal customers. And they're built on different technologies. There's a pipe between them just to make sure that data is exchanged in real time.

Then we have teams that are developing applications on top of these data platforms, including data science applications, creating machine learning models using GenAI, using that data. And the fact that we have the data in one place makes it super easy to do all of that. And also, data is huge in our case. We are the market leader in that with respect to company data, intent data, contact data, first-party data, third-party data. Then we have on top of it what we call product foundations teams.

For example, search team, identity team, admin hub team that deal with settings, permissions, and all that, platform UI team, for example. And all the teams are what we call product foundations. And then on top of it, now we have the product teams that are interacting with those. And the idea is basically to use these platform components and use the data that is already there so we can build products and features very fast. So through the APIs, third parties can also do that.

But more and more, we are trying to also make it easy for third parties to build new experiences, new user journeys on top of our platform.

We call it ZDP, ZoomInfo Data Platform.

Okay. Yeah, internal name is ZDP.

I mean, with respect to, let's say, you're going to get first-party data. If first-party data is, let's say, in Salesforce or HubSpot and platforms like that, then we already have these pipes set up for you. So that's one way for us to get data. With respect to third-party data, we are definitely continuously crawling company websites, merger and acquisition data, financial data, SEC data, all kinds of information that are about people, including news that's coming out of those. That's also continuously happening.

So all the public sources, basically, we will have it. We have a contributor network. For example, if you're using ZoomInfo Lite, then as part of that free access, we get some data from our customers.

Information about your company, your email or phone number, things of that nature. If you're also tied into, let's say, your website traffic, then we can also use that to identify why people are visiting your site.

Yeah. Also, these are not personal; it's all business-related. So our context is these are all needed to make a sale, for you to find customers, to retain your customers, as well as your potential business interactions that you're going to have, just to help for that. So data is completely specified for that.

Publicly available, but customers definitely have first-party data, right? So some companies, let's say they're using Copilot. As a result of that, they want to actually combine their first-party data with our data. Let's say they have a set of contacts in these companies, and the profiles for each of these contacts that they have may not be complete. So they want to use, let's say, our data and tech to enrich that, to either correct these, to remove duplicates and all that, but on top of it, maybe to get additional attributes that they can use for that purpose.

So for that, we can get that data on their behalf, completely dedicated to them as a single tenant. We combine that so we can extract additional signals for them for selling.

Depends on the type of the data, definitely. If you're, let's say, getting public data from public sources, first of all, you got to crawl it, right? You got to crawl it at a certain frequency so that you are going to get all the changes that are relevant for our purposes. You have to put the right tech over there so the data that you are extracting from those sites will be correct. We also have a good number of researchers, about 300 of them.

If there are specific cases that we need a human touch for them to actually correct the data, it's very difficult to find, definitely in certain cases, to verify that this is accurate data. Then also people might be involved in the loop as part of that.

Almost for every attribute that we are getting, we have a machine learning algorithm or model to tell us how to interpret it, whether it's correct or not, what kind of anomalies we should be sending over there.

Yeah, we use some crawlers we might be using from third parties, but many of these components are homegrown, yes.

Yeah, ZDP is built on Google BigQuery right now. So all these pipes are continuously pushing the data to the platform, or sometimes we are pulling the data. Then you have to definitely store it on a platform like that, and then you got to do processing on top of it, right? So you have to sometimes join this data or extract certain attributes for each of these, create a graph because these are related to each other. A person works in a company; this company maybe sells a product to another company.

All these connections are there that we are extracting out of that, and that is done through different technologies, either Spark code or Dataflow that GCP has, all kinds of basic processing. Machine learning, definitely, right? Model building. Some of these are just traditional machine learning. Let's say you're going to have a tree-based model that uses that data to create, I don't know, classification or prediction. Also GenAI, right? Using that data, either create an additional context for GenAI, also just to make sure that you can prevent almost hallucination, saying that, do not go beyond that.

Here is your data and try to extract insights out of that. So all that processing is done over there.

So we're using some third parties on that. And part of the catalog, we also do internally, so it's homegrown, to make sure that we can build it in the way that we can. It's not a single system, right? We have a cataloging solution that we are using from a third party, and there's a version of that internally. And there are different cataloging systems, component cataloging, for example, pipeline cataloging. And these are also connected to each other so that they're almost like a unified source for whatever you're looking for.

It's accessed through that. And with respect to processing on top of it, catalog is telling you exactly where things are, a single source on that one, all the metadata. Then how are we going to actually create these schemas in such a way that they themselves are in a way that you can process them very easily, right? So creating all the readers and writers automatically, writing the processing in such a way that whoever is going to write a processing engine internally, I don't want them to literally go that deep into every schema, every data type, and it should be very unified in a way that we may store the data in, let's say, one or more tables, but they should see it as a unified, let's say, 360 view as a person or as a company, right?

So that it's a logical representation, and it gets mapped to a physical representation underneath. And through the cataloging system, actually, you have access to that one, right, with respect to attributes and how the schema looks, owners, who are the owners, right, what kind of updates that we have on top of it. That logical layer, actually, as well as an API on top of it, simplifies basically application writing and creates a very fast turnaround.

Those applications are using additional platform components. For example, search. If you go to SalesOS, there's lots of filters: companies of that size, that revenue, people of that title, and intent of those categories. So you do that search, which means the data that we just extracted should be pushed to search indexes. So we are using Solr internally. There's also some Elastic used for different purposes. If you're going to join the data, then they go through what we call entity resolution.

So there's a separate pipeline for that one because you cannot just perform a regular join. As you know, that join is pretty involved, and lots of machine learning and rule engines are used for that. So that's one place. For example, let's say you need access to—you find what you're looking for through search, but you need more data about that key that you just found out—then we have a separate system to give you additional information about that entity, whatever that is, let's say company or contact, right?

So all the right places it has to go. Then through the APIs, our products are accessing those. So it's also unified in that sense. So there's no need for creating a separate search for talent versus sales, because it's a big investment. It's very real-time indexing, real-time search, very fast, like milliseconds. It's also expensive to run something of this scale. It's already elastic. So we try to basically push it into the right platform foundation components so products can access it very easily.

And even APIs, external APIs, are supported through those systems.

I think across—so, creating a data platform with respect to its components, right? All the integrations on the input side, output side, storage, processing, governance layer, with respect to cataloging, security, permissions, definitely pushing the data for business intelligence, Tableau, and systems like that. So most of these are well known, but it just gets learned and relearned every time, right? So I'm lucky that I have created huge platforms multiple times in the past.

It helps.

But choosing those components, making sure that data flows, let's say, in real-time streaming mode, you can be assured of the quality of that data if you're going to create, let's say, analytics out of those BI systems or custom analytics interfaces, right? All of these should be easy. I think the lesson is this, again, that gets learned every time, is you have to pay attention to all these pipes and the boxes and how these things are connected to each other so that you can do it in a way that you can trust that data, you can turn it around.

And time to market is very important over here. You don't have to start everything from scratch. Let's say you're going to do analytics. There are lots of analytics databases, different choices. They are proven with respect to these capabilities. I mean, in the past when we were doing data platforms, there was no Kafka, for example. We created a Kafka-like system ourselves, but we are lucky that we created an HBase—sorry, Hive-like system ourselves with 100 petabytes of data, right?

So there's no need for that right now. I think many of these are very mature. You just have to know which one to start with, or just start with one. And you can adapt to that. I think the lesson is time is of the essence. Productivity is key. Just use all these very mature components and connect them in a standard way. There are lots of best practices on that.

Yeah, definitely. For example, our internal productivity dashboards are driven through DevLake. It's an open-source Apache project.

Almost every key AI technology or model-building technology, we had to have it. Because when you're dealing with data of this nature, some of this data is, for example, unstructured, right? How do you extract that? Or if you're dealing with calls, you have to do natural language processing, which simply, you have to figure out what's being said, who's saying it at what point, analytics, sentiment extraction, right? All of these you cannot do without AI. So the company was deep into AI before GenAI.

We had a team almost from the beginning, but all these models were deployed across the board, and many of them are traditional models, right? Natural language processing models, tree-based models, classification models, and SVM and things like that. Even with Chorus, right, Chorus still runs our own homegrown NLP technology. We are doing it because our quality, we believe every time we compare, is actually pretty good. We are beating well-known vendors. Plus, it's cheap for us, right?

We are running it with just a couple of people. It's really well done. Now, when GenAI came, especially in our business where these AI models were working pretty well, it created a revolution almost. And the first thing we released actually last year was applying GenAI to Chorus. So you go to a call, let's say you have a customer call for one or two hours, you are talking about things, and you want to share this with your team. Well, what's the best way to share?

They can look at the transcript, they can watch the video, but then they're going to spend two hours. And one of the best applications and first applications of LLMs were summarization, right? So we actually provided call summarization and action item extraction. And the amazing thing was, literally, as soon as it got released, as you know, increasing NPS is difficult. Sometimes every time you just increase just a couple of points, and the absolute increase in NPS was just amazing, almost like 20 points.

Because people loved what they were getting with that. So that was the first experience.

Yeah, for Chorus, it was an addition, not a rip and replace. We are still using natural language processing with respect to converting recorded conversation to transcripts. It was extracting summaries and action items from transcripts. We are using GenAI, so it was an addition.

With respect to extracting from recorded calls, we have not actually tried it in that sense. There were some experiments on that. I think it's going to be very expensive to run this unless we have our own model. Now we are running on GPUs that we have created. Probably this could be one of the areas that we can do. But since our NLP tech for creating transcripts is working really well compared to what's out there, it's not the first priority to replace that at this point.

But anyhow, GenAI, that was the first use case and wow moment. Then with Copilot, as you know, right?

Chorus. For that one, we used Anthropic.

So yes, we are using Anthropic, we are using OpenAI, we're using Google. But for that one, we released with Anthropic.

Yeah. Then Copilot came. Copilot is GenAI, even beyond GenAI. The whole point is to completely turn around the interaction of sellers with our products.

But the name fits really well because it's really a copilot and helper for sales.

So the problem that sellers will have, right, and they want to retain their customers, they want to find new customers, they have to be on top of so many signals. And if you look at what best sellers are doing, they reach out to different data sources, they extract insights from different places. They do, let's say, lots of searches on ZoomInfo and combine the data that we are providing with what they were doing. But then they had to do all of these. And if you are not as good of a seller as these best sellers, then you were at a disadvantage, right?

So Copilot completely turned this around, right? So, at the fundamental technology level, it's a recommendation engine, right? But what it did is, it will tell you exactly what you should be focusing on: which accounts, what are the signals that you should be paying attention to, what action you should be taking on top of it. We are actually helping you to take those actions. So if you're going to send an email, we will actually help you with that email. If you go to what we call Account AI, all the information across all the data sources, all Chorus calls, right, all that extracted data boiled down to a version that you can actually pay attention to.

And the chat box is also connected to it. So you can literally ask questions about that account. Let's say you want to get up to speed, you're going to have a conversation with a potential customer, and you have half an hour. There is no way you can look through all that data to extract what is the essence of that, right? And we help you to do that, right? So it just turned this completely around, and people just love it because of how it enables you as a seller to be super effective.

It will definitely magnify the skills that you have. I mean, you still have to take extra action with respect to what questions, for example, you're going to ask.

Closes most of the gap. Yeah, definitely. Actually, some sellers were jokingly complaining about it, saying that, hey, I am a good seller. I'm doing all the work, and now you're going to help these lazy sellers to compete with me.

No, it was released in production. So in May. Before May, actually, it was in beta.

Twenty-seven or so thousand customers. I mean, people were saying that they save about, like, 10 hours, literally one day a week they were saving by using Copilot. And with respect to the use of GenAI, it was not only, like, Account AI or summarization, email generation, but actually the models underneath for certain signals were done using GenAI. It actually simplified our life because the typical way of extracting the features and creating a model and the training and all that. So just in some of the cases where we had more than a dozen models, that completely were regenerated using GenAI for Copilot.

So one is the data, definitely, right? Without data, and that's actually the power. So you see lots of mentions of AI: I just throw AI and I am getting good results. But AI without the data that we have, not only its comprehensiveness, accuracy, how you're actually making sure that the model, you are providing the right context to the model. Not only do you get better results, but you prevent hallucination for that context. So all that actually is a requisite for the model to run really well.

Yeah, let's say Account AI use case. So how am I going to extract that account information? Right. So this is my account. Let's say I'm the seller. What are the pain points of this customer? How many conversations have happened so far? And in each conversation, what was the summary of the conversation? Was there a need mentioned or a competitor mentioned, or was there a problem that was mentioned that I can pay attention to? I know exactly what kind of discussions happened about their problem needs as well as pain points and what kind of responses we have given, right?

All that has to be extracted. And this is going to be extracted with not only the information that we have about this account, but also the first-party data. Let's say you're going to bring, because let's say you were using Chorus and all that, you own that data. Yes, our technology is recording it on your behalf, but you own that data, and we extracted this on your behalf automatically. So it's in one place. So this is just one use case of Copilot.

The others are actual signals themselves, right? So it's going to tell you, let's say you work with John Smith in this company and the person has moved to another company, right? This new company, not only is the news about that, news by itself is just one signal, but more valuable is this new company maybe is in the same industry or they are in the market actually trying to buy something that you are selling, right? All that has to be extracted and presented to you.

You don't have to even search for that anymore.

Is that correct? So for each of these use cases, you have a model we created. We're using Anthropic and OpenAI, right? We have our own. We use Llama, basically anything that is—

So the one that we deployed right now is 7B, I think. Okay, the 7B. Yes.

There is always a trade-off, definitely. Right, accuracy and cost trade-off. So let's say it is a small-scale run. Let's say I'm going to extract, in that context, a simple use case. Go with the email generation, right? We also provide through the UI, like, you can actually customize it even further for the person. It's not just some random email, actually an email based on all that information that we have extracted. It's to the context, and then you can customize it even further.

So for that use case, it's going to be, let's say, you deal with 10 accounts or 100 accounts, it's 100 emails. Let's say using Anthropic or OpenAI for that one, it's cost-effective, right? Plus, accuracy is, since they are on par with the top models, and then you can just use them for that purpose. But let's say another use case is I'm going to extract the scoop or news item from these websites, and I have 100 million websites. So using 100 million LLM calls, Anthropic, they can do it, but it's going to be expensive.

So for that one, we can actually use our own, right? And plus, we are actually renting the GPUs in Oracle Cloud. Okay, so your Llama runs on Oracle GPUs.

Exactly. Then we can run it at any scale because we already paid money for it, right? And for those use cases, even if it's comparable for most of the experiments that we have done, but let's say even if it's just a little bit below in terms of accuracy, at that scale, the advantage that you are getting is still a big magnifier to whatever else you can do right now. If you do not have that technology like that, what are you going to do?

You have to crawl the data, extract all those insights, create all those models. Maybe you have it. If you don't have it, you have to create them from scratch, do feature extraction. All that time saving is given to you by using that.

I think you have to create a layer on top of this at a minimum, right? In the end, whether Anthropic or OpenAI, they're companies, right? They sometimes have stability issues, right? If they're down, what are they going to do about that? Since our responses, Copilot, is all real-time, we have to respond to that. There is security: whatever we are sending and whatever we are getting as a result of that has to be scanned, right? So that layer actually includes that.

There are companies today offering even further on that one with respect to accuracy and cost trade-offs. So right now we have not implemented that one, but that's one thing that we're actually considering doing within that layer. There's also, as you know, either GCP or AWS providing a layer by themselves for you to make it easy to choose different models for different use cases, right? So we are thinking, like, how much of that layer we should be creating ourselves and how much we can actually rely on what Google or AWS or Oracle will be providing as a layer for us.

And we can just use that. And given, I think, how valuable these products are, how well they work, and the enterprise use cases will force enterprise support structure around that security and privacy and model selection, all that. So as a result, probably we've got to be careful with respect to how much investment we are going to do versus what's going to be given to us a little readily through these cloud vendors.

Yeah, I mean, we are very technical ourselves, so we can evaluate them, I think, pretty well. We work with so many vendors, small or large. The point is how it is helping me. Am I definitely saving time? Productivity is top of mind for us, as well as, if I were to do this ourselves, right, we can do anything. In the end, it's software, but how many people are you going to allocate, how much time is it going to take, and all that?

If there is a solution out there that we can use, we will do that. But one key is it has to be very easy to integrate and it should work, right? And we are going to make sure that through evaluation, that's the case.

So one is, I think we have an internal motto that we always use: how fast do you move is probably the only advantage of any company, including ZoomInfo. So as a result, I will actually give a talk today about productivity on that one and share some of these. And I think a couple of months back, at FirstMark CTO Summit, to plug this, and the CTO Guild, again from FirstMark. Thanks for that. Thank you for the invitation. So I had actually given a talk on using Copilot, GitHub Copilot, for developer experience.

So again, since productivity is top of mind, we try to use these tools to speed things up, right? So the first easy use case was GitHub Copilot or OpenAI or Anthropic, literally to generate code, generate testing, help with documentation. And people love it. We do regular quarterly developer satisfaction surveys. People love what they're getting out of that. There was a big push, actually, for us to deploy these as soon as possible. We made sure that it is secure, went through all the basic enterprise checklist items before actually deploying this.

Now we are trying to go beyond that. It's not only helping while you are coding and writing tests. We are trying to, for example, speed up code reviews using this technology. And Google had an internal, I think, solution that they had a paper about. I'm not aware of any startups doing that. Probably there are, but we are trying to develop ourselves internally to speed this up because code reviews are a best practice for improving product quality, software quality. But at the same time, it takes time.

It takes time from both sides, from the coder who wrote the code as well as who's reviewing the code. So if you can speed this up, even the nice thing about these things is you don't have to completely eliminate humans. Even if you save, let's say, 10 minutes, when you multiply it by the number of code reviews you are doing and by the number of developers that are doing it, the savings will be immense, right? Trying to do the same thing for incident detection or creating, for example, telling you exactly maybe what could be wrong.

With that incident. And response time for an incident is not only to discover, but to resolve, mitigate the issue, is super important. Again, GenAI can help a lot over there. So we are getting initial signals around that, some good use, but we have not deployed a solution yet. But it's being developed.

Actually, they love it. Yeah. I mean, not everybody started using it right away, but there was a push to get this deployed as soon as possible. So that was a nice thing to hear from the developers, that they also want to move as fast as possible. So we actually deployed this, I think, last year. We first did a trial with a number of developers just to make sure it works out. Good signals were there, but just to prove that it actually works out for the use cases, again, it's secure and we are going to protect our intellectual property.

All of that, we checked, confirmed, then we deployed to all of our engineering. So everybody's using this daily in their interactions. And our acceptance rate for code suggestions was around 25%, actually over 25%, which is interestingly in line with the recent research that GitHub published. I think they were mentioning 26%, which was good that we were actually aligning with the number that they mentioned.

Really appreciate it.

Yeah, this was fun. Thank you for inviting me. Thanks for having me.

Yeah.