How P&G’s AI Factory is Powering Scalable Digital Transformation - CIO Seth Cohen | Technovation 978

The nice part about an AI factory approach is it's a platform that we can allow people to have instant access to the data within the data repository, but then also instant access to the AI algorithms. Therefore, the developer spends a lot less time having to worry about how to scale it, because that comes out of the box in terms of what that experience is ultimately going to deliver to the organization.

Hey Peter, it was great to see you, and I appreciate the opportunity to talk with you.

Well, it's my pleasure.

Listen, it's a phenomenal company. I would say that the term business, we can talk about the products. There are 11 different categories, ranging from paper towels to diapers with Pampers, to my favorite product, which is the Gillette razor, as you can probably tell, but many, many others. It's an absolutely fantastic business. The thing that I've really grown to appreciate coming to P&G from some of the other companies I've been at is P&G's just maniacal focus on understanding the wants and needs of the consumer, as well as trying to figure out how to address those needs in ways that haven't been done previously. It's such a fun place to work, to really spend that time focusing on what we call our vectors of superiority, to ensure that we have the best product to meet the consumer at every step of the way.

Yeah, so it's a fantastic role, and I could not be more pleased. I'm not just saying that because I'm being recorded; I really do believe that to be the case. P&G is very much a highly matrixed organization, and the matrices go a little bit different than how I've seen it at former companies. It's multi-dimensional in the sense of the categories, which is kind of our baby. The category is largely at the top of the house in terms of really driving the ownership of the performance of the company. But then beneath the category, you have both geographic cuts as well as functional cuts. So this very complex matrix organization forms around it. Now, I would argue, as I'm sure most people are aware, P&G has a very storied background of promotion from within, so people knowing where to go and doing things almost becomes part of the DNA of the company. As a new person coming in, I was a little fearful of that level of complexity in terms of the matrix of it, but I found quite the opposite to be true. In terms of my responsibility, direct accountability, I have accountability for all of P&G's infrastructure, security, and then the data analytics, as well as what we call more of the commercial capabilities into the organization. From a matrix perspective, there are people that sit across all of those different matrices as well. In terms of setting the strategy for technology for the company, I engage with that full organization to ensure that we are driving against our key strategic planks that we want to drive for the company. So it's a bit of a complicated way to describe it, but I would tell you it works, and it works quite well.

Yeah, unlike what I think we see in the press and in the YouTube videos, it doesn't just snap your fingers and it comes alive. There's actually some pretty heavy lifting involved to ensure that you have the ability to have very nimble and accessible capabilities you can drive into the organization. So first and foremost, I would say we make sure that we're leading not with technology, but we're leading with capability. Then from that, we peel that back into what is the needed technology to bring it to life. But if I were to dive a little bit deeper into the technology, it starts in some of the spaces that probably are not the most exciting to talk about but are very important, which is the systems of records, whether it be our ERP solutions or our Salesforce automation solutions that sit within there. Those are where all the transactions are happening. But then what I believe is the secret sauce of any company I've been part of is freeing the data out of those systems of records into a common data repository. That is critical because I think what we're finding more and more is the use cases of the capabilities we're trying to drive don't just end where a function begins and ends; it actually ends across the entire value chain. So if you think about a capability I'm trying to figure out, why might I have out-of-stocks at the shelf in the retail store? Well, the problem or the opportunity could be in the selling organization, it could be in the distribution organization, it could be in the product supply organization, it could be in the raw material purchasing organization. To see all that come together is so critically important. Then I would say the next layer after that data layer is what we call our AI factory. It's probably one of the biggest capability drivers that we have as an organization. The nice part about an AI factory approach is it's not a bunch of one-off answers; it's a platform that we can allow people to have instant access to the data within the data repository, but then also instant access to the AI algorithms, to some of the models. If you think about some of the generative AI work, it's all there. Therefore, as a developer, the developer spends a lot less time having to go through the security stage gating and those sorts of things, spends a lot less time having to worry about how to scale it, because that comes out of the box in terms of what that experience is ultimately going to deliver to the organization.

Yeah, it's a really good question, and it's a really hard answer. I don't know if anyone's ever going to be done with the answer. I have found in former attempts at this, in former days people would say, okay, you build this data repository, people will come to it. I found that that is often not the best answer. Let me explain why. To be blunt, data is hard. This isn't easy. It's hard not necessarily just from a technical perspective, but it's hard from a change management perspective. What do I mean by that? If you're a finance functional person, you have data. You may not be very excited about that data being shown to other functional areas, not because you don't trust them, but rather you've grown up with the DNA that that's my data, I don't want that data to be used for other things. So what we have found, and this is something that we feel is a fairly significant unlock, we start, as I mentioned earlier, with what's the capability we're trying to drive. By starting more from the capability lens with the right partnership with our key stakeholders, if we come up with capabilities that span functional silos, we then can go into the detail of how do I source the data for that. We have the traditional data modalities. We want to make sure data, if you think about the gold, silver, bronze, we use a different naming convention, but it's roughly a gold, silver, bronze layer. When you go from that bronze to silver, what does that mean? It means I've cataloged the data, it means there's metadata attached to it, it means it's discoverable. But if I do this now more from a capability building perspective, I can then ensure I have the right level of access control as well as the right quality of data that I'm looking for. If it's not there, I can work with the stakeholders to make it there to bring that capability to life. The reason why I think this is a really interesting approach is it creates a flywheel. The good news is if we take this approach that I just described, a data lake that then feeds an AI factory, while the first use case might be very specific in an area over there, the second use case comes along. Guess what? All the data from the first use case is still there. So now all you have to do is start to mix them up, and maybe there's some incremental data required. Great. But I'd rather deal with an incremental 5 or 10% to make that next use case come alive than start from ground zero and have to go to 100% immediately with a second use case. So it's a very leveragable flywheel that we see start to develop over time. But Peter, I would be lying to you if I told you this just happens. This is hard work. I think once you do the work and maintain that data quality, maintain that data lineage that I'm describing, watch out. You have a lot of capabilities you can drive into the organization.

I can give a couple of examples. There's one example that internally they get so annoyed with me because I always bring it up, but I just love it because it does such a great job of articulating this idea of removing functional silos. I mentioned earlier that baby or Pampers is a big product for us. The nice part about that product is we have a very active engagement with the purchaser, the moms or the dads or whoever is doing the purchasing of that product, such that we have what we call Pampers Club. Pampers Club is an app, and people are very excited about engaging with the app for a couple of reasons. They love to get the information if they have questions about the sizing of the diaper or which product they should use for their baby, but they're also excited because there's a loyalty program attached to it. What we've been able to do is create this end-to-end flow where we can have a consumer at the consumer unit level. We're talking millions and millions of diapers that go through the lines. The consumer scans it when they buy it to get credit for their loyalty program. We are able to trace that product all the way through the value chain. That product, we know what store it came from, we know how long it took to go from our DC into that store, we also know if there's a problem with that product. If you think about the supply chain, we know what time of day, what microsecond that product hit the glue gun for some gluing thing. So we can quickly diagnose based on a consumer complaint with simply the QR code that that consumer has, that full lineage all the way back. Back to my earlier example about freeing data across, think about what that really means. That's combining product information, POS information, it's providing retailer distribution because retailers will distribute the product within their own domains, it's also our product supply data. We have that full end-to-end visibility, which is just such an amazing unlock for us as we bring this to life. So that's one example. I'll give a second one which I'm equally as excited about. I'll focus on our customer, which is the retailer. We have capabilities now where our salespeople can walk into a store and be able to get all kinds of insights about that store. Is that store what we call planogram, how the shelf is set, is it correct, is that store having the correct products in there, are we out of stock on products, what might we think to do differently? We actually have done some stuff where we can look at how that store is performing versus other stores in the vicinity. As we know, store managers will be a bit competitive. So why is this SKU doing so much better over there at that other store versus this store? We can help work with the managers around what might we learn from those other stores to allow the current store to be even better and more productive in terms of driving turnover of the product. So we have tons and tons of capability now. Similar to the last discussion, that doesn't just happen. There's all kinds of information that comes in around that, whether it be demographic data about what's going on around the store, whether it be POS data, whether it be shipment inflow data or outflow data. That's all coming together to create those fantastic insights.

Peter, it's a brilliant question. One of the things I did when I first joined the company was I sat down with the leaders, both internal to the function but also our stakeholders that we engage with, to understand where might we think about doing things a bit better. As you would absolutely suspect, people is a big part of that agenda. We absolutely have to do it. P&G, similar to most of the companies I've had visibility to, went through a period of time maybe five or six years ago, and actually for the decade before, of this big outsourcing of IT talent. The problem that I think most of us have encountered with that is we kind of turned the keys of the kingdom over to someone else to manage. Now, I'm not suggesting it means there's no room for external people to be engaged; there are. But when we're thinking about some of these critical areas that we're talking about, such as AI, insights capabilities, data science, data engineering, these are capabilities we need to control and own internally. So two things. First, fortunately, my predecessor has been doing an amazing job, similarly to what we were doing at my last couple of companies, of insourcing. So we've gone through a reversal, not across the board. There are definitely tasks and capabilities that we do want to continue to have our external partners help provide, but we want to own and control things like data science. Now, does that mean there are no data scientists that happen outside? No, there is, but we have a fair number of data scientists, a fair number of data engineers, a fair number of people involved in cloud engineering to allow us to really control our destiny. So what we're spending a great deal of time on is we've insourced a lot of roles, but then we're also working on upskilling programs for both the internal IT organization, but also, I would argue, part of that upskilling is for the broader organization. If you have a reluctant organization, a digitally reluctant organization, it's very difficult to introduce new capabilities to that organization if they don't want to do things in a more digital way. So we also spend a great deal of time. We've partnered with Harvard Business School, we've partnered with Boston Consulting Group on broader training for the organization to raise the digital acuity of the entire organization so that we can try to solicit new ideas for capabilities that can drive the most meaningful impact into the organization.

It's a great question, and it's a topical question. Let me start by saying I don't think AI replaces the human. What I think AI will do is help the human be much more productive and much more impactful into the organization. Listen, unfortunately, I'm a bit old. When I started in the workforce, there were no spreadsheets. I remember when the first spreadsheet, I think it was VisiCalc, came into play. The industry was abuzz that there would be no more need for analysts anymore because you now have a spreadsheet to do everything. Well, lo and behold, 30-plus years later, analysts are still around. They're very productive with the spreadsheet. They're driving a lot more capability than they ever could have dreamed of doing if they were doing things with pen and paper. I view this AI revolution similarly. I think it's a tool, and ultimately it's going to be a requirement of our folks to embrace the tool, similarly to that spreadsheet example. Can you imagine now, 30 years later, if you're an analyst and you have no ability to do anything other than pen to paper? You're not going to be able to be at a comparable level to your peers, and you're not going to be at a competitive advantage for the company you're working at. I think AI is the same way. My very strong encouragement for everyone is we really need to embrace this. So if you're not spending time engaging with AI now, make sure you're meeting your company's policies. In case anyone from P&G is listening, do not go to the open platforms; use the internal platforms to do it. But engage with the solution and start to experiment. How could it help you make yourself better at doing your job and more impactful to the organization?

Absolutely, yeah.

Yeah, so I think it goes in many different directions. Being 11 different categories, supply chain is not just one term here, because you have different parts of the supply chain depending on the products you're working on. But it starts with ensuring that you have the most resilient supply chain operation possible. That resiliency is not just is the system up or down, but with things like real-time vision of the production line. As you start to see razor blades being made or paper towels being made, can you start to detect where things become out of tolerance during that very fast process? Our industry is often referred to as fast-moving consumer goods. Many of our products, fast-moving is an understatement; it's blurred. When you see it on the production line, how fast things are actually going, being able to start to catch before a problem becomes a problem is a big idea. That's something that we've been focusing on quite a bit. It gets back to how does that come alive? It's data, vision data, insight, and we're able to drive things. We also spent a great deal of time around distribution, like where's the best place for demand planning, network planning, to ensure we're producing the product where it needs to be produced to reduce the friction to get that product ultimately to the retailer or to the consumer. There's a great example we have in Brazil. Brazil has a very interesting practice. Brazil's customers or the retailers will typically give us an order for the month in advance, and their comment is just deliver the order when you're ready. The problem we had up front, if you think about how we can bring insights to life and how we can improve the resiliency of the supply chain, the problem we were having was we basically were fulfilling orders based on requirements that were more internal focused, like do we have excess of this SKU or was this SKU in a high run? But we weren't really focusing on what were the demand signals that the consumers had. Brazil had probably one of the largest out-of-stock situations that we had in our company. We were fulfilling the orders as they were asked, but the problem was we might have front-loaded a whole bunch of SKUs, and then they sold through the SKUs and the SKUs weren't there anymore, or we might have back-loaded a bunch of SKUs and people came in for the SKU and it wasn't there. Through the use of AI, leveraging all kinds of things such as supply chain planning, knowing our network, what plants we're producing, what weather patterns, time of year patterns, holiday schedules, all kinds of stuff, we were able to create insights into that such that when that bulk order comes in at the beginning of the month, we now have a much better way to articulate which SKU should go to what store at what time. That comes back to feeding our production schedule so we can meet it. If you don't know our industry, the percent I give might not sound like much, but we improved our out-of-stock by 15 percentage points. This is huge. Usually in out-of-stock, we talk about things in like one or two points. This was a 15% improvement. It was amazing. Again, it's because we freed the data up across all that value chain to bring that to life. It's an amazing project. I could not be more proud of the team for delivering it.

Yeah, I'll call it Horizon 1.5 if you use the McKinsey horizons, and one's more probably Horizon 2 to 3. The first is these new reasoning models that are coming into play. What we are finding is that the reasoning models and the new cool kid way to describe it is agentic AI, all of that attached to it. We think there's something there, and we're starting to get very involved with bringing some capabilities to life in this space. What we're finding, to be blunt, though, is the reasoning models or the agentic AI is often not going to be an out-of-the-box situation. There are all kinds of things we need to think through. One example might be, I am a big proponent of why would you ever in the future need a dashboard? Why can't you just talk to the data? That sounds very interesting, but here's the problem. Today, if you were to put an agentic system on top of your data set, it doesn't know your vocabulary. Every company I've been part of has its own vocabulary. So you ask it about something around a specific category and a specific time of the month, it might not necessarily understand exactly what you mean by it, and therefore it's not able to. So having the ability, we're starting this in earnest, having robust ways to give feedback back to the models to allow them to become more and more robust. That's not really agentic AI yet, but that's allowing the reasoning models and the general models to understand that interaction in that context. Once we're able to do that, the next step for us would be, as you probably are very well aware, in former years we would talk about robotic process automation. It was very deterministic, rule-based models that would do things. Once we are able to have a good vocabulary, a good way for the models to really understand the information below, we believe we can move into more automation capabilities, agentic capabilities that would allow us to do things that in the past the rule sets would just be far too big for us to try to bring to life. That's the vision. The opportunity with that vision, though, is we still have to worry about security, we still have to worry about identity and access management. I need to understand from the bot's perspective, whose persona is that bot taking on? When we think about identity and access management, we often think about it in three different layers. There's the human, there's the people that developed the agent, but then there's whose persona is that agent taking on? If that agent is acting on behalf of me, I want to know that that was an agent, not me doing that interaction, but that interaction was done this way through my approval. So that's part of the work that we have going on. That's the nearer end. I think if you fast forward maybe 6 to 12 months, we're going to see some amazing capabilities come alive. We're already working on those. Longer term, I got to think quantum. I think quantum is a big idea for us. Quantum is not going to have all the answers to all the problems, but what we believe quantum is going to have a significant capability for us is in the ability to optimize in ways that we may be limited today. If you think about supply chain optimization, transportation optimization, or financial optimization, the main constraint we have today is time. We have a set number of variables that feed into the engine, and based on the time required, because there's only 24 hours in the day and the models have to be done, we will limit it. I believe as we move into the future, and I think this is a bit further out, whether it be 5 years or 10 years, we'll see, I think that's going to be a huge unlock for companies, not just P&G, but for many companies.

No, absolutely. So in terms of the team, we actually have a very esteemed group of people we nominate every year. It's a two-year rotation that we call the Technology Advisory Board. They work on these very far-looking process areas or technology areas and start to come up with ideas. What we're trying to figure out is, in some spaces, we might have to wait five-plus years, and it's good for us to have a good understanding. The reason why quantum is obviously an important one is quantum-safe encryption is also something we need to start getting our arms around. What does that even mean? How do we bring that to life within the organization? That's something that we've been working on very actively. But also, I would say there's probably more near-term ones as well. We talked about this agentic stuff, the fact that we can now start to talk about these personas in ways and have perspectives of how we can bring it to life. We have a team of people that are now focusing with part of this Technology Advisory Board. For myself, lots of outside-in looking, asking people. The one thing I find fascinating in these spaces is there are so many different topic areas that we could be probing on. To not spend the time to understand it better, I think you're not doing yourself a service. To really have a good understanding of these spaces is critically important. It evolves. My perspective of generative AI has probably evolved 18 times even in the past six months. But it's fascinating to see how the technology evolves.

Don't quote me, I think it's about 20 to 25 people. Maybe a bit more, but it's not a huge group. But they're focusing on these different areas that are really helping to bring all of this to life. It's amazing.

Oh, I love it. I always joke, I walk into the reading and I feel so humbled because this group knows so much more about the topic than I do, but I walk away smarter. So I love it.

Oh, that's a good question. Many, but here's one that I would suggest. If you have time, this is not for the faint of heart because it does take a couple of three-plus hours to go through. There is an individual, Andrej Karpathy. He was a co-founder of OpenAI, he was at Tesla for a while. He's done these very long-formatted YouTube videos on the whole transformer model. I would tell you, if you have an interest to learn about how these models work, even if you're not an IT person, you'll understand. He does an amazing job. The reason why I find it so helpful is he gets into some of the details of how the models get smarter and how they get trained, which actually, as you think about how you can apply it into your organization, I think he definitely opened my eyes to a bunch of things. One of my peers, a CIO, recommended I watch it, and I'll tell you, I think it's brilliant. So I would definitely encourage you to spend some time on his videos.

No, I appreciate your time as well, Peter. This has been fantastic. I look forward to the next time we can connect.