They reflect on what Southern Company has learned from putting AI into practice, from improving data quality and defining success measures to balancing innovation with governance. They also look ahead to conversational data access, multimodal AI and digital twins as utilities build on these foundations.
For Joyce, one of the biggest lessons has been the importance of investing in data quality early. Utilities increasingly rely on data to inform operational decisions, making its accuracy critical to the value AI can deliver.
Looking back, she also sees an opportunity to establish the conditions for success earlier in the journey.
Southern Company is now applying AI and machine learning to improve the process itself, including using synthetic data to fill gaps in meter data that previously required manual intervention.
Doug points to another lesson: build for reuse from the beginning Investing early in infrastructure as code, automation and repeatable deployment frameworks can make it faster and easier to deploy new AI use cases as adoption grows.
Not every AI use case delivers as quickly as initially expected. Some require more time for fine-tuning, prompt engineering and iteration before they achieve the intended outcome. What can initially look like failure may instead reflect overly optimistic expectations about how quickly a use case can move from concept to measurable value.
As AI adoption expands, balancing the speed of innovation with governance and trust becomes increasingly important.
The key is to separate experimentation from authority. Teams need room to test ideas quickly in controlled environments, while production controls increase with the consequences of an AI system’s decisions.
A tool summarizing internal documents, for example, carries a different level of risk from an AI system influencing customer service, field operations or safety.
Joyce sees education and collaboration as equally important. With sensitive AMI data, helping employees understand why guardrails exist—and how they enable trusted use of the data—has become an ongoing part of adoption and change management.
Southern Company is also changing how employees interact with its AMI data.
As part of a broader modernization effort, the company is rebuilding a long-standing reporting experience used by nearly 900 employees. The new self-service platform incorporates Databricks Genie models and a knowledge assistant, allowing users to ask questions directly of governed data and learn more about AMI concepts and the information available to them.
Instead of relying on a new report or dashboard for every question, employees can interact with data conversationally. Human involvement and user feedback provide a way to continue tuning the agent and improving its responses.
For Joyce, this creates an opportunity not only to make information easier to access, but also to help employees build their knowledge as they work.
The next opportunity is to expand the types of utility data AI can work with.
Doug sees significant potential in multimodal AI that brings together tabular data with engineering drawings, geographic information system data and other operational information. This could give utility teams new ways to interact with complex information without requiring every answer to be built into a traditional report or dashboard.
Joyce is also looking toward digital twins that combine data, analytics and AI to simulate future grid conditions and support more prescriptive decision-making.
Realizing that potential will still depend on trusted data, reusable technology, appropriate governance and close collaboration between business and technology teams.
For Doug, one of the most important lessons from the journey is that AI value can compound. Investments in trusted data, reusable patterns and governance make individual use cases possible, but they can also make the next use case easier and faster to deliver.
Joyce sees that evolution ultimately serving a broader utility objective: bringing data, AI and people together to improve reliability while supporting sustainable affordability for customers.
As AI capabilities evolve, these foundations can support a more intelligent energy ecosystem, with people continue shaping how the technology is applied, governed and improved.
- 1. What utilities can learn from scaling enterprise AI
-
Peter Warren:
Hello everyone and welcome to part two in a series that we're doing with our friends at Southern Company, specifically Joyce Solomon. I'll ask Joyce to introduce herself and Doug Leal from our AI and data side. Joyce, how are you? Say hello to everyone.
Joyce Solomon:
Hi everyone. It's good to be back again, Peter. Thank you so much. I am the Data Analytics and Data Science Manager in the AMI space at Southern Company. Happy to be here.
Doug Leal:
Absolutely. I'm glad to be here as well. I'm Doug Leal with CGI, and I help our clients turn their data and emerging machine learning and AI models into measurable business outcomes. I'm really looking forward to the conversation today for us to explore lessons learned, what is working, what is not working and where the next chapter of our AI will take us.
Peter Warren:
That's a great point. We had fun in the first talk, and that spawned us to do this one. A little time has passed, Joyce. You've been building on a lot of things. You've learned a lot about your systems, you're using your AMI data, you're using AI data, you're connecting things. But to Doug's point, let's talk a little bit about what you might do differently and maybe what didn't work at all. What would you have changed if you had a chance to start all over again?
Joyce Solomon:
I think one of the things that we really underestimated was the effort around the data quality piece and getting the use cases operationalized. With some early use cases and the excitement around them, that was not a perspective that we looked at.
But if I could do everything all over again, I would spend more time up front defining what the success metrics are and making sure the data quality is in place because we are building insights today for operational decisions. We want the right decisions to be made based on the insights that we provide, and also to establish business ownership. Those were the things that I sit today and think through. I wish I did better in these areas.
Peter Warren:
And Doug, did you see something similar from your side?
Doug Leal:
Yes, I would say I completely agree with data quality. Data quality is always a gotcha. You think that you can trust the data, and then once you see it, it's, "Oh no, I need to do some type of data cleansing here. I need to work to enrich this data."
But I will also add the reusable framework for deployment. As we think about infrastructure as code and moving things faster, innovating faster, being able to press a button to deploy the solution to higher environments, make the solution available to business users faster and get that feedback faster. The whole concept of fail fast and learn even faster.
That would be something that we would invest in. We would definitely repeat what we did, which is invest in infrastructure as code and invest in automation on day zero, but we would double down on that to automate it as much as possible in the beginning because that will pay off long term.
- 2. How AI can improve data quality for utility operations
-
Peter Warren:
Hindsight is always 20/20, so this is an interesting exercise. Let's talk about data. Did you use AI to help cleanse your data? How did you get your data up to speed?
Joyce Solomon:
Yes. Today we are using AI to help cleanse our data so the quality becomes better. As a matter of fact, I am running a separate program to fill meter data that has gaps with synthetic data using AI and data science with machine learning models.
Maybe three years ago we were manually inputting it so that we could get a good, solid data set, but that took a long time. I'm excited to see how AI can help accelerate our journey on a project. I'm super excited about this new program.
Peter Warren:
That's very cool. Do you think you could have spent time up front to get that closer and more accurate, as Doug was talking about? Or is that something you really had to work through? As you go to another department, do you think you've got a faster, better way of doing it or maybe coaching somebody externally on that?
Joyce Solomon:
Yes, absolutely. I think we figured out that you always have to crawl, walk and run. Today, if we were to think about that topic, it would be walk and run for us because we have figured out what's the easier method to use and how we can use AI to help us move forward as well. So yes, definitely, I think that's the way we want to go.
Peter Warren:
And maybe just to open it up to Doug, were there any parts that just absolutely didn't work? Are there things that we thought were going to be gold, but just didn't produce what you thought they were going to?
Doug Leal:
It's a good question. I cannot think of anything.
Peter Warren:
Okay.
Doug Leal:
It might take longer for us to implement. I think there are a few things that took longer than we initially thought, but in the end, either through fine-tuning or prompt engineering, whatever technique we leveraged, we were able to reach the goal, to get to the metric that we wanted.
Peter Warren:
So it wasn't really a failure then. It was maybe being too optimistic. Joyce, anything to add to that? Do you see anything where you think, "I wouldn't go back there again. That's something I wouldn't want to do"?
Joyce Solomon:
Yes. Trying to take meter data and fill it one by one is what I don't want to go back to doing. But that is the most important data, if you think about it for a utility. That is the data that connects us to billing. So we've got to have the highest-quality, accurate data for billing.
If there's one thing I don't want to do, it's mess with that data because when you build something on top of it, that insight is going to be totally wrong. That's what I'm very cautious about. I don't want to go and impute it one by one, but we're trying to use data science and AI to do that imputation with a lot of business rules in place so that we can be successful.
Hopefully, if we ever have our next series, I can tell you whether I failed at it miserably or actually succeeded.
Peter Warren:
I look forward to that. That'll be fun, and I do hope you do well. As Doug said, you may just have to wrestle it to the ground, but we'll get there.
- 3. How AI governance can support innovation and adoption
-
Peter Warren:
Looking at this, we talked in the first episode about the human factor, how people were working and how, Doug, I think you mentioned change management kind of happened organically because as the departments wanted the outcome, they adapted. Has that continued? Are people adapting and changing the way they're working?
Doug Leal:
Yes, I believe so. And I believe governance plays an important part in it. There was always that conflict of balancing the speed of innovation with governance and trust. For us, we don't see speed and governance as opposites.
I think the key is to separate experimentation from authority. Things should be able to test ideas quickly in a controlled environment, and that helps with adoption because you give that freedom and people are comfortable innovating. Most of us love to innovate. We love to test different things.
But the controls required for production increase with the consequences of those decisions. A tool summarizing internal documents does not carry the same AI risk as an AI system influencing, for example, customer service or field safety.
Having that balance between the speed of innovation and governance to move things to production and to the end user also helps with AI adoption.
Peter Warren:
And Joyce, do you see your folks behaving differently than they did before this process got going? Maybe some silos don't exist anymore, maybe people are teaming up differently.
Joyce Solomon:
Yes, I definitely see us moving to the advantages of this, Peter, but I won't say that it happened overnight. Going back to what Doug said, everybody wants to innovate and everybody wants to win quickly.
From our perspective, AMI data is sensitive data, so we put guardrails in the form of governance around it. It did take us some time, and I'm still in that journey. I'm not saying that I've successfully completed it. I'm still in that organic adaptation and change management journey to educate folks that governance is put in place to help them instead of being a roadblock to using our data and being successful.
That's definitely a journey of collaboration, training, understanding and teaching. In order for us to help them with adoption, we have to educate them on the journey that we are taking. It's definitely a lot of collaboration and partnership that has helped us move that needle of change management. I'm committed to continuing that journey because I don't think it's going to stop today or stopped yesterday. It's a continuous journey.
- 4. Why leadership and culture matter when scaling AI
-
Peter Warren:
You both mentioned time. Some of these things took longer and the change was slower. At any point has your executive sponsorship waned? What was their viewpoint? Were they expecting something immediate? Are they seeing the benefits and outcomes from their perspective?
Joyce Solomon:
I report up a stream of leadership who loves innovation. Quick to achieve, quick to recognize failure, and a lot of ideas have always been cultivated.
I feel very safe in this culture where they let you fail quickly and they also want you to be successful. There have been a lot of programs and projects that we have completed successfully and we continue to evolve to bring to a successful closure.
The billing system that I was talking about earlier took four years. It is a great celebration that we hope to have together because the leadership has leaned in, they have helped and they have talked about how we are doing this journey, especially with AI incorporated.
Peter Warren:
It's great to see culture in action, isn't it, when it's all swimming in the right direction?
- 5. How multimodal AI can unlock more utility data
-
Peter Warren:
Following up on that same idea, where do you see this going in the future? AI has evolved since you started. Doug, Genie wasn't really a thing when you started with the Databricks platform. That's an evolution, but where do you see it going from a technology standpoint? Then Joyce, we can come back from a business standpoint.
Doug Leal:
I believe multimodal approaches, where you can leverage solutions that will not only work on tabular data. You mentioned Databricks Genie, which is a great tool. It is used by a wide range of organizations that are on the Databricks platform.
But going to the next step, which is this multimodal approach, where it's not only tabular data but also engineering drawings and GIS, I think there is untapped potential to leverage different types of models that work very well on different types of non-tabular data to drive insights out of it.
For listeners who are not familiar with Databricks Genie, it is an AI assistant that allows you to ask questions of your data instead of developing a report or dashboard. That's a very simplistic way to describe it.
It works very well with tabular data, but expanding that to other types of data sets, I think that's the next untapped potential that we have.
- 6. How conversational AI can make utility data more accessible
-
Peter Warren:
And Joyce, do you see the same thing? We talk about headless operations too, where you're not typing into a fixed tablet, but people are having conversations in the field with data and doing things. Do you see that in the future for Southern Company? Are you going to be talking to your data and interacting with it in a totally different way?
Joyce Solomon:
Maybe not the future, Peter. It's now, and that's why I'm super excited about it.
As part of our billing system go-live, I had the opportunity to look at a dashboard, a reporting tool that had been in place for maybe the past 15 years. I'm working with a project team to rebuild that and get into the latest features, but also incorporate AI Genie into it because we found that we have close to 900 users using that reporting dashboard on a day-to-day basis.
I wanted that dashboard to be converted to a platform where they can come and self-service. I have a super agent running that not only has an ensemble of Genie models that goes to the table when they ask questions, gets the answers and provides them, but I also have the knowledge assistant embedded into it.
If they come to the portal and want to learn about AMI or each of the columns in a table that they have been exposed to, they can ask that question and the knowledge assistant will help them.
For instance, when I came to this team, I didn't know what a momentary outage was. But today, I've enabled new employees who come into the power delivery environment at Southern Company to ask that question of Genie and get their output: What is a momentary outage?
What does that help? Self-enablement. Just like you said, interacting with Genie. There's also a human in the loop. There is also a feedback tool that can continue to tune the agent so that it will be more productive for folks to use. So yes, I'm super excited about that one, Peter.
But of course, it is going live on Labor Day weekend together alongside the Hugues Project.
- 7. How digital twins and AI could support future grid decisions
-
Peter Warren:
That sounds very fun. Sorry your Labor Day weekend is going to be a little busy, but that's what we do in this industry.
If we looked back at the beginning of this conversation and what you would have done differently, looking forward three years and then turning around, where would you expect to be? If you had a chance to jump forward and turn around, what kinds of things would you expect to be doing in three years?
Joyce Solomon:
That's a tough one. I came from the asset management group, so I was very exposed to the new cutting-edge technology then called the digital twin.
I'm hoping to use the techniques, tools and data that we have today to help with a digital twin that could help us simulate future grid conditions and even what sentiment might look like when you engage with the customer on an individual basis.
I think that is the next best thing as prescriptive analytics. We're trying to see how we can roll ourselves into that, Peter, but that's just my ambitious mind thinking through what we can do three years from now.
- 8. Why AI value compounds as utilities strengthen the foundations
-
Peter Warren:
I love that idea. I think that's a great ambition, and we're at time here. I'll give you each a final thought. Doug, just final thoughts on the future and where things are going.
Doug Leal:
Absolutely. What the journey shows is that AI value compounds. Investing in the foundation, the data that makes the use case possible, the reusable patterns that allow us to move faster and the governance that allows us to move things from experimentation to production safely. Those are compounding values.
As we continue to invest in those, I strongly believe that the speed at which we deliver those AI-driven use cases will increase exponentially over time.
Peter Warren:
Good. And Joyce, a final closing thought from you?
Joyce Solomon:
Yes, Peter. I'm very, very [unclear] towards reliability and affordability, so I probably will niche it that way.
I think for us as a utility, the real opportunity isn't just about building models, but creating an intelligent energy ecosystem that involves data, AI and humans in the loop together to not only improve reliability, but bring sustainable affordability for the customers that we serve.
Peter Warren:
Thank you for that. Both of you are doing brilliant work, and it's great to talk to both of you. I wish we could do this on a daily basis. But thanks for making time for us today. And to everyone listening, we'll catch you in episode three. Bye.
