The Quiet Revolution: Why Azul's AI strategy wasn't about AI at all

Transcript

Greg Buckner:So we are the prerequisite AI talk, but it's not going to be about hype. We're going to be talking about real-world examples that Azul has implemented that have created tremendous value for them. So very grounded, very simple. My goal here is to really show how easy it is to get started and also what sort of an advantage you can accrue by beginning to do these sorts of things. You know, this is— Azul has been able to create millions in additional weekly revenue by implementing AI intelligence on top of their existing systems. So this is without replacing their infrastructure or their stack, the way that they organize things. It's simple, it's efficient, and it's driving tremendous results for them. And, um, I'm very excited about this because I find, you know, we'll be talking about network planning, we'll be talking about revenue management, we'll be talking about predictive maintenance, we'll be talking about all of these things. And I find this really fascinating because I just think airlines are like this perfect version of economics and capitalism, the way that these markets work and the way that you can manage these systems. So with me, I have André Americo from Azul. Very excited to jump into it. Do you want to give a little bit of your background?

Andre Americo:Yeah, I have been in the airline industry for more than 20 years now, 15 of them being in Azul, always leading revenue management and planning teams. And I'm very excited to be doing this transformation that we have been doing.

Greg Buckner:So, obviously, a lot of time in the space. You know, we started on this journey a few years ago. Can you kind of describe a little bit about where Azul was at before you started bringing AI into the system and seeing these returns?

Andre Americo:Yeah, I think the most important point to note is that we were not struggling. We had our systems in place. We had a good revenue management system. We had a good pricing system. We have good network systems, but we are doing something additional and incremental. We wanted to step up a notch higher and be a little bit above what we were already doing. And through many different use cases and many different technologies, let's say, we were able to implement it and give our systems better weapons to fight and to do their jobs. So that's— that was quite amazing and very exciting to do.

Greg Buckner:What was kind of the inkling that you had that AI was going to be able to create these sorts of results? If things were going well, you had modern systems, a good team behind it, etc.

Andre Americo:We do have modern systems and good tools, but the world is changing faster than any system can capture. And working with the AI tools and having another layer of intelligence on top of what we already have was definitely a game changer for for this transformation. So I think the pace at what demand is changing, seasonality is changing, the whole demand profile is changing. I think the whole world is seeing now much more premium demand showing up and becoming a higher share of their revenues, and no system can capture it immediately. So we need— we needed more data to back us up and to help us in that process.

Greg Buckner:Yeah, you and I spoke a little bit about this. AI is enabling you to learn faster than the market is changing, and that's really what it's about, is this intelligence layer creating speed within the organization.

Andre Americo:That's correct. Yeah, we, we are not just implementing AI, but we are deepening our learning capabilities and possibilities with the ecosystem that we already have. I think that's key because we have learned so much more about our own data and our own behaviors, our own underlying behaviors that were driving both revenue management and network decisions. So I think this was a big part of the whole process itself, not to mention the revenues, which were in excess of $3 to $4 million a week right now at this point.

Greg Buckner:And it didn't start with, hey, we need to rip out our RMS system, we need to start from scratch. It was just building these things simply on top of what existed already, building adaptive pricing, building flight-level profit forecasting, predictive maintenance, natural language analytics. All these things are just built on top of the existing systems, and it's taking small steps to get to big wins and being experimental. You don't need to boil the ocean. Yeah, you can just take these small steps. So, So I want to start by talking a little bit about the adaptive pricing model.

Andre Americo:Sure.

Greg Buckner:Which is kind of where this journey started. And it is kind of the most, you know, as a finance and economics geek, I think it's kind of extremely interesting what you did. So you're running multiple models, correct?

Andre Americo:Correct.

Greg Buckner:Yep. Okay. It's a little bit of kind of this like internal Darwinism, competition between these models. Can you talk a little bit about where that started and then how it scaled?

Andre Americo:Yeah, we started very small with very quick wins. I think once in the very beginning we had 3 models running and we have up to 16 of them running in parallel right now. All of them, each one of them handling a different problem that we were seeing after all the optimizations were done. So each one of them handling a different chunk of revenue. And 15 out of the 16 models that we have running right now, they are all priced— they're all pricing upwards, not downwards. So we are creating, on top of better decisions, we are creating a healthier industry, fair environment as well. We started small, small steps. And I think the beauty of it is that no systems were replaced. No big implementation, no big bang had to be done. We were implementing on top of stuff that we already have and trying and learning along the way. Anything that works, we kept investing more time and resources in that together. And anything that we were looking for that didn't really show as very promising, we just keep it aside for a later moment. So small steps and continuously improving. And very quick wins that add up.

Greg Buckner:And I know you even pivoted in that initial first experiment. You were pursuing one thing, but it turned out the data showed a different place, which is where kind of those first models came from. What, what was that? Where did you start and how did you pivot?

Andre Americo:Yeah, I think we started with the more obvious problems like flights that are too full, demand is not showing up for flights that are too empty. And eventually we realized that we should be focusing more on the inputs that we were providing to other systems. So I have a great example that I love, which is the fare class concentration. The fares were optimized, the inventory was optimized, but we were still seeing a lot of bookings happening on the same fare class. Does that mean that demand could be potentially upselling? We discovered later on that yes, the answer is yes. Therefore, we created more fare classes to cover the intermediate points of demand. So we were— I think this is the most interesting part about learning that I talked before. Learning is important because it shows us what's not in the data as well. Not only the story with the data, but also what's not being told Now let's put some numbers behind this.

Greg Buckner:How soon after you began this process did you start seeing significant revenue returns and uplift?

Andre Americo:I think the first ones were really fast. I think between 6 and 8 weeks we were already up and running and ramping up. So it took us a little bit of time to set up the environments, access to data, and— but then after all the data was in your hands, it was pretty, pretty fast. Every new use case takes 2 to 3 weeks to implement at most.

Greg Buckner:And you were seeing a revenue uplift within, I think it was either 3 or 6 months, of like $2 million a week, correct?

Andre Americo:Yep, correct. Yes.

Greg Buckner:And where is that now?

Andre Americo:That's 3 to 4 on the revenue side, plus $2 million on the network side by forecasting future P&L and taking network actions in advance.

Greg Buckner:Okay, and we'll talk more about the revenue planning in a second, but $3 to $4 million in weekly additional revenue.

Andre Americo:That's correct, yes.

Greg Buckner:From implementing these models?

Andre Americo:Yep.

Greg Buckner:Okay, that is obviously very impressive, and seeing results within 6 to 8 weeks.

Andre Americo:Yep, that's correct.

Greg Buckner:So let's talk a little bit about— you just made reference to this— but the network planning piece, which was the second area that we started focusing on. Where was Azul in that process before implementing AI?

Andre Americo:Okay, so P&L figures on a per-flight basis, they take a very long time to show up in the data. So we are in December now. I'll probably— every planning team is probably looking at October, on a very good scenario, November P&L figures. That's not enough because I'm planning ahead. I'm planning March, April, May schedules already. So waiting for the data to tell me the story about the P&L for the past 2 months or 3 months does not really help me looking into the future. That's when we came up with this idea of forecasting revenues, costs, of course, that's the easiest part, but elasticities as well. So what is the impact of adding a new flight in my network on all the flights that are already selling and scheduled? Or on the opposite side, if I am to do a schedule reduction, What is the expected impact on everybody on all the other flights? And we have been able to make better decisions in advance. And the most interesting part is we are making non-obvious decisions. Every schedule planner knows which markets work in a given season and which do not. But we are doing more non-obvious stuff like connecting dots that we never really did and doing I think one very counterintuitive move that we did was adding business frequencies in months with leisure seasonality. Why did that happen? Because we did not have enough time to sell those flights. And the models captured that I'm better off selling business markets in leisure months because of the, of the number of days out that I had to sell. I did not have much time to sell them, and I was better off adding these counterintuitive business frequencies in leisure months rather than doing the obvious move of adding leisure markets once again.

Greg Buckner:There was also a non-obvious finding around kind of routes that looked like they would only be marginal until you factored in ancillaries and connections, correct?

Andre Americo:That's correct. We're taking the whole spectrum of both revenues and costs into consideration. as well as aircraft utilization. So it works like a real accounting P&L for the past, but looking into the future.

Greg Buckner:So across both the revenue management piece and the network planning piece, where are the humans in this system? What does that look like? Because you had a strong team, you layered on AI. What is that kind of collaboration?

Andre Americo:I think the AI helped us most in the avoid firefighting, having our analysts spend thousands of hours in work that doesn't add a penny to our revenues. They have more time to strategize and to think, and in the end, the final decision always comes from the specialists of their area. So the planning team gets the recommendations and they have the final word into saying Does that make sense or not? Let's try it or not. So we add intelligence to the process, but we keep the final judgment. This is still ours. We're still in control. So that's beautiful. It opens up a spectrum of possibilities for us, of new options. But the final decision is the analysts or the specialists.

Greg Buckner:So if anybody's worried about AI in terms of, you know, replacing human capital, in this case, you're taking a strong team and you're actually giving them More time to think strategically, and more insights that they just would not have been able to generate on their own.

Andre Americo:More time and more tools, more resources, more insights, and more learning. Yes.

Greg Buckner:And that is what allows you to move faster than the market and to take advantage of these new technologies as things continue to— volatility increases, as things continue to change in the market.

Andre Americo:Yes. It enables us to capture movements, capture the market dynamics much earlier than what we did beforehand. So we are anticipating more changes in behavior, both on the demand or on the macroeconomic side, as well as taking into account the restrictions of our own environment, fleet maintenance, I don't know, everything else, airports, airport constraints. So we're, we're, I think we're doing a really good job together into taking all the variables into account while looking forward in a very dynamic and changing environment.

Greg Buckner:So what do you think, before we move on to predictive maintenance, across revenue management and network planning, what would you say to like this room are the just 2 interesting nuggets that others should begin thinking about?

Andre Americo:I think we're getting to a point where We are taking— we are making the best global decisions among revenue and planning. So instead of just optimizing the revenue for a specific flight, we think beforehand, aren't we better off by adding more capacity or changing some of the capacity we already have instead of keep optimizing for revenue? So the best decision is the best global decision among revenue and planning. And I think that's one of my favorites. Because if the teams are working separately and independently, sometimes you have your revenue management team optimizing a flight like crazy, trying to squeeze every penny from it, and then the planner comes and adds new capacity to it. So he could have sold a little bit slower to get a better global result, but that was not the case. So I think we're getting to a point where we can do the best decision globally and in a timely manner.

Greg Buckner:I mean, the best possible decisions you can make globally obviously create huge, huge value.

Andre Americo:Yeah. The less siloed we work, the better off we are. And I think we have seen this in the projects that we are running, where we see the dots connecting among customer data and bookings data, bookings data and network planning decisions. We started seeing those cross-relationships among different departments. And the good thing is I have the same company running all those projects. And so it's so much easier to link the dots, to connect the dots down the road than if we were doing it ourselves in every— in each department independently. We don't need to have AI analysts in each department. We need an intelligence that is looking at the whole picture, that's seeing the big picture.

Greg Buckner:Well, it's been very fun for us to be a part of that. What's— so before moving on, on the network planning side, what's the financial impact been?

Andre Americo:For the future P&L predictions, $2 million a week.

Greg Buckner:Another just handful of million dollars in the bucket. Yeah.

Andre Americo:You know what's up.

Greg Buckner:I think this predictive maintenance story is very, very interesting. Everybody knows maintenance logs are incredibly chaotic. You've got handwritten notes, ATA codes, inconsistent formats, et cetera. So we started working on unifying everything. And I want to get to the funny story in a second, but talk a little bit about what the predictive maintenance journey looked like.

Andre Americo:So we were fighting with having a lot of aircraft on the ground because of many different issues. And we were not able to say what were the causes of the next AOGs. So we got together and we tried to look at the data that we had and predict based on the last failures that we had, what are going to be the next failures for those aircraft. And if we can plan in advance or do something in advance, it turns out that we could influence both the supply chain and the, and the maintenance schedules to account for that. And we, we end up avoiding many of the AOG and increasing aircraft utilization and availability.

Greg Buckner:And this was just by modeling past performance and then understanding what was going to probably— what are the leading indicators of aircraft being grounded, et cetera?

Andre Americo:Yeah. As a good example, we've seen that when certain aircraft had a specific failure in a specific component, the next component that would see a failure was another one. And I think I'm coming to the funny part of the story. But I placed a bet with the maintenance and my team that an aircraft that had this failure was supposed to be AOG in the next 2 weeks or so. And I wrote down the tail of that aircraft on the glass, and it turns out that it happened. So by trusting the data, I was able to see 2 to 3 weeks in advance an AOG that happened.

Greg Buckner:And the maintenance team did not believe this before it happened.

Andre Americo:Yeah, we had to trust the data, but Nobody believed me at first. It turns out that I was right. We were right, actually.

Greg Buckner:And with this, you've been able to predict like 150 events per month.

Andre Americo:Yes.

Greg Buckner:And prevent them.

Andre Americo:Yes. Or at least be better prepared, like dedicate more teams and more hours, resources, or parts where I need them, where the aircraft are going to be. We could do better planning because we had a better forecast.

Greg Buckner:Okay. And now this is the— this 4th example I want to talk through. This is kind of the most AI-y because everything we've talked about before is really around machine learning. But this is a combination of, you know, machine learning, but importantly, LLMs, the kind of newest AI wave, natural language analytics. So the funny story that you told me is that you originally built this just for your analysts.

Andre Americo:Yeah.

Greg Buckner:to use it, but now everybody in the company is talking to AI.

Andre Americo:Yeah.

Greg Buckner:What does that look like? How did it start?

Andre Americo:And I was very clear that it was a pilot that we were testing, and I gave it to the president of the company, and he was using it himself like crazy and asking questions. The beauty of it is that everybody gets consistent answers for anything they, they might think of. So we were able to ensure better information flowing across different departments. And I told him, if you want, you can put my face in there and you can ask any questions you want. It will feel more, more human, but you have all the answers you can think of. And it helps a lot us in not losing time putting together queries or asking analysts obvious and straightforward questions.

Greg Buckner:A lot of enterprises have actually failed to implement these sorts of solutions. What are some of the things that allowed you to succeed such that everybody can actually use this and trust it?

Andre Americo:I think having the data, all the data that we could get in one place and give you access, and you were already testing cases with the data, so you knew your way around. I think that helped a lot in connecting the dots and realizing which questions would be relevant or would be made. And I think that it transforms language into queries in the end. But you did a good job in translating what was going to be asked into usable terms.

Greg Buckner:And there's kind of like a templating system involved, correct? And like flagging questions that are very likely to be correct, those can be exposed to executives. Ones that may have some uncertainty or haven't been verified yet, Those kind of go into like a yellow zone that only analysts can use, etc. So you create this robust system by merging human with AI.

Andre Americo:Yes.

Greg Buckner:And creating this availability and creating kind of a culture of learning in the organization.

Andre Americo:And also for the questions that were asked multiple, multiple times, we created different use cases for that. So everybody's asking about this. So why bother? Why even bother asking? Why not show up to everybody who needs this information? beforehand instead of waiting for them to ask those questions. We— I just realized that I like contests because I put together a contest in my department of who would come up with the best question to the system. And it turns out we got really good ones.

Greg Buckner:What, just as a concrete example, what are some of the questions that are getting asked very frequently and kind of are driving the most value for, you know, executives?

Andre Americo:Yeah, I think when the data has to be very segmented, for example, how many bookings in Y class I had in the past 2 weeks that were applying a specific promo code type, or how many of those were connecting passengers, how many of those have a lower than average yield for this exact AP. So that kind of question became very easily answered now that we have this model in place.

Greg Buckner:And the system also actually guides the user. So if they ask a question that doesn't make sense, it'll actually try to predict, hey, did you mean this?

Andre Americo:Yes.

Greg Buckner:So that people can ask questions and they will actually get coached through what is the right question to ask to get accurate data.

Andre Americo:Yeah, not to take anything for granted. They do another why. They ask back a question to the user. Just to clarify and make sure he's asking— he's going to answer— it's going to answer what's being asked and not something else.

Greg Buckner:What's been the impact of this system within the organization?

Andre Americo:I think the biggest impact was in the communication and having the information flow on all levels— C-level, directors, analysts— and better information leads to better decisions.

Greg Buckner:I mean, it sounds at the end of the day, you know, this was— we call this talk a quiet revolution because this was not, you know, tipping over, ripping out the machines, starting from scratch. This was small experiments getting applied in different parts of the business that snowballed into creating substantial value. But with all of your existing systems, with all of your existing people.

Andre Americo:Yes.

Greg Buckner:Without changes, but just harnessing this intelligence layer to learn faster and move Faster.

Andre Americo:Yes. I think the biggest value is this additional layer. It's monitoring and watching all our systems. So we have the same network, the same aircraft, the same systems in place, but it's looking at each piece of this ecosystem and providing better solutions or better inputs. So I think that's where the value is. And I think that's what brought us to this incremental revenue.

Greg Buckner:Is there anything that you're working on? I'm not sure if you can share or not, but anything you're working on now that's coming down the pipeline that you're excited about?

Andre Americo:I think this connection between revenue management and planning and making the best decisions, this is my favorite one because I happen to manage both teams and I have never seen anything like that. I have seen teams working separately, And the whole source of problems for revenue management teams are planning decisions and vice versa. So I envision us getting to a point where they will live in harmony and the best decisions overall will be made.

Greg Buckner:And this is obviously incredibly important because we all know that volatility is not going to decrease. Revenue management system upgrade cycles are not going to change. We're going to continue to have more and more uncertainty, and that is where AI applied on top of these systems creates tremendous competitive advantage.

Andre Americo:Yeah, that's right. And I think volatility will always be present. We won't be able to predict the future, but we will have hints of what is about to happen so that we can plan better and prepare in advance.

Greg Buckner:So I think people in this room can maybe ask themselves, you know, where are Where are their analysts fighting more fires instead of steering and having that strategic direction? What decisions are being made slowly because maybe you're waiting for better data or you're drowning in the data that you have now and you don't have the team to do that? How can AI actually be leveraged for that? And importantly, because we have the entire time started small, made, you know, you've said across all of these, it's 6 to 8 weeks to proving value and then scaling those things. What are experiments that could be done next month instead of next quarter or the quarter after that?

Andre Americo:Yes.

Greg Buckner:Because these things compound. And for Azul, they have compounded tremendously, giving you an advantage that others don't have.

Andre Americo:Yes, that's correct. And without having to undergo a full system implementation that can take anywhere from 1 year to 2 years to have it fully up and running. So being able to capture those low-hanging fruits and quick wins was decisive for our success. And I think the ability to learn more. As we learn more, we think about more and new use cases that, like you said, they compound. So having started this journey 2 years ago has definitely been good as a starting advantage in this, with this new technology.

Greg Buckner:And I'm sure it doesn't hurt at the end of the day that we're talking about $6 million in incremental revenue weekly, 150 cancellations prevented, and everybody in the team getting to access all of this data and accelerating the learning across the organization.

Andre Americo:Yes, the numbers speak for themselves. So great.

Greg Buckner:Thank you, guys.

Andre Americo:Thank you.

Greg Buckner:Thank you, André.

Copyright policy: All transcripts on this site are the copyright of CAPA - Centre for Aviation. Our reproduction policy is as follows: you may quote up to 400 words of any transcript on the condition that you attribute the transcript to CAPA - Centre for Aviation and link to the original video page. All other use is prohibited. While we aim for 100% accuracy in the transcript, there may be some minor transcribing errors.