2025 Franz Edelman Award finalist: Amazon

Ashley K:

Welcome to the latest in our special series of Resoundingly Human podcasts, highlighting the finalist teams for the 2025 Franz Edelman Award. The Nobel Prize of Analytics, which will be awarded at the upcoming 2025 informs Analytics plus conference in Indianapolis this April. These finalist projects represent teams from around the world who have leveraged advanced analytics to transform their organizations, address their most significant challenges and better serve their customers and communities. Today I’m joined by members of the team representing Amazon to discuss their finalist project in the lead up to the Franz Edelman competition. So to start, can you briefly introduce your team and the project you’re being recognized for in this year’s Franz Edelman Award?

Nick McCabe:

Yeah, hi Ashley, and thank you for having us. My name is Nick McCabe. I lead Amazon’s a network design planning and engineering team. My team was responsible for the implementation of the regionalization project. We’re here to talk about that required us to work across many teams from planning and operations to software and execution teams, but it was really the close partnership with Amab here and his science org that really got this initiative going.

Amitabh Sinha:

Thanks Nick, and thanks for joining. Ashley. My name is Amitabha. I led a lot of the science and operations research modeling work that went into regionalization in close partnership with Nick, and as you mentioned, a bunch of other teams including software finance, operations, et cetera. The project we are here for is called Regionalization, and what it’s about is Amazon’s continued focus on our customers to offering large selection, low prices and fast delivery speeds. And if you go back to Amazon’s history and you think about how we’ve fulfilled customer orders so far it’s been well before regionalization. It’s been a national fulfillment model. And what that means is that if a customer anywhere goes to our website and orders an item, we have the entire network of fulfillment centers across the country from where we decide how to fulfill the customer’s orders. And as we’ve grown over the past several years, grown in customer demand, grown in our network size in terms of the number of fulfillment centers, last mile notes and so on.

What we found was that that national fulfillment model was not working the way we wanted it to work. We were not getting the speeds and the cost structures that we wanted out of that. Instead we were seeing this economy, the scale coming from the increasing complexity of the number of connections between fulfillment centers and last mile nodes leading to our customers. And that’s what made us really want to radically rethink how we wanted to structure this fulfillment network. And that’s what regionalization is and that’s what we’ll talk more about today as well as in the conference.

Ashley K:

So you’ve obviously already touched on this, but I’d love to hear more about what inspired your team to take on this particular challenge and was there a big aha moment that led to your solution?

Amitabh Sinha:

Yeah, so the inspiration for this was again, Amazon’s continued customer obsession on selection, low prices and fast speeds and the fact that the large national network was not working. And to address this, we built a whole suite of network optimization, network design tools, leveraging math techniques, analytics techniques including optimization, simulation, machine learning, AI statistics, a whole bunch of things. And we tested a lot of different network design models through this suite of tools. And the aha moment came when we realized that this regional structure, I’ll tell you in a minute what this regional structure is, this regional structure, which the regional structure was what allowed us to get to a place a better equilibrium with faster speeds and lower cost structure. So what is this regional structure? The idea is that we are going to break up the United States into a number of regions and then as much as possible try to fulfill each customer’s orders from facilities within their region.

They would still have access to the entire national network, but as much as possible they’re going to try to fulfill from within the customer’s region. And if you think from an optimization perspective, we are actually taking the national network and constraining it, we are adding constraints that say that to a large extent, orders should be fulfilled from within the customer’s region. And from an optimization perspective, typically when you constrain the system, you’re going to get a less optimal solution and the more you remove constraints, the expectation is that you’ll find a more optimal solution. And for us, that was the aha moment that actually adding these constraints actually breaking up the network was getting us to a better equilibrium. And we’ll talk more about this in the Edelman competition, how this actually helps us get to a better equilibrium in a dynamical system. But that was the aha moment and then that was just the initial hypothesis that this is an idea that could work.

It still took a lot more work to then come up with the actual operating model. So for instance, we had to figure out how many regions, right, we ended up with eight, but should it be six, should it be 10, should it be 15? That required a bunch of analytics. What does the connectivity structure inside the region look like? We wanted inside the region connectivity to be really fast. Well, what does that look like? What does the connectivity across regions look like? Right? We still have connectivity to the full national network, but the priority is inside the region, but what does the cross region connectivity look like? So we did a lot of work fleshing all of that out. January, 2023, we launched a pilot in two regions. Nick will talk much more about the pilot and the execution. Where I want to go with this is a small personal story.

So my high school daughter had her spring dance on Saturday of this last week through the high schooler. We spent the whole week in a sort of miniature fashion parade at home where she tried out a whole bunch of different dresses. Thursday evening she finally picked a dress and now it was like, Hey, now we need to get shoes. And this was a Cinderella dress, so she needed Cinderella shoes. I’m not kidding. I have pictures on my cell phone, right? So she needed these Cinderella shoes and where she going to get them Thursday night dances on Saturday night. So she looked at Amazon and lo and behold, she was able to order get them home on Saturday afternoon. And this is just a personal story about my family, but this is the level of customer delight that we want to enable for all of our customers through, again, large selection, low prices, fast delivery speeds.

Ashley K:

That’s a lovely story. Amazon really saved the day.

Amitabh Sinha:

Indeed.

Ashley K:

What are some of the biggest hurdles your team faced when tackling this project and how did you overcome them?

Nick McCabe:

Yeah, there’s really two that come to mind. The first one was really just the sheer magnitude operational magnitude of deploying this change. It’s worth noting this change touched almost every aspect of our network from software to operations. So it required us to work across many teams from transportation, logistics, truck scheduling, planning even into the fulfillment network. And then second was how do we deploy this change while maintaining the high service levels that we have across our network, given the interconnectivity that am atop talked about in the national network. So that really led us to the pilot structure and where we can design a pilot focused on a small part of the network, two regions in this case where we can monitor the performance, monitor our systems, monitor the key KPIs that we’re looking, and then ensure that we’re at a point where we can proceed with the rest of the pilot. When we turned it on in January, we were amazed at how fast our systems started to adapt and react to that regional structure. We were so impressed that it actually led us to pulling forward our national deployment plans actually deploying to the remaining six regions or the rest of the country one month faster than what we had originally planned going nationwide by March, 2023.

Ashley K:

If you could sum up the journey of your project in one word or phrase, what would it be and why?

Nick McCabe:

I think transformational is probably the best word. As I mentioned earlier, this wasn’t just a small change. It really impacted every aspect of the way we think about network design, network planning, our fulfillment and software systems just really building from the ground up the way we think about customer fulfillment within Amazon and really the performance of the change in the KPIs really speak to the transformative nature of this project. Once we launched, we saw in region fulfillment improve from 62 to 76% we’ve seen customer fulfillment distance. So travel distance of a package improved by 10% since deploying this. And really what am atop talked about is we continue to set and then break all time speed records for inventory selection and availability next day or same day for our customers.

Ashley K:

In your opinion, how do you see analytics continuing to evolve in your industry over the next five to 10 years and how do you plan to stay ahead of the curve?

Amitabh Sinha:

The amount of data that we are able to collect nowadays is just going to keep increasing. It is continually increasing and it’s going to keep increasing, right? And what that does is it enables the opportunity for a lot more advanced analytics to come into play at Amazon. We don’t want to be just keeping up with the curve. We want to be setting the bar ourselves and investing in research and algorithms ourselves. So there’s a lot of work we are doing in that space. Everything from generative AI to optimization algorithms, network design algorithms and so on. Also thinking about things like, well, not thinking about but looking into things like new chips, like graphical processing units, using them for optimization and how we can speed up optimizations using that. Amazon is a large physical system in terms of fulfilling inventory, fulfilling packages to customers, and we still are in the journey of understanding how that large physical system works.

And a lot of the work that we are doing leveraging data and analytics is really understanding that system and then making it work better to touch upon a few specific things that we are thinking about. One is personalization and predictive modeling in our journey to get customers meet customers needs and intent as delightfully as possible. It comes down to being able to anticipate what our customer is going to be looking for, what is their intent and positioning inventory in the right place so that when the order comes in, we can fulfill it as quickly as possible. Another one is automated decision making. When you think about Amazon and you think about our scale and our complexity and the fact that customer orders need to be satisfied fast and the whole thing needs to work in real time, decisions have to be largely automated. And this comes down to minute microlevel decisions like the specific route chosen for a specific package, but also aggregate network level decisions such as a weather event is coming and how do we want to rewire that network.

And we do that with humans in the loop to various degrees and various levels, but we want to empower human decision makers with automated decision making to drive this machine in the right direction. What it really comes down to in the end of it all is really people, how do we enable people to grow in their analytical capabilities and try new things and experiment and fail and keep learning and so on. So we need to be able to, and we do this consciously, make the time for people to learn, make the time through things like hackathons and conferences and research presentations to try new things and keep up with what’s happening and develop new things on our own. And I’m talking not just about scientists in my team, I’m also talking about supply chain managers and operators and people in our warehouses to be able to do these things. In fact, I’ve been coming to the INFORMS annual meeting for over 20 years, the analytics conference as well, a few times less frequently, but a few times. And we come there not just to talk about our work, but to actually learn from others as well, to learn from our industry peers as well as all the academic presentations.

Ashley K:

What’s one thing you wish more people understood about the role of analytics in solving real world problems?

Nick McCabe:

Yeah, I think the one thing I wish people understood is analytics isn’t just about gathering data or creating metrics, it’s really about building confidence in your decision making. That led us to really three key findings and approaches as part of this project. First analytics helped show us that we had inefficiencies in our network and there was this new opportunity, so it surfaced those inefficiencies as our network has scaled over the years a second, the ability to simulate and predict gave us some initial insights into how the network would perform as we started to make these changes. Then finally, and probably the biggest benefit here is it gave us the confidence in making this decision to proceed forward. So analytics really empowered us and gave us that comfort in proceeding forward with this initiative, even though we knew it would be extremely disruptive to our network. So I would say that confidence in decision-making that analytics empowers is really the piece that I hope people take away from this,

Amitabh Sinha:

And I love that answer, and I’m going to play off one specific word that Nick used, which is the confidence, right? So from my perspective, when we think about, we talked about how much data that is available and how much we are advancing our algorithmic capabilities, the temptation is to build analytical solutions that are extremely complex, right? That we have modeled every single feature and we have found the absolutely optimal solution. But if you do that, what you end up with is a black box that no one understands, and that sometimes behaves in unpredictable ways, and that does not lead to confidence. And you also don’t want to go to the other extreme where you have something that’s extremely simple, but it’s obviously missing the important real world features that are needed when this model hits the ground. And so finding that right balance between simplicity and complexity where you’re capturing the right features, but you are still in a place where you can enable business leaders to have that confidence and test the model and see how it behaves and everything makes sense. That’s one of the keys to making real world impact successful. And to do that, we have to have close partnership between science and business, right? It’s not just between me and Nick, but between our teams, right? My team and Nick’s team, and then our leaders as well. And that close partnership and making sure that the analytics work streams are tied to the real world and the business work streams, and using that to find the right balance in terms of how you build these analytical models. That’s the key to driving real transformational business impact.

Ashley K:

Thank you both so much for taking the time to share this special insight into your finalist project. I wish you the very best of luck in the Franz Edelman competition and look forward to meeting you and the rest of your team in Indianapolis for the 2025 informs Analytics plus conference.

Amitabh Sinha:

Thank you very much for having us. We are looking forward to this as well. Lovely combination.

Nick McCabe:

Thank you, Ashley.

Ashley K:

If you’d like to learn more about today’s episode and guest, visit resoundingly human.com and check out our show notes. The podcast is also available for streaming and download on Amazon Music, apple Podcasts, Google Podcasts, and Spotify. Wherever you listen, please be sure to leave a five star review to help others find and enjoy the podcast. Until next time, I’m Ashley Kay and this is Resoundingly Human.