Nov 17, 2022 · 39m · mad
Leveraging People Data at Scale | ADP Chief Data Officer, Jack Berkowitz
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
Jack Berkowitz, Chief Data Officer at ADP, joins Matt Turck at Data Driven NYC to discuss processing workforce and financial data at global enterprise scale. The fireside chat covers ADP's cloud data architecture, predictive machine learning applications, AI ethics governance, and strategic advice for early-stage startups partnering with enterprise leaders.
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 12.1% of the talking time here. How this is scored →
Matt as informed peer 2.6 Guest teaching 3.7 Guest disagreement 0.3 Matt pushing back 0.0
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 0:18 Vendor Neutrality Correction
Jack explicitly interrupts the host's premise that ADP is strictly a 'Snowflake shop', setting the record straight that they avoid endorsing any single vendor.
Hardest push from Matt ▶ 0:05 Challenging Black-Box Model AssumptionsMatt questions whether ADP uses black-box deep learning models where predictive factors cannot be explained, forcing Jack to defend their use of explainable ML.
Biggest teaching moment ▶ 0:22 Enterprise Legacy Realities vs Cloud StartupsJack educates the audience on why cloud-native startups miss crucial enterprise needs like connecting to legacy DB2 databases on-premise, justifying their choice of Informatica.
Matt holds his own ▶ 0:13 Connecting Architecture to Data MeshMatt demonstrates deep knowledge of current data engineering trends by directly challenging Jack on how ADP's hub-and-spoke setup compares to Data Mesh.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Title Cards and Speaker Introductions | 3 | 3 | 0 | 0 | Matt opens by quoting specific ADP operational metrics such as client count and revenue to ground the interview in scale. Jack elaborates on the company's massive transaction volume and data processing breadth. | |
| Predictive Models, LLMs, and Automated Ontologies | 4 | 4 | 1 | 0 | Matt asks technical questions regarding explainability in deep learning models and automated ontologies. Jack clarifies why explainable models are preferred in HR and describes ADP's automated skills graph. | |
| Defining the Role of a Chief Data Officer | 2 | 3 | 0 | 0 | Matt frames the evolving definition of the Chief Data Officer role across tech companies. Jack describes his background in product development and how ADP created its first CDO post to balance data security and product innovation. | |
| Organizational Structure: Hub and Spoke Data Model | 2 | 3 | 0 | 0 | Matt prompts Jack on organizational data structure, leading Jack to outline ADP's hub-and-spoke architecture where a core team maintains the platform while 30 distinct spokes contribute semantically declared data. | |
| Data Mesh, Governance, and Access Control | 5 | 4 | 0 | 0 | Matt demonstrates strong domain knowledge by raising the Data Mesh paradigm to see how ADP aligns with it. Jack details which Data Mesh principles ADP adopted, such as declared semantics, and which they rejected, like federated query. | |
| Tech Stack: AWS, Databricks, Snowflake, and Custom BI | 4 | 4 | 1 | 0 | Matt probes ADP's cloud tech stack and asks about BI tools, prompting Jack to politely clarify that ADP avoids vendor exclusivity and builds custom reporting engines to handle massive enterprise scale. | |
| MLOps, Supply Chain Security, and Vendor Infrastructure | 4 | 5 | 1 | 0 | Matt asks detailed questions about MLOps and vendor choices. Jack educates on ML supply chain security risks and explains why legacy on-prem mainframes made Informatica necessary over cloud-only startups. | |
| How Startups Can Successfully Partner with Large Enterprises | 4 | 4 | 0 | 0 | Matt draws on his venture experience to ask how startups can successfully partner with large enterprises. Jack explains that enterprise buyers prioritize executive trust and long-term relationships over marginal technical features. | |
| Enterprise Innovation & ADP's AI Ethics Board | 3 | 3 | 0 | 0 | Matt references ADP's AI Ethics Council from his prep research, allowing Jack to share how ADP structures monthly review meetings with internal teams and external advisors under NDA. | |
| Audience Q&A: Managing Algorithmic Bias and Diversity | 0 | 4 | 0 | 0 | With Matt stepping back to moderate audience Q&A, Jack advises sales reps on pitching enterprise CDOs and explains how ADP measures algorithmic bias using EEOC standards. | |
| Audience Q&A: Candidate Evaluation, ATS, and Skills vs. Pedigree | 0 | 3 | 0 | 0 | Matt acts purely as a passive moderator while an audience member asks about recruitment algorithms. Jack explains that modern candidate evaluation systems weigh verified skills over traditional academic pedigree. | |
| Audience Q&A: Big Data Evolution & Client Data Privacy Rights | 0 | 4 | 0 | 0 | In the final audience segment, Jack contrasts corporate contractual data rights with individual employee privacy rights, highlighting ADP's 5-in-10 data aggregation threshold. |
Statements from this episode (24)
Berkowitz: ADP pays about 40 million workers globally every month
“Some of the numbers are kind of astounding, so we pay in any given month about forty million people around the world.”
Jack Berkowitz says ADP moved $3.2 trillion in money transactions last year
“So last year we moved about 3.2 trillion dollars.”
ADP's benchmarking database includes data from around 950,000 clients
“So we've got about 950,000 clients in this database.”
ADP uses ML to reduce 21 million job titles into 7,000 categories
“So in any given month, right, those forty million people, there's about twenty one million job titles. We run some pretty sophisticated, ah, machine learning to boil that down to about 6000, ah, or now 7000 key job categories.”
ADP's turnover prediction model runs on 57 attributes
“The turnover predictor, for example, uses 57 attributes.”
ADP operates large language models in full production
“We're in full production on top of some of those large, large language models.”
ADP's skills graph ontology is 95% automated
“Instead of having it all done by hand, it's, pfft, 95% automated.”
ADP pays roughly 20 percent of all workers in the United States
“We pay in the US, you know, something like 18 or 20% of people, and then we do the taxes on another You know, similar size.”
ADP rejected federated queries to maintain a centralized analytics repository
“The piece we didn't bring from the data mesh was the notion of federated query. Because the problem with federated query is the latency of the query is the worst performing member of the federated group. Right? And so we still have a centralized data repositor…”
Jack Berkowitz: Enterprise data governance requires runtime dynamic access over RBAC
“So we need a more flexible way of dealing with data that can actually be interpreted at runtime. Or during the latency of the query than sort of the straight hierarchies, role-based access control that's existed to date.”
ADP uses Databricks for its data layer and Snowflake for analytics workloads
“We use Databricks as our, as that data layer, and we use Snowflake to do very specialized analytical database workloads that it's very, very good at.”
Berkowitz: ADP pays 20% to 30% of the French population
“In France, we're also, we pay something like 20 or 30% of the French population.”
Berkowitz: ADP delivers reports to around 30,000 downstream systems
“We actually have something in our system like 30,000 downstream systems. You know, Fidelity, Blue Cross Blue Shield. They all get reports.”
ADP uses MLflow, SageMaker, and Robust Intelligence for MLOps and security
“We do use an awful lot of MLflow for our MLOps. We do SageMaker work. We're working with this really interesting company called Robust Intelligence in California that's helping us with bias monitoring, compliance, supply chain attacks on machine learning”
ADP chose Informatica over cloud startups to support legacy on-premise data systems
“We went with Informatica, and the reason why we're on Informatica is very clear. Not, as Spencer was saying, not everything is in the cloud. And unfortunately for startups, they're making decisions about where to build, and so they're building in the cloud. In…”
Berkowitz: ADP can scale a new product to 5,000 clients within a year
“So, I can launch a product today and have 5000 clients on it in a year.”
ADP includes external experts on its AI Ethics Board under NDA
“We have external people on that as well. We have industry experts. We invite external representation. People that aren't associated with the company, except for the ethics board participate. They're under NDA. They see exactly what we're doing in terms of prod…”
Startups only get a meeting if ADP already knows their work
“At least in my team, they don't get the meeting if we didn't actually already know what they were doing.”
ADP aims to measure and reveal AI bias rather than eliminate it
“And what we said was, is that we don't believe that bias doesn't exist. In fact, we believe it does exist. And so our role isn't to mitigate all bias, but our role is to at least measure it, understand it, and express it, and make sure that people understand i…”
ADP uses EEOC's Test of Means internally to gauge machine learning systems
“So the EEOC has a thing called the, ah, Test of Means, and we actually use that internally to gauge our machine learning systems, and then we'll test and monitor for that.”
Berkowitz: ADP handles recruiting for 15% to 18% of US workers
“If we're paying 18% of the country, we're recruiting about 15 to 18% of the country as well through those systems, okay?”
Berkowitz: Skills predict job performance better than academic pedigree or titles
“Education's an interesting factor, and it's in our system, but skills turn out to be a much more interesting factor as well as, not job titles, but the nature of the job.”