Susan Etlinger (00:05):
Welcome to Leading the Shift, a new podcast from Microsoft Azure. I'm your host, Susan Etlinger.
(00:11):
Have you ever been in a meeting where people were talking about data and you wished you could hit pause, ask your own personal expert all the burning questions in your head and then hit play again? That's what I was lucky enough to do Recently with Teresa Tung, Global Lead of Data Capability at Accenture.
(00:29):
Teresa is a researcher whose work spans breakthrough cloud technologies, including the convergence of AI, data and computing capacity. She's a prolific inventor, holding over 225 patents and applications. I sat down with Teresa to learn more about the changing nature of data and its value to an AI strategy. Think of it as a master class packed into a tight half hour.
(00:55):
Teresa, welcome.
Teresa Tung (00:57):
Thanks Susan for having me.
Susan Etlinger (00:59):
Of course. So folks, we've been talking, Teresa and I, for the last few weeks, and we have some kind of a really interesting conversation to get into. Obviously the world of data has changed so much and particularly, in the last 18 months. And one of my favorite questions to ask people is sort of what are you seeing? What's going on in your organization?
Teresa Tung (01:19):
So as you just mentioned, what we've done as an industry over the last decade in data has completely changed. Generative AI has made some of those skills that we've built around governance, data supply chains. It's all been pre-generative AI. And so data capability is looking at the vision as to where we're going. It is building up our talent and our assets to be able to meet those visions and really to be able to help our customers really achieve their ambitions.
Susan Etlinger (01:51):
Absolutely, and what's interesting to me is that if you look back about 10 years, and we look at the maturation of the internet, and we talked about 10 years ago, we were talking about big data and things like that, and social media data and all the different kind of data signals that were coming into the enterprise, today, I think you're really excited about and organization's proprietary data and what the opportunity is given generative AI. Can you kind of expand on that?
Teresa Tung (02:17):
Yeah. I think we've always been investing in proprietary data, but with generative AI, it seems that some of the focus has shifted. And I want to say that that's still really the key. With generative AI, it almost looks like magic with some of these pre-trained models. They're so powerful. They've been pre-trained on this corpus of internet data, so it could translate languages, it could write poems, it could write code. It's able to summarize emails and text and even take exams. So it seems almost magical that it could do a lot and it certainly can.
(02:54):
But for enterprise use, without your proprietary data that's data about products, your people, your processes, your customers, how are you going to do anything that's meaningful? So for some of the studies that we've done, for enterprises, some of that corpus of internet knowledge, anything that a third or a fifth-grader would know how to do that, that's good, but you're going to have maybe 30% accuracy for the things that you need to do in your business. That means it's 70% plus inaccurate. And so to be able to actually tap into the power of generative AI, just like the AI of the generation before, we really need to invest in your proprietary data.
Susan Etlinger (03:36):
So much of that data is unstructured now. We have documents, we have images, we have audio files, video files. So how do you see clients approaching, trying to understand that?
Teresa Tung (03:50):
Yeah. So 80% plus of enterprise data is unstructured, and that's the fastest growing class of data. And so you listed a whole bunch, documents and video and voice, and these are things that we have. Clearly there's a lot of enterprise. That's the biggest class, but we haven't really tapped into it at scale and for AI.
(04:12):
And so generative AI, what it does so well is it's able to extract value from this unstructured data. It could take even this video and transcribe it and summarize it, maybe write even a social media posts for both of us. So that ability to be able to use unstructured data, it's really unlocking this whole new generation.
(04:34):
I mentioned, and I think we started with saying that for a decade we've been working more on structured data, data that's in a tabular format with columns and rows. It's been structured in that way. It's to pre-define answer questions that we know we want to ask. That table in the set of columns is really saying, this is the sort of thing and the features that I want to ask about to power a report or to create a dashboard or to run analytics on it. This newest generation, because generative AI made it so easy to work with unstructured data. Now I can ask any questions. So I don't have to pre-define that structure. I can ask the question about something that wasn't created and it could even generate the graph in the chart for me just based off of what I want to know in the moment.
Susan Etlinger (05:24):
So clearly with unstructured data, so much of that you were talking about how with structured data, we know that it's in columns, it's in rows, we pre-defined it. So much of unstructured data though is undefined, right? It's surprising. It's human. Context is so important. So how do we understand that data?
Teresa Tung (05:46):
Yeah. So I like to look at what we've seen with the internet. Anytime we do internet search, and all of us have done, that's really bringing in context. So some of us do remember internet search as it was coming out. So when it was coming out maybe over a decade ago, it had the links, right? So if you search for Microsoft or you search for Accenture, you would get all these links and it would be our job to click into those links to be able to extract any information from it. Today, if you search for Accenture, it will give you a little card in the front that actually gives you the office closest to where you are because that context. It knows where you're searching from. It's guessing that most people who are looking for a company, companies have physical locations, want to go to that company. That's bringing in context. And internet search is using something called a domain knowledge graph. It's that context based off of anticipating what the user's looking for.
(06:43):
In the same way that we have seen with internet search, we're going to see with generative AI. So generative AI, people talk a lot about vector databases. Vector databases act very much like internet search. It's going to give you similarity like the internet search that I described maybe a decade plus ago, lots of links. As we have a lot, lot more volume of data, eventually that's going to be too much, and you don't want to click through all those links. So just like internet search, having that knowledge graph or the semantic layer that adds the domain context, semantic means meaning. So the meaning that's needed, just like how we were able to tell Accenture as a company, I want to go there, so it's going to show me on that. Maybe it's going to show me how many it should cost to get there via one of the ridesharing services. All of that is from the context.
(07:31):
So in business now, what it means is that I might have the same customer data, but I'll have a different context. If I'm looking at a perspective customer, if I'm doing a customer sales transaction, if I'm doing customer service, all of them is still looking at the same data, but the context you want to give in each of those use cases will be different. So in the same way, a lot of what we now need to invest in as companies to use generative AI is the data for sure, but it's also capturing this context.
Susan Etlinger (08:00):
Yeah. And there are some situations in which you may not have that data, that data may be sensitive in some way. For example, healthcare data or data about children, and you're trying to solve a problem. So one of the things I've certainly heard and read a lot about is the idea of synthetic data. And I'm curious to know how do you think about synthetic data? And what do you think the right use cases are for it? And a little bit too about how do you approach synthetic data?
Teresa Tung (08:31):
Well, Susan, you just listed two great use cases. One is where it's sensitive. So the data is so sensitive that I wouldn't want to share it with you in the clear, and sometimes legally I'm not allowed to. So if I could create synthetic data that has the same characteristics so that you could do AI and in some cases fine-tune or train a model, that's one sort of case around sensitivity.
(08:56):
I think the other case that you talked about is just where there's gaps in the data. I wish I could do AI for it, but I can't RFID-enable everything. So an example would be with self-driving cars. Self-driving cars use a lot of synthetic data. You could bet that you don't want the car to get into all sorts of dangerous situations so that it can have enough data so that it could train the model to drive. And yet you need to make sure it covers all those situations so that we stay safe when we're using the self-driving car. So again, using synthetic data to fill in gaps.
(09:30):
You mentioned how do I see this different than before? So I think things that we did before, digital twins, engineering simulations, those still hold true. But again, here's generative AI for generative AI's sake. So when we did a fine-tuning of a model, we didn't have enough good quality data to fine-tune that day one. We had a lot of data and we had our human expertise who knew what this should look like. And then together, they're using generative AI to create the best quality data that covers all the use cases in the same way that the self-driving car would cover all these use cases. And now you have a corpus of really high-quality data that you could use for model validation, you could use for model fine-tuning and model training. So it is really powerful now with generative AI.
Susan Etlinger (10:26):
So do you have any best practices about working with synthetic data? Are there things that, for example, people in the business or technical decision makers who are not experts in synthetic data or even in unstructured data should be thinking about?
Teresa Tung (10:40):
I think really thinking about treating data as a product is how we're looking at this whole space about what to invest in within your proprietary data. We mentioned proprietary data is your competitive advantage. So look across your business, look at the business processes in which you engage. What data do you need to be able to better do that process or do it completely differently? And what sort of data comes through that process that is really unique, that is going to actually be something that might be of use for other people? So if we're thinking about supply chain as an example, do you have something about the demand or about the safety and the risk that is really unique to be able to both improve your processes but also act as a, almost a new revenue stream and a new stream of optimization? So whether we thinking about synthetic data, structured data, unstructured data, really thinking about it as a product just like we have other products that you can buy. The data itself should be assetized as a product.
Susan Etlinger (11:44):
Years ago I was speaking with somebody and he came up with this idea or shared this idea of a compounding data advantage. The idea that when you have data, you think about it as money in the bank or money in your investments, that over time, you invest in it, you invest in it, but then over time it accrues value that you might not have anticipated at the beginning. And so I really like that idea because I think that there are so many ways of understanding processes, of understanding people, of understanding organizational dynamics or the behavior of products and the behavior of the world, everything from climate to elevators to cars to children's toys. And it feels as though, it seems as though there is this notion of a compounding data advantage is valuable because you don't necessarily always know what you're going to be using it for in the future. And having access to it and understanding it and being able to work with it, feels like that's the sort of origin of a lot of value that organizations can unlock in the coming years.
Teresa Tung (12:55):
Yeah, a hundred percent. I like to think about it as an organization's collective intelligence. It is all the digitized assets, whether it's structured, unstructured. It is the knowledge... I mentioned things around the semantic layer and the knowledge graph. That's the knowledge within people's minds. That's the context that, given my experience, I know when I make a decision, here's the data that I need and the format in which I need it in which I'm going to make a decision. So all of this intelligence is a really key part that when we digitize it, that's where it's compounded.
(13:25):
If you only had data that takes... Again, if we were a products company, if you only had data around one part of the business, you might have the best supply chain or you might have the best R&D. But once you have data across the business where everybody can see, now as if I'm sitting in R&D, I could see exactly the issues with service so I can make a better device next time. If I'm sitting in the supply chain and I can see what R&D's planning to do next, I might be able to better procure and plan my procurement so that I can meet that demand. So all of this by being able to see more. That's that compounding effect that you were talking about. And even more so when you are working within a interconnected world where all our partners are now sharing data.
Susan Etlinger (14:13):
So sharing data of course brings up the governance and guardrails. Talk about that. What are you seeing with clients? What are they thinking about? Obviously, the regulatory landscape is changing. There's a lot to think about in terms of the way we use data. What kind of counsel do you give your clients and how do you think about it at Accenture?
Teresa Tung (14:34):
I think this is one really key part to get right. So regardless of if you're a company that says, "We're not ready for generative AI at all. We don't need it. We have data governance," I'll say that you still need to look at this particular dimension. One is because generative AI in apps is so easy to use. My kid's probably doing their homework using ChatGPT at the moment, and you can bet a lot of our people are doing the same thing. So really understanding your governance landscape and how it changes within generative AI, whether you're not planning anything, all the way through to if you're going to be creating your custom models, that's really key. And as part of that, you mentioned a whole bunch of things like the regulations, some of the risks, because generative AI, I just mentioned how easy it was to be able to find relationships across data.
(15:28):
That could be a super scary thing to say to someone, "Oh, it's so easy to find relationships across data. Now I can actually stitch together. This is who we're talking about." And so adding back in those governance controls, which again have been defined for structured data largely. So any sort of access control largely has been done for structured data. If it is for unstructured data, it's can I use it or not in its entirety? Once you've indexed something against generative AI, we don't want to say that we only limit it towards things that are not sensitive. That's currently how we're starting. But eventually, some of the promise is being able to bring in some of your IP so that generative AI can help you find new relationships and new ways to monetize. So we really need to be able to add that back in. So I think this is a really key area.
(16:22):
Some of the things you mentioned, synthetic data is one way that we get around it. Another way that we look at tackling this issue is around confidential compute. So this is some of the new programmatic capabilities that allow data to be processed in encrypted fashion. So typically, it's encrypted when it's at rest and in motion, so when it's stored and being transmitted. But when it's computed, you have to decrypt it. This new capability keeps it encrypted. So one thing you could do is you can encrypt all the data and then you're using that with generative AI only when you need to actually see the result, then you decrypt it for those specific use cases. So it actually punts a little bit of that decision about when to use it, when to see it for a later time.
Susan Etlinger (17:09):
Yeah, that's really interesting. It just makes me wonder though, are there, in terms of themes, in the conversations that you have with clients, are there any sort of common misconceptions or even just common strategies that you think are really useful? I'm just curious to get a pulse on what you've been seeing.
Teresa Tung (17:30):
Well, definitely we mentioned treating data as a product. Part of that is a pivot in terms of op model. So it's thinking about the value and right sizing investment. So when we looked at data, sometimes we started as an industry to say, "Let's bring it all into a modern platform first. And then after we've done that migration modernization, then the magic happens." Sometimes just bringing it into that single platform, it could take months if not years.
(17:58):
And so instead, you might actually start in a more agile fashion. So when we think about it as a product, not every product has to be done in the same way. You're going to right size the investment. So some things you might need to bring centrally right away, some things maybe it's fine where it resides. And so being able to now have a more federated approach to data so that there's a hub, and that's a central team that acts as a design authority, they might have the central platform that I mentioned, and that's where the best sort of sets of data that's needed the most. And you want to be the most efficient and the most accurate and really guarantee the quality of that. The other data, we're going to move it as it's needed. Some data, it might never move, and some data that we're going to want to move it right away because the business needs it right away, and now we can match the investment to that business need.
Susan Etlinger (18:55):
And do you see the operating model shifting over time? I mean, it feels like at the beginning of something, especially when there's a lot of unknown, you want to have some more centralization, you want to have a lot more control. Because one of the issues I think, or one of the challenges with data is that we're making data widely available to people. And at the same time, there's a lot of science and a lot of statistical understanding and sort of we're all learning. Every one of us is learning, from the most sophisticated data scientist to somebody in a different department who is just trying to build a product or who's trying to create a marketing campaign. So I just wonder, when you look out a few years, do you see the operating model shift? And do you think that that's going to be static or are we always going to be sort of more in the mode that we're in now?
Teresa Tung (19:45):
I think I'm seeing a shift towards a hub and spoke. So whether you start from a very distributed siloed op and you need some more of that centralized function because it's such a new space and you want to maximize the learnings to minimize risk and to maximize opportunities, or if you started very central as a very strong central team, and now generative AI is making data so much more democratized for use around the business.
(20:12):
So regardless of which direction you're coming into the hub and spoke, where the hub is really led by tech. It is a central design authority. And I am very clear. It's a design authority. They're saying, here's the standards of what good looks like, but the investments and the actual how you carry out which data products you invest in, how much should I invest, even I'm going to use it for, that's within the business domains. And so in this case, HR, finance, supply chain, commercial. All of these different domains could own a lot more of their data, again, because generative AI makes it much easier for anyone to be able to ask questions about it and to eventually be able to work with it.
Susan Etlinger (20:55):
Yeah. That makes a lot of sense. And so at the top of the conversation, you said that generative AI is jumpstarting data capability. So you made that comment. I'm curious if there's anything else that you want to add to that now that we sort of walked through your six principles.
Teresa Tung (21:12):
Well, this is the sixth one. So I think I talked about five principles for [inaudible 00:21:17] data, synthetic, unstructured contextualization with that semantic layer. And then we talked about risk.
(21:24):
So the last and the sixth principle is around jumpstarting. So I just listed a whole bunch of things, and Susan, you can say, "Oh, that sounds like so much work." The good news is that generative AI can be applied to data itself to be able to jumpstart data and reinvent how we do data, just like how it's reinventing everything else. And in fact, it's a really great place to do it because data practitioners, we understand the risks, we understand we're a bit more about where the technology gaps are. So it's a great place to be able to start experimenting with this at scale.
(21:59):
But we've been able to use generative AI for everything from writing the code that's needed to create some of these pipelines. Really important to do some of the code conversion. So I mentioned things like migration and modernization, how it used to take months if not years, being able to accelerate that because generative AI, it could help parse through the data and the logs and the code and give some understanding, well, what's there? We call that reverse engineering. I'm going to reverse engineer what's deployed. And then generative AI to do the forward engineering to write what it needs to be in that new target system. All of that's going to make things faster. So whether it's doing this sort of modernization story, it could be less fixed reports and dashboards and more using generative AI to ask any question you want on the fly. It might be with writing your data product descriptions or finding mappings to new questions that you want to ask. All of this is really accelerating how we did data before.
Susan Etlinger (23:01):
That's amazing. Thank you. Wow, this has been such a fun conversation, and it's just an incredibly... You got me even more excited. It's such a great time to be thinking about this and working in this field. So thank you so much for joining us.
Teresa Tung (23:14):
Oh, thanks for having me.
Susan Etlinger (23:21):
There is so much to unpack here. And please check out the show notes for a deeper dive into Teresa's thinking. But in the meantime, here are a couple of her insights that really resonated for me. First is the opportunity of unstructured data, especially because as she mentioned, it now constitutes over 80% of enterprise data. Unstructured data isn't exactly new. Some of you may remember the big data moment of the mid-2010s. I know I do, but it wasn't always easy to extract meaning from it, which is something generative AI is really, really good at. The second is the importance of context. And of course, it makes sense that if you're working with unstructured, undefined, often human and messy data, that how you interpret it is so important. We've all had the experience of trying to figure out the tone in a text or an email, so imagine that but now with your customer data. What can you learn about their experiences with you? What leading or lagging indicators can you find? And most importantly, what will you do with them?
(24:25):
But here's my favorite quote of the day. "Data is an expression of an organization's collective intelligence." And that brings us back quite elegantly to the central theme of our conversation, that proprietary data is key to an organization's competitive advantage.
(24:44):
We hope you've enjoyed this episode. And as I'm sure you know, new podcasts live and die on engagement so please like, comment, share, tell your friends, and let us know what you'd like to hear about. We're listening. If you'd like to learn more about how people and organizations are innovating with Microsoft Azure, visit azure.microsoft.com.
Credits:
Leading the Shift is a production of Microsoft Azure. Susan Etlinger is our host. Our executive producers are Hailie Meehan, Aaron Russell, and Erik Williams. Production support for this episode provided by Indigo Slate and RUN Studios.
Welcome to Leading the Shift, a new podcast from Microsoft Azure. I'm your host, Susan Etlinger.
(00:11):
Have you ever been in a meeting where people were talking about data and you wished you could hit pause, ask your own personal expert all the burning questions in your head and then hit play again? That's what I was lucky enough to do Recently with Teresa Tung, Global Lead of Data Capability at Accenture.
(00:29):
Teresa is a researcher whose work spans breakthrough cloud technologies, including the convergence of AI, data and computing capacity. She's a prolific inventor, holding over 225 patents and applications. I sat down with Teresa to learn more about the changing nature of data and its value to an AI strategy. Think of it as a master class packed into a tight half hour.
(00:55):
Teresa, welcome.
Teresa Tung (00:57):
Thanks Susan for having me.
Susan Etlinger (00:59):
Of course. So folks, we've been talking, Teresa and I, for the last few weeks, and we have some kind of a really interesting conversation to get into. Obviously the world of data has changed so much and particularly, in the last 18 months. And one of my favorite questions to ask people is sort of what are you seeing? What's going on in your organization?
Teresa Tung (01:19):
So as you just mentioned, what we've done as an industry over the last decade in data has completely changed. Generative AI has made some of those skills that we've built around governance, data supply chains. It's all been pre-generative AI. And so data capability is looking at the vision as to where we're going. It is building up our talent and our assets to be able to meet those visions and really to be able to help our customers really achieve their ambitions.
Susan Etlinger (01:51):
Absolutely, and what's interesting to me is that if you look back about 10 years, and we look at the maturation of the internet, and we talked about 10 years ago, we were talking about big data and things like that, and social media data and all the different kind of data signals that were coming into the enterprise, today, I think you're really excited about and organization's proprietary data and what the opportunity is given generative AI. Can you kind of expand on that?
Teresa Tung (02:17):
Yeah. I think we've always been investing in proprietary data, but with generative AI, it seems that some of the focus has shifted. And I want to say that that's still really the key. With generative AI, it almost looks like magic with some of these pre-trained models. They're so powerful. They've been pre-trained on this corpus of internet data, so it could translate languages, it could write poems, it could write code. It's able to summarize emails and text and even take exams. So it seems almost magical that it could do a lot and it certainly can.
(02:54):
But for enterprise use, without your proprietary data that's data about products, your people, your processes, your customers, how are you going to do anything that's meaningful? So for some of the studies that we've done, for enterprises, some of that corpus of internet knowledge, anything that a third or a fifth-grader would know how to do that, that's good, but you're going to have maybe 30% accuracy for the things that you need to do in your business. That means it's 70% plus inaccurate. And so to be able to actually tap into the power of generative AI, just like the AI of the generation before, we really need to invest in your proprietary data.
Susan Etlinger (03:36):
So much of that data is unstructured now. We have documents, we have images, we have audio files, video files. So how do you see clients approaching, trying to understand that?
Teresa Tung (03:50):
Yeah. So 80% plus of enterprise data is unstructured, and that's the fastest growing class of data. And so you listed a whole bunch, documents and video and voice, and these are things that we have. Clearly there's a lot of enterprise. That's the biggest class, but we haven't really tapped into it at scale and for AI.
(04:12):
And so generative AI, what it does so well is it's able to extract value from this unstructured data. It could take even this video and transcribe it and summarize it, maybe write even a social media posts for both of us. So that ability to be able to use unstructured data, it's really unlocking this whole new generation.
(04:34):
I mentioned, and I think we started with saying that for a decade we've been working more on structured data, data that's in a tabular format with columns and rows. It's been structured in that way. It's to pre-define answer questions that we know we want to ask. That table in the set of columns is really saying, this is the sort of thing and the features that I want to ask about to power a report or to create a dashboard or to run analytics on it. This newest generation, because generative AI made it so easy to work with unstructured data. Now I can ask any questions. So I don't have to pre-define that structure. I can ask the question about something that wasn't created and it could even generate the graph in the chart for me just based off of what I want to know in the moment.
Susan Etlinger (05:24):
So clearly with unstructured data, so much of that you were talking about how with structured data, we know that it's in columns, it's in rows, we pre-defined it. So much of unstructured data though is undefined, right? It's surprising. It's human. Context is so important. So how do we understand that data?
Teresa Tung (05:46):
Yeah. So I like to look at what we've seen with the internet. Anytime we do internet search, and all of us have done, that's really bringing in context. So some of us do remember internet search as it was coming out. So when it was coming out maybe over a decade ago, it had the links, right? So if you search for Microsoft or you search for Accenture, you would get all these links and it would be our job to click into those links to be able to extract any information from it. Today, if you search for Accenture, it will give you a little card in the front that actually gives you the office closest to where you are because that context. It knows where you're searching from. It's guessing that most people who are looking for a company, companies have physical locations, want to go to that company. That's bringing in context. And internet search is using something called a domain knowledge graph. It's that context based off of anticipating what the user's looking for.
(06:43):
In the same way that we have seen with internet search, we're going to see with generative AI. So generative AI, people talk a lot about vector databases. Vector databases act very much like internet search. It's going to give you similarity like the internet search that I described maybe a decade plus ago, lots of links. As we have a lot, lot more volume of data, eventually that's going to be too much, and you don't want to click through all those links. So just like internet search, having that knowledge graph or the semantic layer that adds the domain context, semantic means meaning. So the meaning that's needed, just like how we were able to tell Accenture as a company, I want to go there, so it's going to show me on that. Maybe it's going to show me how many it should cost to get there via one of the ridesharing services. All of that is from the context.
(07:31):
So in business now, what it means is that I might have the same customer data, but I'll have a different context. If I'm looking at a perspective customer, if I'm doing a customer sales transaction, if I'm doing customer service, all of them is still looking at the same data, but the context you want to give in each of those use cases will be different. So in the same way, a lot of what we now need to invest in as companies to use generative AI is the data for sure, but it's also capturing this context.
Susan Etlinger (08:00):
Yeah. And there are some situations in which you may not have that data, that data may be sensitive in some way. For example, healthcare data or data about children, and you're trying to solve a problem. So one of the things I've certainly heard and read a lot about is the idea of synthetic data. And I'm curious to know how do you think about synthetic data? And what do you think the right use cases are for it? And a little bit too about how do you approach synthetic data?
Teresa Tung (08:31):
Well, Susan, you just listed two great use cases. One is where it's sensitive. So the data is so sensitive that I wouldn't want to share it with you in the clear, and sometimes legally I'm not allowed to. So if I could create synthetic data that has the same characteristics so that you could do AI and in some cases fine-tune or train a model, that's one sort of case around sensitivity.
(08:56):
I think the other case that you talked about is just where there's gaps in the data. I wish I could do AI for it, but I can't RFID-enable everything. So an example would be with self-driving cars. Self-driving cars use a lot of synthetic data. You could bet that you don't want the car to get into all sorts of dangerous situations so that it can have enough data so that it could train the model to drive. And yet you need to make sure it covers all those situations so that we stay safe when we're using the self-driving car. So again, using synthetic data to fill in gaps.
(09:30):
You mentioned how do I see this different than before? So I think things that we did before, digital twins, engineering simulations, those still hold true. But again, here's generative AI for generative AI's sake. So when we did a fine-tuning of a model, we didn't have enough good quality data to fine-tune that day one. We had a lot of data and we had our human expertise who knew what this should look like. And then together, they're using generative AI to create the best quality data that covers all the use cases in the same way that the self-driving car would cover all these use cases. And now you have a corpus of really high-quality data that you could use for model validation, you could use for model fine-tuning and model training. So it is really powerful now with generative AI.
Susan Etlinger (10:26):
So do you have any best practices about working with synthetic data? Are there things that, for example, people in the business or technical decision makers who are not experts in synthetic data or even in unstructured data should be thinking about?
Teresa Tung (10:40):
I think really thinking about treating data as a product is how we're looking at this whole space about what to invest in within your proprietary data. We mentioned proprietary data is your competitive advantage. So look across your business, look at the business processes in which you engage. What data do you need to be able to better do that process or do it completely differently? And what sort of data comes through that process that is really unique, that is going to actually be something that might be of use for other people? So if we're thinking about supply chain as an example, do you have something about the demand or about the safety and the risk that is really unique to be able to both improve your processes but also act as a, almost a new revenue stream and a new stream of optimization? So whether we thinking about synthetic data, structured data, unstructured data, really thinking about it as a product just like we have other products that you can buy. The data itself should be assetized as a product.
Susan Etlinger (11:44):
Years ago I was speaking with somebody and he came up with this idea or shared this idea of a compounding data advantage. The idea that when you have data, you think about it as money in the bank or money in your investments, that over time, you invest in it, you invest in it, but then over time it accrues value that you might not have anticipated at the beginning. And so I really like that idea because I think that there are so many ways of understanding processes, of understanding people, of understanding organizational dynamics or the behavior of products and the behavior of the world, everything from climate to elevators to cars to children's toys. And it feels as though, it seems as though there is this notion of a compounding data advantage is valuable because you don't necessarily always know what you're going to be using it for in the future. And having access to it and understanding it and being able to work with it, feels like that's the sort of origin of a lot of value that organizations can unlock in the coming years.
Teresa Tung (12:55):
Yeah, a hundred percent. I like to think about it as an organization's collective intelligence. It is all the digitized assets, whether it's structured, unstructured. It is the knowledge... I mentioned things around the semantic layer and the knowledge graph. That's the knowledge within people's minds. That's the context that, given my experience, I know when I make a decision, here's the data that I need and the format in which I need it in which I'm going to make a decision. So all of this intelligence is a really key part that when we digitize it, that's where it's compounded.
(13:25):
If you only had data that takes... Again, if we were a products company, if you only had data around one part of the business, you might have the best supply chain or you might have the best R&D. But once you have data across the business where everybody can see, now as if I'm sitting in R&D, I could see exactly the issues with service so I can make a better device next time. If I'm sitting in the supply chain and I can see what R&D's planning to do next, I might be able to better procure and plan my procurement so that I can meet that demand. So all of this by being able to see more. That's that compounding effect that you were talking about. And even more so when you are working within a interconnected world where all our partners are now sharing data.
Susan Etlinger (14:13):
So sharing data of course brings up the governance and guardrails. Talk about that. What are you seeing with clients? What are they thinking about? Obviously, the regulatory landscape is changing. There's a lot to think about in terms of the way we use data. What kind of counsel do you give your clients and how do you think about it at Accenture?
Teresa Tung (14:34):
I think this is one really key part to get right. So regardless of if you're a company that says, "We're not ready for generative AI at all. We don't need it. We have data governance," I'll say that you still need to look at this particular dimension. One is because generative AI in apps is so easy to use. My kid's probably doing their homework using ChatGPT at the moment, and you can bet a lot of our people are doing the same thing. So really understanding your governance landscape and how it changes within generative AI, whether you're not planning anything, all the way through to if you're going to be creating your custom models, that's really key. And as part of that, you mentioned a whole bunch of things like the regulations, some of the risks, because generative AI, I just mentioned how easy it was to be able to find relationships across data.
(15:28):
That could be a super scary thing to say to someone, "Oh, it's so easy to find relationships across data. Now I can actually stitch together. This is who we're talking about." And so adding back in those governance controls, which again have been defined for structured data largely. So any sort of access control largely has been done for structured data. If it is for unstructured data, it's can I use it or not in its entirety? Once you've indexed something against generative AI, we don't want to say that we only limit it towards things that are not sensitive. That's currently how we're starting. But eventually, some of the promise is being able to bring in some of your IP so that generative AI can help you find new relationships and new ways to monetize. So we really need to be able to add that back in. So I think this is a really key area.
(16:22):
Some of the things you mentioned, synthetic data is one way that we get around it. Another way that we look at tackling this issue is around confidential compute. So this is some of the new programmatic capabilities that allow data to be processed in encrypted fashion. So typically, it's encrypted when it's at rest and in motion, so when it's stored and being transmitted. But when it's computed, you have to decrypt it. This new capability keeps it encrypted. So one thing you could do is you can encrypt all the data and then you're using that with generative AI only when you need to actually see the result, then you decrypt it for those specific use cases. So it actually punts a little bit of that decision about when to use it, when to see it for a later time.
Susan Etlinger (17:09):
Yeah, that's really interesting. It just makes me wonder though, are there, in terms of themes, in the conversations that you have with clients, are there any sort of common misconceptions or even just common strategies that you think are really useful? I'm just curious to get a pulse on what you've been seeing.
Teresa Tung (17:30):
Well, definitely we mentioned treating data as a product. Part of that is a pivot in terms of op model. So it's thinking about the value and right sizing investment. So when we looked at data, sometimes we started as an industry to say, "Let's bring it all into a modern platform first. And then after we've done that migration modernization, then the magic happens." Sometimes just bringing it into that single platform, it could take months if not years.
(17:58):
And so instead, you might actually start in a more agile fashion. So when we think about it as a product, not every product has to be done in the same way. You're going to right size the investment. So some things you might need to bring centrally right away, some things maybe it's fine where it resides. And so being able to now have a more federated approach to data so that there's a hub, and that's a central team that acts as a design authority, they might have the central platform that I mentioned, and that's where the best sort of sets of data that's needed the most. And you want to be the most efficient and the most accurate and really guarantee the quality of that. The other data, we're going to move it as it's needed. Some data, it might never move, and some data that we're going to want to move it right away because the business needs it right away, and now we can match the investment to that business need.
Susan Etlinger (18:55):
And do you see the operating model shifting over time? I mean, it feels like at the beginning of something, especially when there's a lot of unknown, you want to have some more centralization, you want to have a lot more control. Because one of the issues I think, or one of the challenges with data is that we're making data widely available to people. And at the same time, there's a lot of science and a lot of statistical understanding and sort of we're all learning. Every one of us is learning, from the most sophisticated data scientist to somebody in a different department who is just trying to build a product or who's trying to create a marketing campaign. So I just wonder, when you look out a few years, do you see the operating model shift? And do you think that that's going to be static or are we always going to be sort of more in the mode that we're in now?
Teresa Tung (19:45):
I think I'm seeing a shift towards a hub and spoke. So whether you start from a very distributed siloed op and you need some more of that centralized function because it's such a new space and you want to maximize the learnings to minimize risk and to maximize opportunities, or if you started very central as a very strong central team, and now generative AI is making data so much more democratized for use around the business.
(20:12):
So regardless of which direction you're coming into the hub and spoke, where the hub is really led by tech. It is a central design authority. And I am very clear. It's a design authority. They're saying, here's the standards of what good looks like, but the investments and the actual how you carry out which data products you invest in, how much should I invest, even I'm going to use it for, that's within the business domains. And so in this case, HR, finance, supply chain, commercial. All of these different domains could own a lot more of their data, again, because generative AI makes it much easier for anyone to be able to ask questions about it and to eventually be able to work with it.
Susan Etlinger (20:55):
Yeah. That makes a lot of sense. And so at the top of the conversation, you said that generative AI is jumpstarting data capability. So you made that comment. I'm curious if there's anything else that you want to add to that now that we sort of walked through your six principles.
Teresa Tung (21:12):
Well, this is the sixth one. So I think I talked about five principles for [inaudible 00:21:17] data, synthetic, unstructured contextualization with that semantic layer. And then we talked about risk.
(21:24):
So the last and the sixth principle is around jumpstarting. So I just listed a whole bunch of things, and Susan, you can say, "Oh, that sounds like so much work." The good news is that generative AI can be applied to data itself to be able to jumpstart data and reinvent how we do data, just like how it's reinventing everything else. And in fact, it's a really great place to do it because data practitioners, we understand the risks, we understand we're a bit more about where the technology gaps are. So it's a great place to be able to start experimenting with this at scale.
(21:59):
But we've been able to use generative AI for everything from writing the code that's needed to create some of these pipelines. Really important to do some of the code conversion. So I mentioned things like migration and modernization, how it used to take months if not years, being able to accelerate that because generative AI, it could help parse through the data and the logs and the code and give some understanding, well, what's there? We call that reverse engineering. I'm going to reverse engineer what's deployed. And then generative AI to do the forward engineering to write what it needs to be in that new target system. All of that's going to make things faster. So whether it's doing this sort of modernization story, it could be less fixed reports and dashboards and more using generative AI to ask any question you want on the fly. It might be with writing your data product descriptions or finding mappings to new questions that you want to ask. All of this is really accelerating how we did data before.
Susan Etlinger (23:01):
That's amazing. Thank you. Wow, this has been such a fun conversation, and it's just an incredibly... You got me even more excited. It's such a great time to be thinking about this and working in this field. So thank you so much for joining us.
Teresa Tung (23:14):
Oh, thanks for having me.
Susan Etlinger (23:21):
There is so much to unpack here. And please check out the show notes for a deeper dive into Teresa's thinking. But in the meantime, here are a couple of her insights that really resonated for me. First is the opportunity of unstructured data, especially because as she mentioned, it now constitutes over 80% of enterprise data. Unstructured data isn't exactly new. Some of you may remember the big data moment of the mid-2010s. I know I do, but it wasn't always easy to extract meaning from it, which is something generative AI is really, really good at. The second is the importance of context. And of course, it makes sense that if you're working with unstructured, undefined, often human and messy data, that how you interpret it is so important. We've all had the experience of trying to figure out the tone in a text or an email, so imagine that but now with your customer data. What can you learn about their experiences with you? What leading or lagging indicators can you find? And most importantly, what will you do with them?
(24:25):
But here's my favorite quote of the day. "Data is an expression of an organization's collective intelligence." And that brings us back quite elegantly to the central theme of our conversation, that proprietary data is key to an organization's competitive advantage.
(24:44):
We hope you've enjoyed this episode. And as I'm sure you know, new podcasts live and die on engagement so please like, comment, share, tell your friends, and let us know what you'd like to hear about. We're listening. If you'd like to learn more about how people and organizations are innovating with Microsoft Azure, visit azure.microsoft.com.
Credits:
Leading the Shift is a production of Microsoft Azure. Susan Etlinger is our host. Our executive producers are Hailie Meehan, Aaron Russell, and Erik Williams. Production support for this episode provided by Indigo Slate and RUN Studios.
