Earley AI Podcast - Episode 99 Data Governance, Business Context, and Why AI Makes the Old Problems Worse with Zoher Karu

Why the Same Data Problems That Existed Before AI Still Exist - They Just Get Expressed Faster, With More Confidence

Guest: Zoher Karu, Founder and President at ZiZi Advisors

Host: Seth Earley, CEO at Earley Information Science

Published on: September 14, 2026

Podcast Transcript: Data Governance, Business Context, and Why AI Makes the Old Problems Worse

Transcript introduction

This transcript captures a conversation between Seth Earley and Zoher Karu about why the most fundamental data discipline - governance - has become more important in the AI era, not less. They cover why ungoverned data produces bad AI answers faster, the difference between data quality and business context, why the models themselves are commoditizing while organizational knowledge becomes the differentiator, how to think about a platform approach rather than use-case-by-use-case vendor accumulation, and what the brakes-in-a-car analogy tells us about why governance enables speed rather than limiting it.

Transcript

Seth Earley: Welcome to the Earley AI Podcast, brought to you by VKTR. I'm your host, Seth Earley. In each episode, we explore how artificial intelligence and data are reshaping business strategy and operations.

Today, we are talking about data governance, AI readiness, and one of the most persistent problems in enterprise AI - the gap between the promise of intelligent decision-making and the messy, siloed, poorly governed data reality that most organizations are actually working with. The tools have never been more powerful. The data behind them has never been more consequential.

The organizations that treat AI like magic are the ones that are getting burned, because the same old problems - the data silos, the multiple sources of truth, the missing business context - do not disappear. They just get expressed faster, with more confidence.

My guest today has spent his career on the data and analytics side of exactly this problem, running enterprise-wide data and analytics teams at Sears Holdings, eBay, Citibank, and Blue Shield of California, and building personalization programs before personalization was something a large language model could attempt.

Zoher Karu has a phrase for what he is seeing right now, and I think it is one of the best summaries of the moment I have heard. Data governance has become sexy again. Zoher, welcome to the show.

Zoher Karu: Thanks, Seth, thanks for having me.

Seth Earley: Give our audience the version of your background you gave me - technical training, McKinsey, then a run of enterprise data and analytics leadership roles most people recognize. How did you end up thinking about AI the way you do?

Zoher Karu: My background is varied across multiple industries and technical by education - bachelor's, master's, PhD in electrical engineering. Then I went straight into McKinsey and did management consulting, where you learn a lot about a lot. One of the things you certainly learn is making decisions on partial information - which is a lot of what AI is doing these days. After McKinsey, I worked at several smaller startups doing interesting things like tracking customers in retail stores using video cameras, or analyzing phone conversations in call centers. Then, as you mentioned, leading large enterprise teams at bigger companies - Sears, eBay, Citibank, Blue Shield of California - trying to use data analytics to help accelerate business objectives. I have seen data being used for a while in different forms, whether that is people walking in a store, or conversations, or numbers in a database. Typically, organizations use data just to make reports, but not really to make decisions and drive their business.

Seth Earley: You told me data governance has become sexy again. That is a great line. Unpack it, because most people think generative AI made governance less relevant rather than more relevant.

Zoher Karu: We are in the middle of a transformation. There is certainly a hype cycle around the latest version of AI - generative models and transformer models. And it will be transformational despite the hype. But at the end of the day, it is based on data. Where is that data? Does everybody understand what it means? Because machines are also trying to understand what it means.

In the past, there was a lot of effort around data governance - in the banking world, for example, making sure critical metrics were calculated correctly, making sure there was traceability. And today, you can point powerful tools at your data. But if the data is not clean, you will generate answers and you will not be sure the answers are correct. These AI models have a certain knack for always sounding like they are right, but you never know if they actually are.

The classic problems - data silos, multiple sources of truth, ambiguity about how things connect together - always existed, they still exist, and just because you are pointing more powerful engines at them does not make them go away. Otherwise you just make bad decisions faster.

Seth Earley: What are the misconceptions you run into when executives think about AI and data?

Zoher Karu: The misconceptions usually start somewhere like: we need to be using AI, let us just use some AI, without really understanding what problem we are actually solving. And then just because you point a large language model at part of your database does not mean it can figure out what is what. There might be four columns called sales. How does it know which one you actually meant?

When I was at Citibank running analytics across multiple countries, everybody had a different definition of active customers. You could not compare across countries because everyone had a different definition of what active meant. Until you actually tried to bring it together, you realized you need to standardize. Executives can say go compute this, and you can compute something. But you are not sure it is right unless you look through how it got to that answer. If there are multiple versions of the truth, it may have picked the wrong one.

The analogy I use is data is like blood. The body is like an enterprise, with many organs - each one a different department - each needing the blood for a different purpose. But there is only one version of the blood. It is shared, it is not contaminated, it is not leaking out. That is the mentality you need when managing your most valuable enterprise asset.

Seth Earley: And the challenge is compounded when you recognize that different departments transform data as they use it. Engineering creates specifications, marketing creates specifications, support creates specifications. Whose specifications are right? It depends on context and audience. But someone needs to be the source of authority. Somebody needs to be responsible for adjudicating that.

Zoher Karu: The root cause of a lot of data problems is human behavior - people will use data for their own need and will not care about data for somebody else's need. If they have to create their own little access database on the side to do their task, they just will.

I asked someone at a bank once: how many loans did we sell yesterday in Malaysia? Apparently, the hard word in that sentence was yesterday, because they only had monthly reports. When we got the number, I asked for it by time of day. They did not understand why time of day mattered for their commission calculation. But marketing cared, call center cared, many others cared. People are focused on their own silo, and that is how you run into different definitions and versions of the same number.

The power of AI, and of the recent transformer models specifically, is connecting across domains. If you are trying to solve within one domain, fine, you will get some value. But the real power is asking questions across areas of your organization - how does call center behavior compare to that person's login behavior on the website? You can start asking those questions if you can connect the dots. If you have different versions of what a customer is, it will be hard to do these kinds of things.

Seth Earley: Talk about what the models can and cannot do when it comes to business context. Clean data is necessary but not sufficient.

Zoher Karu: The models themselves are moving toward commoditization. They are ingesting all of the same data and the differences between them are narrowing. The differentiating factor will be how well you have given the model context for your business.

Imagine hiring a brilliant analyst off the street. She knows how to query data like nobody's business. But then you say, go manage my supply chain. She does not know how to do that. You have to train her - this is how you handle VIP customers, this is how you handle international shipments. All of that knowledge is somewhere, usually in somebody's head, and you have to get it out and get it trained.

When you work with an LLM, you can give it all the data you want and it can be pristine, but if you do not tell it the context around the way to use that data, that is the next wave of problems. Your data is your enterprise asset - but so is the way you run your business. And that is typically captured loosely in documents, Slack messages, emails, or not captured anywhere at all. It is tribal knowledge. That is what is going to start differentiating organizations, because yes, you can use LLMs for efficiency gains, but the real value is helping you make decisions that help you run your business - and that is not always captured in just data.

Seth Earley: I was at the Snowflake conference earlier this summer, and a year ago it was all about agents. This year it was all about context.

Zoher Karu: Right. Because agents can go act, but if they do not have the context in which to act, what are they going to do? You can get brand new, super smart people off the street - they still do not have enough information. You still need to train them. In the same way, you need to give a model more than just data.

That information is often locked in one piece of code. VIP customers get 45 days to pay and everyone else gets 30 - that rule is probably coded somewhere, and it runs, and it operates. But that knowledge is not disseminated across the organization. Customer service and marketing cannot take advantage of it. They all recode the same rule over and over. Platforms that allow organizations to start capturing business context, ontology, business rules - that is the next wave of what we are going to see.

Seth Earley: What should executives do differently as they are deploying AI and evaluating vendor claims?

Zoher Karu: Many vendors will say I can solve this problem for you - and the next sentence is: give me all your data. Hold on. The easiest place to start with AI is productivity improvements. Automating call center notes after a phone call. Getting code written in five minutes instead of five days. You will get benefits and you should start there, because it demonstrates value quickly.

But the real value of AI is not process improvement. It is business process change. On the call center notes example - yes, you can automate the notes. But have you thought about why you are taking phone calls at all? That is a different question, and you will get there eventually, but start small and build toward it.

Be very careful of solving point problems one at a time with a different vendor each time. That creates a large mess. Think about a platform perspective - what is the suite of capabilities I am looking for - so you at least have a roadmap rather than throwing use case darts at the wall. You are solving classes of problems, not individual symptoms, and the platform you choose should have coverage across those classes.

Seth Earley: If there is one thing to take from this conversation, it is this: the tools are getting dramatically more capable, but the underlying disciplines do not go away. They have gotten more important.

Data governance is sexy again not because the technology demands new governance, but because the old governance was always the actual bottleneck - and now the cost of skipping it shows up faster, with more confidence behind the wrong answer.

Zoher Karu: And there is even more to govern now. One way to think about governance is that controls are not meant to be internal bureaucracy - not mud you have to slog through. Think of them as the brakes in your car. The reason you are able to go around a curve at high speed is that you know you have brakes. If you did not know you had brakes, you would inch your way around the curve, cautious, going slowly. You need controls and governance to know what the absolute limits are - and operate at the limit.

Seth Earley: Zoher, thank you for joining me and sharing what decades of working at the intersection of data, analytics, and enterprise decision-making actually teaches you about what it takes to make AI work. And a final thank you to our sponsor, VKTR. To our listeners, thank you for tuning in to the Earley AI Podcast. You can also find us on dLogos at dlogos.xyz/podcasts/earley-ai-podcast-271271ce, where you can vote on upcoming guests and submit questions and topics for future episodes. Be sure to subscribe for more conversations on how AI is shaping the future of business.

Zoher Karu: It was great, I enjoyed it, thank you.