Unlocking data strategy success: Practical AI, robust governance and agile management

In the latest episode of our Energy Transition Talks series, CGI Vice-President, Consulting – Data and Analytics Doug Leal discusses with Peter Warren the evolving landscape of data use in the energy and utilities sector, particularly in light of new AI applications. In the first instalment of this two-part conversation, they explore the challenges of scaling AI models, the move away from experimentation toward practical solutions and two key approaches to data management: the Data Lakehouse and the Data Mesh—both of which are shaping the future of data strategies’ success.

Utility organizations are facing increasing pressure to leverage data effectively for decision-making. This involves the integration of various data sources, such as Advanced Metering Infrastructure (AMI) and outage management systems, to enhance operational insights. While some organizations are already progressing in this area, Doug says, many are still in the early stages of their data journey.

Doug and Peter discuss two distinct approaches to AI: one that treats it as a novel tool to explore, and another that focuses on practical problem-solving. The latter, Doug says, is essential for developing a strategic approach to AI implementation, ensuring that solutions are not only effective for immediate challenges but also adaptable for future developments

“We need to be able to build a model or any type of AI solution in a way that will enable the organization to scale—not only scale that model for production, but also for everything that comes after that model, the innovation that comes after that model.”

The challenge of transitioning from proof of concept (POC) to production

Typically, a business unit recognizes the potential of a technology or model and decides to invest further. However, without a well-defined operational process to transition from proof of concept (POC) or proof of value to full production, this can create significant challenges and bottlenecks. 

As Doug shares, only 53% of models successfully progress from POC to production, making it an expensive endeavor when roughly half fail to deliver results.

Shifting focus to Minimal Viable Products (MVPs) and practicality

Peter agrees, citing a current client’s approach that skips the POC entirely, jumping ahead to develop minimal viable products (MVPs) right away. He explains their strategy involves creating solutions that are aligned with their organizational goals and can be effectively scaled. This ensures that the IT team can support the growth of these products and that the business can derive tangible value from them.

Doug has also noticed a shift in mindset among clients. As he sees it, there’s a growing emphasis on how to effectively transition ideas into production rather than just experimenting, reflecting an increased understanding of the importance of assessing the real value and return on investment for these initiatives. Given the substantial costs associated with infrastructure, data scientists and machine learning engineers required for model development, organizations are increasingly cautious about treating these efforts as mere experiments.

Read more in cgi.com

Visit our Energy Transition Talks page

Transcript

Peter Warren (00:01)
Hello again everyone and welcome to another session on our exploration of all things energy and utilities as well as the intersections with other industries. Today we're going to be diving into the data and the last few podcasts we've talked about how important data is but we never answered how do you get to the data? What are some techniques to move forward? So today I'm very privileged to have an expert in that area. Doug, do you want to introduce yourself?

Doug Leal (00:27)
Absolutely. Doug Leo, I'm a technical vice president with CGI and have more than 23 years of experience in the data and analytics area. I'm one of the leads here at our practice at the CGI data and analytics practice and serving our clients here in the Southeast of the United States. It's a very exciting time to be working with data.

Peter Warren (00:53)
Yeah, I think that's an interesting thing. We talked on previous podcasts, I'll just summarize it again. All of the other industries when we did our voice of the client surveys, and for those that are familiar with it, that's a survey we do with our existing clients and other folks, very much like what an analytics firm would do. This year, energy and utilities, there was a dip in satisfaction of data. So Doug, my apothecary

hypothesis on that was that they're rethinking their data because they've got the new uses for AI. Do you concur with that? How do you see data being re -envisioned in the energy and utility

Doug Leal (01:32)
Yeah, so utility companies, are facing an increasing pressure to become more data -driven. And in order to make better decisions, they need to be able to collect, store, analyze, and drive insights of large amounts of data from different sources. Think about AMI or outage systems and workforce management.

So it is a journey, right? Most utility companies, they are just starting. Some of them, they are well ahead of the curve, right? They are positioned to build a scalable platform to support not only the traditional machine learning, but also more advanced AI use cases, such as generative AI that's very popular nowadays, right?

Peter Warren (02:26)
And we see that happening in the Gen. AI and things like that. seems to be a lot of people dabbling in the problem or the possibility. And I guess there's two approaches. One is I've got a great toy. What can I use it for? And the other one is more practical. I've got a problem. How do I solve it? What's your viewpoint on how to approach these issues?

Doug Leal (02:49)
That's definitely a great question because as technologists, sometimes we focus on the technologists that are focusing on the problem at hand. And we definitely need to be focusing on the problem at hand, but not only tackle that problem in a way that is not scalable, right? We need to be able to build a model or any type of AI solution in a way that will enable

the corporation or the organization to scale that, not only scale that model for production, but also for everything that comes after that model, how the innovation that comes after that model. Because usually what happens is one business unit, they see the value of that technology or the value of that model and they say, okay, let's invest more. But if you do not have a operational process to take that model from POC or proof of value, if you will,

to production, it becomes a challenge. It becomes a bottleneck. I have an amazing statistics here that only 53 % of the models actually make from POC, or approval value, to production. So it is a very expensive experiment to only have roughly half of those models to make it to production and deliver

Peter Warren (04:14)
Yeah, I know. think that's a big point. One of the clients that I'm working with right here is heading into the whole area of that. You know, their theory is they really don't want to do POC, the proof of value, definitely, but jump right to minimal viable products, which is what you were talking about. How can I start with something that I can, I have already got the plans and envisionment to take to scale? In other words, it's compatible with my organization. It's not built on something I don't support. My IT team can grow it. My business can get value out of it.

And that's really, think what people are doing is it's less of the playing that maybe we saw even six months ago or even three months ago, much more of how do I take this idea into production? Do you see that as a trend as

Doug Leal (04:55)
Yes, absolutely. It is definitely a trend where anyone that has that proposition, has that use case, it is stepping back and trying to find the real value. It should be spend time and money of our resources investing on this use case. And what is the value behind it? What is the return of our investment behind it? Because it is not...

It is not a cheap experiment when you add not only the infrastructure, but also all the data scientists, the machine learning engineers that you have behind that model being produced. It becomes a very expensive effort to just be an experimentation. So yes, that's definitely a trend that I have been seeing with several of our clients.

Peter Warren (05:49)
So let's dive into the how. So what tools are people looking at? I mean, you've talked about in our prep here about lake houses and data mesh. I think I said that right. I'd like to learn from you as well. How do you see these new tools actually being applied in reality to accelerate success?

Doug Leal (06:12)
Absolutely, absolutely. So let's start with the Data Lake house. So the Data Lake house is a design pattern that combines both the Data Lake and the Data Warehouse. So here's what I mean by that. So the Data Lake provides the flexibility to ingest various different types of data. And what I mean by that is it could be a tabular data, structure data. Think about a relational database. It could be a semi -structured data.

chase zone document or a PDF document, if you will, or unstructured data, images, audio, video. So the data lake enables us to store different types of data set while leveraging the scalability of cloud storage. I don't need to order new hard drives or storage for my data center. I'm using the cloud that's easily scalable.

On the other hand, we have the data warehouse, which is a predictable performance for my business intelligence and reporting. So we combine those two different patterns. We have the lake house, which now, it feeds or support different business organizations.

regardless of that skill set. What I mean by that is, as we ingest the data to the Lakehouse, this data is curated. We apply data quality. We curate this data at different levels. So we have some layers like raw, curated, and aggregated. So on the aggregated layer of this Lakehouse, we have something that is very close to a golden record, where the data is ready for business units with very minimal technical skill sets using a data visualization tool like

Power BI or Tableau, they can just go create their own report, answer their own question. On the other hand, in the raw layer of the Lakehouse, we have the entire data set, which is better suited for data scientists. So it is a scalable platform that allows different teams to consume the data using different tools, but it is a unified.

platform, right? All your data from different systems, think about a utility company, you have outage data, have AMI data, all of this data in one scalable platform feeding the entire organization.

Peter Warren (08:45)
I think that's it. were talking with another client last week and it's about making the data pertinent to that audience and I think that's one of things you just hit it on there. So what about the other strategy, the data mesh?

Doug Leal (08:57)
Yeah, so the data mesh is a social technical approach to share, access, manage, and drive insights out of analytics data. I mention social technical because it is a methodology. There's not a tool out there that will solve your problem and you can implement a data mesh. It is a methodology and therefore, principles.

in Data Mesh, it is a little bit different than what we've been used to because the first principle of Data Mesh is to decentralize your data platform. So usually, we have one big platform, the enterprise data warehouse or the enterprise lake house, if you will, that serves the entire organization. Well, according to the Data Mesh, that's a big problem or a big challenge, per se, to solve with one data platform.

One key principle is to decentralize, not to create data silos, but decentralize your data platforms. Let's think about a utility company here at a very high level. We have the generation, we have the distribution, and the customer. Let's just limit it to three key actors here, if you will, for lack of a better term. So we build a data lake house, a domain -based data lake house for generation.

And the generation team is the owner of that lake house now. We transfer the ownership from the IT team to the business. The business now is responsible for data quality. The business now is responsible for driving insights out of that data. They are close to the data, right? They know the data better than the IT folks. So there is a value there. It's not for everyone, of course, because as you imagine, it is a big

cultural shift, right? Now you are taking the data platform out of IT and giving the keys to the kingdom to the business, right? But it's not for everyone, right? Some clients, this model aligns better than others. And we can talk a little bit about why is that, right? What operating model better aligns with data mesh. But that's one aspect.

So IT still have its role with Data Mesh, which is to build a enterprise data governance, which is very important, and also to support the business with infrastructure as service. So one of the biggest benefit is, just to close it out here on Data Mesh, is business agility. So now that we are transferring this Data Mesh platform, sorry,

domain -based data platform to the business. Now the business has quick access to the data. They can move faster without depending on IT or any other teams to deliver their insights. As you imagine, you need to have a very well established data governance process in place for a data matching implementation to be successful.

Peter Warren (12:12)
Well, that's brilliant. Well, I think that's a good place to pause on this episode. We'll pick up the key point you just brought up a minute ago, organizational structure in part two of this. So with that, Doug, thank you very much for your participation and we'll pick you up in the second part. Thanks everyone. See you then.