Sep 19, 2024 · 45m · mad
AI at ZoomInfo: Superpowering GTM teams | Ali Dasdan, CTO, ZoomInfo
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The MAD Podcast, ZoomInfo CTO Ali Dasdan joins host Matt Turck to discuss ZoomInfo's modern data architecture, multi-cloud ingestion pipelines, and the deployment of Generative AI in ZoomInfo Copilot. He also shares insights on model selection strategy, enterprise RAG frameworks, and driving internal developer productivity.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 27.8% of the talking time here. How this is scored →
Matt as informed peer 3.3 Guest teaching 2.7 Guest disagreement 0.1 Matt pushing back 0.0
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 26:20 Ali reframing the 'rip and replace' assumption
Ali gently pushes back when Matt asks if Generative AI replaced their legacy Chorus NLP stack, clarifying that original NLP was kept for cost-effective speech-to-text while LLMs were added purely for summarization.
Hardest push from Matt ▶ 12:30 Matt pressing on ZoomInfo Light user data contributionsMatt probes directly into the exact data commitments required of ZoomInfo Light users, asking what data customers give up in exchange for reduced platform pricing.
Biggest teaching moment ▶ 35:45 Ali explaining API cost trade-offs vs self-hosted Llama 70BAli educates Matt on model economics, explaining that while commercial LLM APIs work for low-volume tasks like email generation, crawling 100M websites requires self-hosting Llama 70B on rented Oracle Cloud GPUs.
Matt holds his own ▶ 30:10 Matt's analysis of AI leveling up average sales repsMatt demonstrates high strategic insight into sales dynamics, articulating how AI copilots compress the skill gap between average and top-tier reps by automating preparation and account analysis.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Company Overview and Core Value Proposition for Sales | 4 | 1 | 0 | 0 | Matt opens with precise public financial metrics ($1.2B revenue, 28% operating margin) and offers a clear reframe of ZoomInfo's value proposition for B2B sellers. Ali agrees and expands on the core search and contact workflows. | |
| Platform Expansion Across Personas and Operations OS | 3 | 2 | 0 | 0 | Matt specifically brings up the Data as a Service (DAS) business line to explore beyond the core app. Ali explains how customers enrich first-party data via cloud marketplaces like Snowflake and Databricks. | |
| Ali Dasdan's Personal Background and Career Journey | 3 | 1 | 0 | 0 | Matt demonstrates background knowledge by asking about Ali's PhD topic and identifying Confluence as his former business unit at Atlassian. Ali outlines his engineering org structure. | |
| Core Data Architecture and Platform Foundations | 2 | 3 | 0 | 0 | Matt sets up a discussion on data architecture at a data company. Ali details the core separation between internal and customer-facing data platforms. | |
| ZoomInfo Data Platform Ingestion, Pipeline Architecture, and Quality Assurance | 4 | 3 | 0 | 0 | Matt asks detailed questions about the ZoomInfo Light quid-pro-quo model and ML data quality checks. Ali walks through web crawlers, 300 human verifiers, and BigQuery processing pipelines. | |
| Modern Data Stack, Data Cataloging, and Logical Schema Abstraction | 3 | 4 | 0 | 0 | Matt inquires about modern data stack tools like cataloging. Ali explains the unified logical layer that abstracts multi-table physical schemas into a unified 360 view. | |
| Architecture Lessons, Engineering Velocity, and Open-Source Adoption | 2 | 3 | 0 | 0 | Matt asks about architectural failures and lessons learned. Ali reflects on building custom systems pre-Kafka versus leveraging mature open-source solutions today. | |
| Evolution of AI at ZoomInfo, Chorus, and Generative AI Integration | 4 | 3 | 1 | 0 | Matt asks if generative AI was a 'rip and replace' for Chorus's legacy NLP stack. Ali gently corrects him, explaining that custom NLP remains for transcription while GenAI was added for summarization. | |
| ZoomInfo Copilot: Proactive AI Recommendations for Sales Teams | 2 | 2 | 0 | 0 | Matt asks about the launch of ZoomInfo Copilot. Ali explains how Copilot shifts the platform from manual search queries to proactive recommendation engines. | |
| Impact of AI on Sales Productivity and Industry Skill Dynamics | 4 | 2 | 0 | 0 | Matt offers an insightful observation on how AI closes the gap between mediocre and superstar sales reps. Ali enthusiastically agrees and shares user anecdotes and productivity metrics. | |
| Technical Deep-Dive: RAG Architecture, Grounding, and Account AI | 4 | 3 | 0 | 0 | Matt correctly identifies the Account AI architecture as a retrieval-augmented generation (RAG) system. Ali details how grounding LLMs in proprietary context prevents hallucinations. | |
| Model Selection Strategy: Balancing Commercial LLMs and Self-Hosted Open Source | 4 | 4 | 0 | 0 | Matt asks about model selection dynamics between open source and commercial APIs. Ali explains using Anthropic for low-volume precision tasks and self-hosted Llama 70B on Oracle Cloud GPUs for massive web extraction. | |
| LLM Tooling, Middleware Abstraction, and Enterprise Governance | 4 | 3 | 0 | 0 | Matt asks about middleware tools (LangChain, evaluation platforms) and startup vs cloud vendor dynamics. Ali outlines internal middleware needs like uptime failover and security scanning. | |
| Internal Productivity, Developer Tools, and AI Incident Detection | 3 | 3 | 0 | 0 | Matt mentions Ali's talk at FirstMark CTO Summit regarding developer productivity. Ali shares experiments using GenAI for automated code reviews and incident detection. | |
| Developer Adoption, Security Vetting, and Acceptance Rates | 3 | 3 | 0 | 0 | Matt asks about developer adoption dynamics for GitHub Copilot. Ali explains the bottom-up developer pull and cites an internal acceptance rate of over 25%. |
Statements from this episode (14)
ZoomInfo reported $1.2B annualized revenue and 28% operating margin in Q2 2024
“1.2 billion dollars in annual annualized revenue as of Q to 20 24. I should note that unlike some other tech companies you are very profitable with a 28% operating margin, more than 35,000 customers worldwide, 3500 employees worldwide.”
ZoomInfo's technical organization numbers roughly 900 to 1,000 people
“It's roughly, you know, we're at 900 people, close to a thousand people”
ZoomInfo operates two distinct core data platforms on different technologies
“On top of it we have actually two data platforms. One is for all the data that customers, you know, need and our products need, and there's another one that's for internally focused for our internal customers, and they're built on different technologies.”
ZoomInfo employs around 300 human researchers for manual data verification
“We also have a good number of researchers about, 300 of them.”
Dasdan's past teams built custom Kafka alternatives to handle 100 petabytes
“In the past, when we were doing data platform, there was no Kafka, for example, we created a Kafka-like system ourselves, but we are lucky that, yeah, we created an HBase, sort of HHive-like system ourselves with hundred petabytes of size of data, right?”
ZoomInfo claims its homegrown NLP outperforms major external vendors on quality
“Chorus still runs its own, our own homegrown NLP technology. We are doing it because our quality, we believe every time we compare, it's actually pretty good. Actually, we are beating, you know, we've been well-known well-known vendors. Plus it's cheap for us,…”
Adding generative AI call summarization boosted ZoomInfo Chorus's NPS by 20 points
“And the absolute increase in MPS was just amazing, like almost like 20 points. Because people loved what they were getting with that.”
ZoomInfo selected Anthropic over OpenAI and Google for call summarization
“We are using Anthropic, we are using OpenAI, we are using Google, but for that one, we released Anthropic.”
Dasdan: ZoomInfo Copilot users report saving about 10 hours a week
“No, it was released in production, so in May before May, actually, it was in beta. 27 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.”
ZoomInfo replaced over a dozen traditional machine learning models using generative AI
“Actually, it simplified our life because the typical way of, you know, extracting the features and creating a model and the training and all that, so just it, in some of the cases where we had more than a dozen models that completely were regenerated using Gen…”
ZoomInfo rents GPUs in Oracle Cloud to run open-source models
“We are actually renting the GPUs in Oracle Cloud, actually.”
Real-time AI apps require abstraction layers to handle LLM provider outages
“Yeah, I think you have to create a layer on top of these at a minimum, right in the end, whether Entropic or OpenAI, you know, they're companies, right, they sometimes have stability issues, right, if they're down, so what they're gonna do about that? Since ou…”