Leadership Insights: Trail of AI Tears

Transcript

Mike Vizard: Hey, folks. Welcome to the latest Digital CxO video cast. I’m your host, Mike Vizard. Today we’re with Tom Siebel who’s CEO for C3 and an early pioneer in this whole space of MLOps and data science. Tom, welcome to the show.

Tom Siebel: Thank you, Mike.

Mike Vizard: So where are we right now? What is the current state-of-the-art? I feel like a lot of places, for whatever reason, decided to roll out their own and built all these kind of tools and stitched them together. And it kind of feels like the early days of DevOps in a lot of ways. And now, I’m wondering if we’re going to move more towards integrative platforms because that’s exactly what happened in DevOps. So are we having the same journey; just a different set of technologies?

Tom Siebel: So I think AI is very confusing. It’s a – while it is not a relatively new field, we’ve seen a lot of a massive amount of development in AI in, say, the last ten years. You know, a lot of kind of disconnected components that provide – that do useful things. You know? Like pipeline management and machine learning libraries and data visualization and access control and, you know, what have you. Data aggregation.

And so, it’s very, very confusing ’cause we have hundreds and hundreds of AI things out there that, you know, are components that are – people are presented as and perceived of as platforms for building enterprise AI applications; things like Databricks and data robot and Dataiku and DynamoDB and Cassandra and Alteryx and SageMaker and you name it. And TensorFlow, ScaleIT, Digits, Python, R.

And while all of these things, each of these things, are very useful, none of them are in fact platforms for building any kind of application at all. They do things like data persistence or data virtualization, whatever it might be. And the way I think about enterprise AI is, you know, as you know I’ve been around the information technology business for a while.

And I think back to when we started first building enterprise education software in the ’80s. Enterprise application software as think of it today at places like Oracle and SAP and others . And we took – you know, the core was really – so database technology. On top of that, we built a bunch of tools that allowed us to build applications using relational databases. And this was, you know, with forms and reports and screens and whatnot.

And then, using those tools we build a large set of enterprise applications like ERP and CRM and manufacturing automation and human resource management, supply chain management, what have you. And these application snow at Oracle and SAP and others constitute about a $600-billion market. And these applications are very useful, in fact, I think, critical to organizations being able to do managerial and financial accounting.

And so then, we can report with perfect 20/20 hindsight, for example, what our inventory levels were. And we can then record this properly on our assets sheet. So if we’re asset ledger. And so, if we are Boeing and we have Boeing 777 that’s got a million components in its bill of materials, you know, flap actuators and ball bearings and resistors and transistors and integrated circuitry and the landing gear and auxiliary power units.

And they have a, you know, supply chain of these million components that goes from South Carolina – where I believe they build these things – all the way back to Shen Zhen. So a million components that build materials. And, you know, we have so many of these parts in each bucket. And then, we have that for the 767, 787, 757, 737, 7-so-7 what have you. It’s a pretty complex accounting process. And these organizations allow us – these enterprise applications allow us to report what our inventory levels were 60 days or 90 days ago or 360 days ago.

And in the case of Boeing, I think it’s about a $60 billion commercial aircraft company that, until recently, had about $50 billion of in-process inventory. Which is a lot of inventory. Or these applications allow us to tell us if we have a fleet of aircraft or a fleet of tractors or a fleet of automobiles, it’ll tell us, “You know, what was our deployment rate of these assets?” It’ll tell us what our customer churn was at Verizon or Bank of America 90 days ago or 180 days ago with perfect 20/20 hindsight.

This is what enterprise application software do. Okay. Now, if we look at this new class of what I think AI is about, as it relates to the enterprise, as we can provide these – that we can provide these platforms that make these enterprise applications predictive. And so, when they become predictive rather than descriptive, rather than looking at the world in 2020 hindsight, we can look at our inventory.

And rather than – in addition to telling us what we had 180 days ago, it’ll tell you for each part, how many of each of these parts we need to have in the bin of each of the next 180 days to meet the demand function. Rather than tell us what how many customers left us 90 days ago, it’ll tell us which customers are going to leave us in the next 90 days so we can take – now we engage in prescriptive analytics and we do something to retain those customers.

Rather than simply tell us what our non-deployment rate was for aircraft, tractors, trucks, whatever it might be – using AI-based predictive maintenance, predictive analytics – we can tell which of these aircrafts…for example, (we did this through the United States Air Force for like 5,000 aircrafts.) We tell them predictively which aircraft, which components in which aircraft, are going to fail.

And we do this 50 or 100 flight hours in-advance so that they can replace that component; be if flap actuator, auxiliary power unit, propulsion, flight management system, whatever it might be, so that that failure doesn’t happen – increasing the availability rate of aircrafts for the United States Air Force order of 45 percent. So when we applied these platforms, and I think you need a platform to these applications. These applications become predictive and become prescriptive.

So we can look around corners and we can change the future. Okay? We can lower costs. We can lower environmental impact. We can increase customer satisfaction. We can reduce customer churn so we can change the future and we can look forward. So I think the idea is that when we look at these markets for the projections that are out for this market for AI or predictive enterprise applications, this is a $600-billion market 2025.

So this is as large as the enterprise application software market is today. And I believe in one year or two years to three years nobody is going to be satisfied with 20/20 hindsight and everybody is going to want to be – in addition to that, want to be able to change the future. And I think therein lies the opportunity.

Mike Vizard: Seems like a lot of organization kind of get the basic concept and they’re dabbling in AI in various degrees. But they struggle. So when you talk to customers, what do you think is the biggest issue that they’re running into? Is it the data? Is it more the culture? Is it just the lack of skills?

Tom Siebel: We’ve been working with large customers around the world for the last 13 years. And wrestling with digital transformation. This issue of digital. And I believe there’s kind of a Trail of Tears organizations need to go through. And I’m not certain that you can shorten this. And, you know, the first step was cognitive computing – right – where we were going to bring in Watson.

We were going to bring in Predix and these massive, big bangs, and we’re going to place all the physicians at MD Anderson with lots of…you know, so this whole cognitive computing was a lot of hoorah and, you know, “Gee. I spent about $6 billion over ten years and Watson spent God knows most many scores of billions of dollars, and I’m not aware of one of those projects anywhere in the world that were successful.”

And then, we went from there; we brought in a consulting firm of our choice, and we came up with our digital strategy. And the general strategy, “Well, it’s all about Agile. We need to be Agile. We need to push everything down into the business unit and we need to do always proof of concepts with all these kind of toys that are out there.” These kind of interesting tools for citizen data scientists; like Dataiku or Alteryx. And not a lot came out of that.

And so then, we had this epiphany that “Well, it was garbage in, garbage out.” So before, we didn’t really do anything. We need to build the enterprise data lake. How much was spent on building enterprise data lakes. At Haliber, at Exxon, at the United States Air Force? And, by the way, who succeeded at building an enterprise data lake? We were going to take the data for, you know, health and human services or the large hospital system, or whatever it might be, and aggregate those data in a unified, federated, cohesive image.

And, you know, we would bring in the consulting firm of our choice and spend 10 to $20 million a year for five, six, seven, eight years and nothing would get done. So, you know, billions were spent building these data lakes to get over the garbage in, garbage out problem. And so, after nothing came of that, then we were saved by the cloud. And I don’t mean to underestimate the impact of the cloud, ’cause I think the cloud is a secular change in information technology, and it’s hugely important.

But along came Andy Jassy with AWS. Okay? And _____, and now the Google Cloud. and they have this big life preserver that they were sending us. Cause now we can put all our data in an S3 bucket. And again, everything was going to be okay. We’re going to put it in Redshift and S3 buckets. You know, a big table, a big query or whatever it might be; run it in the cloud with all these cloud services. And that was the panacea that was going to fix everything.

So most everyone of our customers spent one, two, three, four, five, six years trying to build these applications using the componentry that was provided by the cloud providers. Which by the way, I’m not diminishing their contribution because I think their contribution is staggering, and it changes everything about computing. But it’s very difficult to solve this problem.. And so, I think these are all the stages that companies are going through. And you got multiple stages of failure, and now you fire a few CIOs and then a few CEOs, and then finally you fire your CEO.

And then you bring in one of the large platform providers and the other companies like NG and Now and Coke Industries, the United States Air Force that are now deploying these massive-scale enterprise AI applications successfully using kind of cohesive platforms from competent professional software engineering organizations, rather than trying to cobble this to get stuff together by themselves. But I think that companies almost need to all go through these stages of failure before they’re ready.

Mike Vizard: Mm-hmm. Do you think we need to kind of bring together those data science folks that have their own culture these days and the application development teams and kind of make that a more cohesive unit? Cause it feels like they kind of work today in a much different cadence. A lot of organizations are lucky if they get a model together twice a year and then an update. And a lot of other folks are updating their applications every two weeks or so, and it’s hard to get those two squads to be in sync with each other. So then, do they need to become one squad?

Tom Siebel: Well, they certainly need to be coordinated, and they need to be working with a common set of tools and a common set of objects. Okay? And whether the objects are, you know – so they need to be with a common set of abstractions. And the abstractions might be machine learning models. Or the abstractions might be whatever the assets the company deals with; people or aircraft or vehicles.

And, you know, they – generally what we’re seeing where people are succeeding, they’re using these model-driven architectures that kind of abstract away all of the complexity of how data are persisted; all the point of where data are persisted; all the point areas across those data. And when you start dealing with, say, a healthcare population for maybe the population of the United States, you’re dealing with hundreds and hundreds of petabytes of data. So the number of points you need to follow are 10 to the 13th. Okay?

And then, you deal with – you abstract all of the processes that act on those data like encryption, most encryption, Access Control Queueing ETL and simply allow a logical aggregate. An aggregate might be a machine, or it might be a customer, or it might be an employee, or it might be an adversary. And you – oaky? And you deal with the aggregates that – so the data scientist, the citizen data scientists and the application developer don’t need to be so concerned with all this complexity of where and how data are persistent and the process is to act on them.

And so, this is kind of the logical extension of the work that the object management group did at the beginning of this century. And the logical sense about these, you know, highly componentized, service-oriented architectures that are called model-driven architectures; where, every time we develop an asset, we had a machine learning model or a tank or a car or a human being, that asset is then reusable by the rest of the enterprise.

And so, we can really accelerate the rate at which we bring these applications alive to enormous social and economic benefit. So if you look at – I believe that we are deploying – I’m confident we are deploying the largest production enterprise applications, AI enterprise applications, on the planet Earth. And, you know, some of these – if we looked at…____, which is the largest utility in Europe, which is a complex smart-grade analytics problem, in that case we’ve aggregated 150 trillion rows of data from 18 enterprises…I’m sorry, 18 instances of SAP, 12 instances of Salesforce.

SEBUL, two different SCATA systems; Atlas Maximo. We go out to the extranet; we have weather terrain and social media updated 62 billion times a day. And then we aggregate the data from 47 million sensors and 42 million smart meters into unified federated image, some of these data arriving at 90-hertz cycles, we process these data at the rate they arrive and then apply ML models to them for grade efficiency.

AI-based predictive maintenance assuming energy resource management; fraud detection; integration of renewable’s; customer churn; what have you. Allowing it now to deliver – this would be the largest utility in the free world – allowing it now to deliver a safer, cleaner, more reliable energy. We’ve had massive integration of renewables.

And this is about an 88-year-old organization currently generating I think €5.5 billion a year in economic benefit from these AI initiatives. So, you know, this is where – this is where we’re start…we’re not horsing around. We’re not toying around with all these kinds of shiny little AI objects that are in the, you know, Apache open-source community. We’re building real applications generating enormous economic and social benefit at NL, at NGS Shell, at United States and Air Force and Coke Industries.

Mike Vizard: Mm-hmm.

Tom Siebel: So that’s what we do.

Mike Vizard: Do you think, then, we’re in danger of going through the proverbial trope of disillusion and then people are going to miss the AI boat because it was so hard going upfront and it’s easy to poo-poo everything. But it may turn out that you’re going to be at a competitive disadvantage for the folks that are making it work.

Tom Siebel: The people who don’t make it work will be at a competitive disadvantage, and they’ll be acquired by those who make it work. But this is – I mean, this is a fundamental technology change. This is like the move from mainframes to mini-computers to personal computers. It wasn’t a question of – I mean, there were organizations that honestly believed they were never going to move to mini-computers.

I mean, there was organizations that believed that – I mean, even digital equipment for goodness’ sake. Dick Olson thought they would never use Ethernet. What did he use the ethernet for? He had the 8-inch floppy disc. Remember? Okay? And a PC. Who could possibly have used that? Whenever we introduced, you know, relational database systems in the ’80s, who needed that when they had ISAM and BSAM?

Well, if you didn’t make the transition to relational database, if you didn’t make the transition to enterprise application software, if you didn’t make the transition to Ethernet, if you didn’t make the transition to the cloud, you basically ceased to exist and you got acquired by somebody and become non-competitive.

So I think as it relates to enterprise AI, there are companies that are being hugely successful at this, there are companies who are, like, making very large-scale concerted efforts to succeed at it, trying to build it themselves, and then there are kind of companies that are kind of tinkering around. You know, those that succeed will be at the top of the stack. Just like all the previous technology change that we’ve seen. And those that don’t succeed will be acquired by those who do.

Mike Vizard: What’s your best advice, then, to business and IT leaders as they look at all of this stuff and sometimes they get a little overwhelmed and they’re not sure what to make of – what the very technical folks are telling them…but, you know, what should they be doing right now?

Tom Siebel: The organizations that we’re seeing succeed are projects and initiatives that are being personally driven by the chief executive officer or the chair of the joint chiefs of staff. Or the secretary of the Air Force; basically the CEO. And so, these are not generally IT-led initiatives or even CDO-led initiatives. They are CEO -led initiatives. There are mandates. I think that people that we’re seeing succeed are not looking at boiling the ocean.

They’re not trying to digitally transform DuPont. Okay? They’re taking a project, maybe stochastic optimization of supply chain; maybe energy efficiency; many ESG; maybe demand forecasting. A project for a product line and bringing it live and bringing it live in a short period of time; like say six months. And there’s only one way to measure the success of these projects.

And that is, “You know, what is the amount of economic benefit that is accruing to our shareholders? What is the amount of economic benefit that is accruing to our stakeholders? And what is the social benefit that is accruing to society at large?” And so, I think that the organizations that succeed are measuring only by those criteria. Okay? And they’re demanding results quickly.

Mike Vizard: Mm-hmm. Hey, Tom, thanks for being on the show.

Tom Siebel: Thank you for having me, Michael. It’s nice to talk with you.

Mike Vizard: All right, folks. Thank you for watching this latest episode of Digital CxO video cast. You can find this episode along with all our other ones on the digitalcxo.com website. And we invite you all to peruse them at your leisure and thank you all for once again spending some time on this. Take care.