#247 Barr Moses: Why Reliable Data is Key to Building Good AI Systems

0:00The way in which the data estate has changed in the last 10 years has been pretty radical. The way in which people manage and store their data has changed dramatically. However, the way in which we manage that estate hasn't changed almost at all. And that's really where Monte Carlo, that's the reason that we were created, is to help organizations that now need to manage this very complex system. Think of it almost as like a car manufacturing, right? You have all these different pieces that you need to put together. How do you know that the car is safe to drive? To put this in simple words, most organizations just build the car, drive it at 120 miles per hour, and then they learn that the door flies out or something, right?

0:40That's really how we've been treating reliability of data. What Monte Carlo does is connect to your systems across your data state, and learn what your data looks like, help you first and foremost detect issues. We can talk about what sort of issues can come up and what does that look like, But first and foremost, allow data and AI teams to be the first to know about data and AI issues. And then the second is help them triage and resolve those issues. What does the future hold for business? Ask nine experts and get 10 answers. Bull market, bear market, rates rising or falling, inflation going up or down.

1:20Can somebody please invent a crystal ball? Until then, over 40 ,000 enterprises have future-proofed their business with NetSuite by Oracle, the number one cloud ERP, bringing accounting, financial management, inventory, HR into one fluid platform. With one unified business management suite, there's one source of truth, giving you the visibility and control you need to make quick decisions. With real-time insights and forecasting, you're peering into the future with actionable data. If I were a larger organization, this is the product I'd use. Whether your company is earning millions or even hundreds of millions, NetSuite helps you respond to immediate challenges and seize your biggest opportunities.

2:14Speaking of opportunities, download the CFO's Guide to AI and Machine Learning at netsuite.com slash IonAI. That's NetSuite, N-E-T-S-U-I-T-E dot com slash IonAI, E-Y-E-O-N-A-I all run together to get the CFO's Guide to AI and Machine Learning. The guide is free to you at NetSuite.com slash IonAI. NetSuite.com slash IonAI. Thanks for having me, Greg. I'm excited to chat with you. My name is Barb Moses. I'm the co-founder and CEO of a company called Monte Carlo. We're the creator and leader of a category called data and AI observability. We started the company about five years ago. We're honored to partner and serve more than 400 enterprises for various industries.

3:15So these think about companies in gaming, media, retail, fintech, technology, financial services, etc. Each of them have lots of data teams, whether it's data analysts, data engineers, machine learning engineers, AI teams, and they're all working to create data products. A data product can be a report that an executive is looking at, it could be a specific dashboard they're using to drive decisions. It can be an internal product, or it can be an external product that a customer is using, or it can be a generative AI application. In each of those instances, data teams are responsible to make sure that those data products are reliable and based on trusted, accurate data.

3:59That's what Monte Carlo does. We help organizations use data by making sure that that data is of high quality and high trust. Yeah, I had a guy on not too long ago, and they had a product called Data Refinery. And I know that there are a lot of other companies doing similar things because the data that we're talking about for training or fine-tuning AI models is largely unstructured. and organizations have just tons of unstructured data in various forms. And these guys take that unstructured data and then transform it into something that's machine readable. Do you guys do that as well? We, no.

4:54So we don't do that. What we would do in that situation, let me take a little bit of a step back first. I think in the last decade or so, most of the work was done with structured data. And so monitoring or observing the reliability of structured data is something that is critical, and we do that very well. And just as an example for the things that you want to look at, you want to make sure that the data, the field level is accurate. You want to make sure that the data arrives on time, et cetera. You can go into more details into what that means. I think now there's sort of a rise with unstructured data, to your point.

5:33I think companies or enterprises are earlier in their journey for that. And one of the important things that folks are doing right now is trying to understand how to monitor and observe the quality of unstructured data. That's a very difficult problem. One of the things where we can help is it is common for, We've seen instances where our customers have structured, unstructured data, perhaps like in the example that you suggested, and then used observability to make sure that that data is accurate. So maybe let me give a practical example to bring this to life a little bit. You know, we work with a Fortune 500 insurance company.

6:12And in insurance companies, there's a lot of support tickets or support conversations that happen. And so, you know, support conversations with end users and support service teams are recorded. And those are basically unstructured data. And so what this company did was used LLMs to actually score or grade those conversations on a scale from zero to 10, where 10 was, you know, the sentiment was very positive of the conversation and zero, the sentiment was very negative. And so that allowed them to have a view of, you know, here are the positive conversations, here are the negative conversations, sort of do an analysis of that.

6:53Where Monte Carlo or where data and AI observability helped in this case is actually observing the output of the LLMs. So imagine that, you know, you got a score of 13 for a conversation. What does that mean? Or a score of negative two. How do you interpret that? When the scale is zero to 10, those outlier values don't mean much. And so actually having observability to make sure that the outputs make sense in this case is how you can start to unpack what does it mean to monitor unstructured data. Does that example sort of make sense? Yeah. Help me make it real? And first of all, why Monte Carlo?

7:32Did you use Monte Carlo Tree Search or something like that? Great question. We actually do not. It goes back, it traces back to sort of the origins of starting the company. You know, we're data people and like to work with, you know, like proud of ourselves to be working with the best data teams in the world. And so we were looking for a name that would resonate with data teams and, you know, be perhaps a bit catchy. And, you know, I think we have some Monte Carlo fans at our company in terms of, you know, the Formula One racing. Yeah. Yeah. And so those are some of the fun origins, if you will.

8:15So it's both, you know, something that data team, data and AI teams know and find familiar, but also sort of a nod to the team. Yeah. And so it's a platform. Is it a SaaS? Is it a service or is it a tool that people subscribe and then can use? How does it work? Good question. So by and large, it's a SaaS platform. And the most common way in which organizations work with Monte Carlo is observing their data estate. And so what do I mean by that? But the way that organizations today process, ingest, generate, create data is through a plethora of systems. And if you think about folks migrating to the cloud, a very common stack, if you will, could include upstream source systems.

9:19You might have a data lake house, data warehouse like Snowflake, Databricks, AWS, Azure, GCP. You probably have orchestration solutions like DBT, Airflow. And then you probably have, you know, either an AI model or you might have a BI solution like Looker or Tableau. so that you can imagine sort of this pretty complex data estate. Now, you know, the whole sort of premise of Monte Carlo is, you know, the way in which the data estate has changed in the last 10 years has been pretty radical. The way in which people manage and store their data has changed dramatically. Like, you know, those solutions that I just mentioned really grew in popularity in the last 5 to 10 years.

10:14However, the way in which we manage that estate hasn't changed almost at all. And that's really where Monte Carlo, that's the reason that we were created, is to help organizations that now need to manage this very complex system. Think of it almost as like a car manufacturing, right? You have all these different pieces that you need to put together. How do you know that the car is safe to drive? You know, most, to put this in simple words, you know, most organizations just build the car, drive it at 120 miles per hour, and then they learn that like something, you know, the door, you know, flies out or something, right?

10:52That's really how we've been treating reliability of data. What Monte Carlo does is connect to your systems across your data state, learn what your data looks like, help you first and foremost detect issues. We can talk about what sort of issues can come up and what does that look like. But first and foremost, allow data and AI teams to be the first to know about data and AI issues. And then the second is help them triage and resolve those issues. Because the reality is that even if you're sort of inundated with alerts about things breaking, if you don't have a good strategy for how to do root cause, for how to do triage and resolution, your team is going to be wasting a lot of time on that.

11:36And so we think about sort of, you know, both of these things as two sides of the coin. Both of them are very hard to do well. And yeah, does that make sense in terms of what we do? How do we go into more depth? Yeah. And do you use, and is this primarily looking at data for AI models or is it looking at any enterprise data? Yeah, great question. So, you know, I think before sort of this AI craze, if you will, most organizations were really thinking about, you know, their data estate and not so much about AI. I think today you can't think about one versus the other. And so I think any efforts to distinguish between the two are probably a mistake.

12:26And I'll explain why. I think, you know, if you think about this like two years ago or so, I think there was really sort of this, and there still is sort of competition in terms of, you know, what's going to be the best, latest and greatest LLM, right? And what's going to be the most important sort of foundation model in the reality? And you're seeing this, you know, with OpenAI and Meta and Anthropic and, you know, with DeepSeek, we sort of appended the whole thing. And, you know, for folks who don't know, DeepSeek is a company that's sort of allowed to run these models at significantly lower costs, like orders of magnitude lower and basically sort of, you know, put into question the whole reason of, you know, how to build these things at scale.

13:14And I think what we are seeing is that every single company has access to the latest and greatest model. And so, you know, every single organization can, you know, use open AI or Anthropic or any other solution. And so in a sense, that doesn't actually present itself as a competitive advantage. because if everyone can do it, like everyone has access to, you know, whatever the best 2 ,000 PhDs can create. And then so the question is, what is my competitive advantage or what is my moat? And I think what we're seeing in working with companies is that the moat is really the data that's powering those models.

14:03Because, and I'll just give a particular example, Intuit is a company that has a financial, sort of AI-driven financial assistant, which basically is based on RAG pipelines and takes information about what kind of car Craig has and where does he live and all this information about Craig and then surfaces back to you recommendations, tax-advised recommendations. And so if, you know, the quality of that AI product is only as good as the quality of the data that it relies on. And so if I have the wrong information about you, Craig, I might be surfacing the wrong information. And that is terrible.

14:49I can talk about some disasters that companies have had by relying on wrong data. So actually, I think UnitedHealthcare in 2024 denied some claims based on wrong data. And so, you know, there's significant sort of societal and brand, if you will, implications for getting the data wrong. And so I think what we're seeing is that a lot of organizations are now rushing to get their data right, sort of be AI ready, if you will, so that the data that's powering their AI applications can be reliable and trusted. Um, we really sort of more than it's actually, you know, just I'll share another anecdote.

15:33We recently did a survey with data leaders, and we found that 100 % of them feel pressure to build AI applications. and 90, I think 96 or 97 % of them actually do, which is great. You know, I guess folks are up for the challenge. However, only two thirds of them actually think their data is AI ready or that their data can be trusted. And so there's this tension where people really want to build with AI, are putting the resources behind that, but don't have the foundational data or the systems to serve that data in a way that makes their AI products reliable and trusted. And because of that, I think we're seeing more and more companies working hard to make sure that their data is actually accurate and reliable.

16:32I can talk a little bit about the reasons for why data goes wrong. I think that's one of the key things that folks are trying to wrap their mind around because it's actually not straightforward to try to understand how to observe your data and AI estate. Let me just pause there. Yeah, well, yeah, I want to hear that. Is your platform or your system, is it a generative AI platform? I mean, do you have a conversational interface that then, you know, looks at the data and can answer, or is it more of an IT solution than that? Yeah, I mean, I would say, so that's a great question. So we definitely have AI features, if you will.

17:22So as an example, you know, when we start working with a data team, the first thing that they want to do is understand what their data looks like and understand what they should monitor. and you know what they should monitor could be for example like let's say if we're working with the sports team or baseball team and they're collecting information about you know player stats for example and so they want to set up monitors about you know the speed at which the ball is thrown right or any sort of statistics about the players as an example. And so actually like coming up with the list of monitors for that is quite hard.

18:06And so what we do is we use AI to come up with a recommendation for which monitors you should have. And so we actually analyze sort of the data, profile the data, and suggest that, make recommendations basically. That's one example. Another example is, you know, we allow data teams to basically write in natural language what sort of monitors they want. We use AI to help them create those. We also use AI to help them basically understand correlations between different things that go wrong. So what I mean by that, if someone has sort of a, you know, detects sort of a data issue or an alert, they want to understand the root cause.

18:51The root cause can be actually an amalgamation of a few different things. So we think of the things that can go wrong as the data, the systems, the code, and the models. So the data, you sort of mentioned, you know, we have both structured data and unstructured data. And, you know, an issue in the data could be like just a problematic data source that sent a bunch of nulls, for example. Systems, you know, it's your model APIs, your orchestrations. We should have talked about that before. And the problem can be like just a job failed, as an example. So that's data systems. Code is basically sort of what engineers write to transform the data.

19:37And you could have an issue with the code that happens all the time. If someone makes a change of the code, has a mistake in the code. And then the fourth is the models. And so you could have an outdated version of the model. The model might not be served on time. or there might be inconsistency in the results of the model. You could have the same prompt sharing a different result each time or maybe the prompt is poorly written and doesn't have all the contextual information to provide a good answer. And so, you know, when there's a data issue, the root cause can be traced to a failure in the data, the systems, the code, or the models.

20:16And oftentimes, it's actually all of the above together at once. And when a data team needs to decipher what went wrong and why, that's where observability can help. And so we use AI to help explain what went wrong and why. So we help correlate between the issues. There's an orchestration job that failed. There's also schema change that happened at the same time. There's a pull request that's related to those. And there's a problem with a third-party data source. And those three things sort of resulted in this model failure or in this report with the wrong data. So we use AI for that as well. Yeah.

21:00You mentioned monitors a few times. I'm not a data person, so what do you mean by monitors? Yeah, great question. So monitors can be monitors of all of, you know, there's a few types of monitors. Maybe the most basic type of monitor could be, I receive data from Craig every morning at 8 a.m. And I didn't receive data this morning. I don't know why. In a real-life example, this might be, I receive information from LinkedIn every morning at 8 a.m. And this morning at 8 a.m., I didn't get information. The API failed. So you can set up a monitor that automatically tells you, let me know if the data did not arrive at 8 a.m.

21:41Another example of a monitor can be, you know, I typically get from LinkedIn at 8 a.m. a thousand rows in my table, a thousand rows of data. And if I'm getting only 20 rows or if I'm getting 20 ,000 rows, I need to know because something's weird. And so you might set up a monitor that tells you, you know, maybe the volume of the data here is abnormally low or abnormally high. and I'll just give a final example you know let's say I'm tracking you know I'm tracking you know what kind of toppings I have in a pizza and suddenly I get chocolate what sort of a topping is that and so you know I might want to set up a monitor that says pizza toppings can only include these five values.

22:36And if I get anything else like that, then please let me know that must be an error. I cannot have chocolate on my pizza. That would be a disaster. So those are types of modern... Does that make sense? Happy to give more examples. Yeah. No, does data observability also go into measuring the bias of the data set in different ways? That's a wonderful question. You know, that's something that we're asking ourselves. We haven't built that at Monte Carlo today. That's not, it's definitely something that I think is going to be very, very important for a variety of reasons. We haven't built it to date, but perhaps in the future.

23:23Yeah. So it's a SaaS offering and enterprise connects to your platform and then does the platform automatically scan your systems and identify data sources or do you have to list them? Great question. So we would typically want or would have the data team point us to their data warehouse or their data lake, their data orchestration and sort of, you know, the various solutions in their data estate. And those can be many and varied. And, you know, we would typically want the data or AI team to direct us to that, and we would make the connection. The other thing that we would typically recommend is to start with a particular use case that you know matters.

24:15So, for example, you know, you don't want to monitor, or I'll explain why, Like, historically, data and AI teams have done a pretty good job of hoarding data. And so there's a lot of stuff out there. Now, you know, whenever you put monitoring or something on observability or something, there's a risk of alert fatigue. And so if there's some data set that I really don't care about, perhaps I don't need monitor for it. But if there's a data set, you know, that's being used every single day by the product teams or by my customers, something that's externally facing, You know, for example, we work with regulated industries where, you know, think about like banks or trading platforms.

24:58You know, you can't afford a stock number to be wrong for, you know, for a very long time or even for a short amount of time. And so that is data you need to monitor very, very carefully. Exactly. You know, on the other hand, if there's data that you're using maybe once a quarter, maybe it's okay, you know, that the frequency of monitoring is less than the severity of the monitoring. So, you know, we like to work with data teams on understanding what's the most important data that you have? What's the data that's really, you know, you're relying on it. It's mission critical to your business. Customers are using that data.

25:34You know, this could be financial data. It could be marketing data. It could be customer platform data. All this, you know, more and more organizations are identifying these critical use cases. And so we recommend, you know, to start with those or to have a strong sense of visibility and monitoring into those for sure to start out with. Yeah. And you've written a book about data. What's the focus of the book? Is it this observability question or is it broader than that? Yeah, it's a good question. So yeah, we wrote a book about sort of data quality and data observability. I think there's sort of an interplay between data quality and data observability.

26:18I think, you know, historically, the problem with data quality has been around for a very long time, around 20, 30 years. I do think the meaning of it has changed dramatically today because 20, 30 years ago, you really could get away with bad data. Very few people, very few organizations were actually using data. And if they were using data, it was very intermittently in small pockets of the organization, maybe just to report numbers to the street once a quarter, for example. But today's world, if you think about the types of companies, just take an example, Pilot is a company that manages truck stops and fuel information and needs real-time information on where are trucks, what's the fuel data, etc.

27:07And so in those instances, you can't afford to look at the data, the quality of the data only once a quarter. And so more and more companies are using data in a way that's very different than what it was five to 10 years ago. And so the definition of data quality needs to change as well. And so we really think of data observability as the next evolution of data quality. Data quality often stops at detection of issues. Data observability helps you take it from detection to resolution and triage or mediation of issues. Data is fascinating because since I've been paying attention, which is only since about 2017, there's just so much data that's being created in every organization.

27:57As you said, people didn't think about it a lot before because it wasn't considered that valuable. But now that you have AI that feeds on data, people are starting to store it and analyze it. And how many different kinds of storage, you know, S3 buckets and, you know, relational databases and all these, is that sort of surface expanding with increasing numbers of data storage solutions and things like that? And can you connect to all of them? Yeah, you're actually spot on. So one of the things that we've seen happen in recent years is this explosion of these technologies. And so oftentimes companies will have multiple from each of those categories that you mentioned.

29:00For different purposes, you know, one can be for analytical data, one can be for real-time data, one can be for data science and AI use cases. So there's actually legitimate reasons for why you would want to have different ways to store and process data. Now, as there's been this explosion of these solutions, it's really hard to make sure that the data is accurate and reliable in each one of those. And so you're spot on. The idea of data and AI observability or what Monte Carlo does is connect to each of those and help teams make sure that the data in each of those is accurate, reliable. And by that, I mean all of the data is there.

29:42The data is actually the right data. It's arriving on time. The volume of the data is what you expected. The connections between the data are the right data. Its output is the right output. There's sort of a lot that goes into making sure the data is trusted, sort of a variety of different techniques to make sure it's trusted. But you're right, it's a lot harder for organizations to do it now that data is spread across a very different number of integrations. So, you know, Monte Carlo, we have over 50 or so integrations that we connect to where our team would deploy Monte Carlo. Within 24 hours or so, we'd start seeing lineage.

30:23So lineage means that we automatically parse the connections between the tables and the fields to understand how is data flowing. And once we have that lineage, which is basically like a sort of think of it as like a graph of connections or like a tree of connections, then we can start understanding, okay, if there's an issue with this particular data set, what are all the different areas that are impacted? by this data set? And if there's a problem here, what is the upstream root cause of those? And so the first thing we do is sort of have that picture of lineage. And then after that, work with the teams to help make sure that there's the accurate sort of monitors on that data.