Stories of AI breakthroughs in business
There’s a lot off buzz around Artificial Intelligence and the opportunities it brings. We are interested in what AI use cases are already in production and how they deliver value to companies. The AI Change Makers by GAIn podcast tells the stories of business leaders who create industry breakthroughs and transform their companies using AI – and how they do it.
#1 Heineken – Building AI & data capabilities to brew better beer
With the large amount of available data at Heineken, the appetite for doing more with AI and machine learning has been growing. In this episode, we hear from Senior Analytics Consultant Sandra Oudshoff and Management Reporting Manager Jasper van Panhuis how AI and data analytics are being applied at Heineken and what value it brings to their organization. Heineken is harvesting their data-driven journey. The goal is to use their data to make improvements across various business domains, ranging from the brewing of beer to delivering it to the consumer – all whilst being more sustainable. Heineken uses the GAIn AI Translator Program to empower this journey and also to identify new growth opportunities. A big thank you to Sandra and Jasper for sharing their stories with us in this episode of the AI Change Makers.
Episode transcript: #1 Heineken – Building AI & data capabilities to brew better beer
Brought to you by GAIn, this is the AI Change Makers Podcast. My name is Wouter Huygen and on this show, I talk to business leaders about how they create industry breakthroughs with AI.
On today’s episode, I have the pleasure to talk to Sandra Oudshoff and Jasper van Panhuis who are both from Heineken and who share their experience in building Data and AI capabilities at a global scale.
Sandra, Jasper, welcome to this first episode of the AI Change Makers podcast. Today, we have our first two guests who are both from Heineken. Welcome!
Let us start with a brief introduction so we can let our listeners in on your background. Sandra would you like to start to tell a bit about your background?
Sandra: “I am Sandra Oudshoff and I have been in Data analytics all of my working life. I studied computer science and graduated in artificial intelligence. After this I worked at a major telecommunication company in many different roles for a long time. For the last 3 years I have applied those skills at Heineken. Basically, I was Heinekens’ first analytical translator. At that time, it was called a functional consultant. My role was to connect the data science experts, the data engineering experts, and the business experts to make sure we bring value out of our analytics experiments at Heineken.”
Jasper what is your background?
Jasper: “I am Jasper van Panhuis. I have not worked with analytics all of my working life, but I have worked with analytics and numbers. I used to work at energy companies and before joining Heineken I did a lot of predictive modelling on unbilled revenue and did analyses to see where customers would be more profitable. Within Heineken, I am responsible for all the managerial reporting to region and group. As such, I am working with a lot of data and I think our DNA is quite comparable to Sandra’s team when it comes to the data engineering part. Of course, we have a different background, which is more functional. I believe that this is where we connected in the experiment, we did a year ago.”
So today we will talk about the Analytics Translator initiative that Heineken has started, and the global training program related to it. Before we dive in, Sandra, could you please share a bit of the background of how the role of Advanced Analytics or AI has evolved at Heineken over the past couple of years?
Sandra: “We started about three years ago with a very small team as a center of excellence to help Heineken in its data driven journey. We started by experimenting in many different business domains within Heineken: brewing of beer, procuring of raw goods, and selling beer to consumers and business customers. We tried to show where there is value in analytics for Heineken by using machine learning technologies. This has created an appetite for Heineken to do more with AI and machine learning. Heineken is seriously on a data driven journey, and this has increased because we recently appointed a Chief Digital and Technology Officer at Heineken. So, for the first time we have someone from IT in our executive team. IT is no longer just a supportive function, but it is central to Heinekens’ business. This has been recognized by having a digital technology function within Heineken. We just had a reorganization affirming that. The mission now is basically to use all the data Heineken has, to brew better, to get the beer to our customers better, and to sell better. Also, to work on our sustainability initiatives, because one of our focus areas has been to create a sustainable brewery. Analytics can definitely help with that by having fewer truck rides for example or using less raw products to brew beer.”
A couple of years ago, you launched a global program for training analytics translators. What was the specific background to that initiative? What led to it?
Sandra: “I think we recognized early on that to be successful with machine learning, it is not enough to have a central team of excellent data scientists and engineers who can make all this analytical magic happen. You need to connect to the business and understand business processes to bring real value with these machine learning technologies. The translator plays a crucial role there. Like McKinsey said: you need ten times more translators in your company than you need data scientists. It helps Heineken to take advantage of these technologies if those translators are not just part of a central team but if they are scattered across the company, across all functional areas, all business domains, and all operating companies where we are active. To get there, we needed to basically help people understand this world of analytical magic. Therefore, we created the Analytics Translator program.”
Could you explain what the program looks like?
Sandra: “It is a quite intensive four-day program where we bring a group of 15 to 20 people together. We take them through the major areas of machine learning so that they understand the machine learning process and what steps you go through. From defining your idea in collecting data to bringing models into production. We talk about the different descriptive and predictive analytical techniques so that they have a solid background in understanding the whole world of analytics that is available. But we also talk about relevant areas like the data domain: what do you need in terms of a data architecture? And we discuss the change management that is involved in getting to real business value. Because when you apply machine learning to business domains, it usually means that people must change their way of working a bit or a great bit. For this reason, change management is a crucial aspect, and translator play an important role.”
At what point do the translators typically get involved? Are they the initiators or are they brought in when the initiative has already started?
Sandra: “It can be both, but what we see now after a year of experience with the Analytics Translator program, is that more and more new analytics ideas are brought forward by people that followed this program. Going through this training, they understand the value of analytics, and they also understand better how to apply it in their own domain. From there they come forward with these analytics use cases as Jasper did.”
So, Jasper, you have been one of the participants?
Jasper: “Yes, a year ago I was lucky enough to join. And I was so enthusiastic, that in the end my entire team participated. It was an insightful course. There was homework upfront and during the course you could not lean back and listen to a few stories, you constantly had to work, bring your own use cases, and refine those along the way. All in all, it was a very good program.”
What were some of the biggest learnings or insights for you personally?
Jasper: “I am in finance and we create a lot of graphs and bar charts, so I thought this is my bread and butter and I know what I am doing. But there was one exercise where we had to map a graph relevant to a question. There I learned about when to show lines, when to show bars, when something is descriptive or predictive. I found this very insightful. Also, it helped me to see all the steps that you need to take when doing statistical work. I used to do that in university about 20 years ago, but never actively used it anymore. Those theoretical parts in combination with the visuals and the examples that Sandra gave won the business over. I think the relation between what an algorithm predicted, and the reality was an eye opener. This got me thinking as well about what we could do with it.”
And did that lead to a concrete application?
Jasper: “It did not start as a concrete thing. In finance, we do a lot of forecasting since the beginning of time, for instance, budgeting or annual plans. That puts a lot of pressure on everyone involved in finance: the reporting team, my team that checks the submission of the Opco’s, the business control team that uses the outcome to challenge the business and do the business partnering. We were wondering whether we could do this any faster or smarter. I made a huge effort to present my idea in the best way possible and entered a competition in which you could win two weeks of a data scientist. I really wanted to win it, and luckily enough I got the opportunity. We started in an agile way of working and as we went, we refined our goal and scope, because to be fair, it was not clear at the beginning. We wanted to forecast revenue, but how we went about it, that changed a lot in the course of two weeks.”
Has that initiative further matured?
Jasper: “It has. After the first instances, we were able to do a fairly accurate prediction of the revenue. We shared the outcome, but it was a bit early in the process. There was interest, and there was a lot of show and tell in the lunch sessions. But how to organize proper funding to take it from an experiment further, I did not know. In the end, I am a middleman, business control is my customer. The middleman is usually not the person initiating these things. But, like Sandra said, Heineken is on this journey of digital transformation. At one point, I think we got the right story and opportunity, we pitched it and we were granted 7 sprints to do implementation. At this moment we are in the middle of that. It is extra exciting, because the predictability was high, but then Covid-19 came. Due to this, the predictability got severely impacted, but the relevance became even bigger, and our finance processes changed. From once a quarter estimating for the next year, to a rolling forecast every month ahead, so, we know better where we are going in the short term. With all the volatility, the accuracy per Opco is not always very high: sometimes it is, sometimes it is not. So it that case, a second opinion from an algorithm could help to challenge or to confirm certain expectations.”
At what kind of a scale is this going to be applied in the near term?
Jasper: “How it is going to be applied exactly, we do not know yet. The scope is known: all Opco’s. We use historical revenue data; from the system my team works with in which all the financial results and planning is submitted. So, in this case the data is already harmonized, for all the Opco’s. Future usage could be central, like a scenario planning or a second opinion. We need to refine that with our customer. Some Opco’s who are less mature could also use it as a second opinion for their local planning. How exactly is difficult to assess from my position. I know that we lack details: we just have revenue. The substantiation behind it, for example which brand of beer (Heineken, Amstel), is something we do not know and that is the starting point for the Opco’s. So that triggers a debate: how can we strengthen each other, as it is not a replacement, that is for sure: it should be an addition.”
It is also an area that can further develop. Are you planning to start with this as a first application and do further iterations on it?
Jasper: “That is what we hope, Sandra did not mention it, but some of the challenges we have is that data in Heineken is not harmonized in every single area. It is of course in finance because otherwise we could not consolidate our results. So it could be possible to go from revenue to margin, but it is much more difficult if you want to go into details about for instance different bottle sizes, brands or sales channels.”
Basically, you state that some other AI application areas might be more difficult to scale?
Jasper: “That is the challenge yes. We have groups of Opco’s with comparable systems on many different functional domains, but it is not like one tool is applied everywhere. So that is why we embarked on this journey for digital transformation, to have that harmonized. It is possible, but it takes time.”
Sandra: “In the transformation that we are in now, to the digital function, there are two key areas: not only analytics but also data management. We need to have more data harmonization, data standards, and data quality across Heineken, and we need the central organization to help facilitate that.”
Sandra, could you give an example of a typical AI application for a brewery? What kind of examples come out of the brainstorms in this program?
Sandra: “We have had many different ideas. The one I like best is my first project at Heineken which was about improving the beer colour of Heineken. We went to the Zoeterwoude brewery which is our largest brewery. There we looked at the beer brewing process, we focused on the quality of the beer and one of the key quality parameters is the colour. I never thought about this before I joined Heineken, but Heineken beer has to be a very specific range of yellow.”
Sandra: “We want it to be “heerlijk, helder, Heineken”.”
Why is the color not constant to begin with? I would say you have factory running and it pops out the same kind of recipe and color every time, but that is not the case?
Sandra: “That would be nice. It would be nice if you could apply total quality management and procedures and then you would get the same beer every time. But beer brewing is not an exact process, it is more an art than a science. For this reason, we also have master brewers at Heineken who have a very important role. Beer color can vary from brew to brew due to changes in the process, changes in the ingredients, and other factors that have impact on the beer color. We sat down with the people that are experts on this topic: the process technologists and the brewing operators. We learned about the brewing process, what kind of information is collected during the process and what impacts the beer color. From that we came up with an idea on how we could use machine learning to help improve the beer color during the beer brewing process. We developed a machine learning model that looks back at what happened at the past 10 brews, it looks at the ingredients that go into the next brew and then it predicts the color of the next brew. It then advices on how much color malt to add to the current brew to get to that perfect Heineken yellow.”
Before this solution was in place, how were the colors of the brew changed if they resulted not to be on spec?
Sandra: “They used to have a morning session where they looked at graphs from the brewing data. Then they would inspect whether the beer color was rising or declining and whether it was still inside the boundaries of the quality parameters. If it was too close to the upper or lower boundary, something was changed in the brewing process. It was a decision that was made once a day, and then they waited for a few days to see what effect the change had, and they would steer again if necessary.”
Did that add additional costs to the process?
Sandra: “Definitely. If the beer is not of the right color it had to be mixed in the cellars of the brewery with beer of another color to get it right before it could be sold to consumers.”
Knowing the color upfront helps you prevent having to make those changes at the end?
Sandra: “Yes. Around 7% of all brews had to be remixed which takes up production capacity that could not be used for other purposes.”
This was one of the first experiments you did. What did this change in the business?
Sandra: “I think it changed a lot of things. As a global team, we had a better understanding of the beer brewing process and everything that goes with it. In the brewery, after the experiment, they had a better understanding of how beer color is impacted. They had a lot of assumptions about the cause and effect, and we were able to substantiate some of these assumptions from the data and on others we saw different results. We helped them understand the brewing process even better. Also, it changed a lot in the way that the brewery was looking at analytics. When we started this process, one of the process technologists said that he did not believe this would result into anything, yet his boss put him on this project. By the end of the 4 months that we worked on the beer color project, he was a big fan of this new analytical way of working and these new machine learning technologies and what effect it could have on the brewery.”
You mentioned that the insights that came out of this initiative improved the knowledge of the team and provided additional insights into how the process works. I can imagine that if you have such an algorithm, it would be worthwhile implementing it and having it operational as an additional tool of measurement in the production process. Did you do that?
Sandra: “Yes, and we are still working on it. When we tested this in real life in the brewery, we found that the beer color of the brews was 30% closer to the perfect Heineken yellow. That is a big improvement. We started discussing how we could make this not only a predictive but also a prescriptive application. In our experimental setup, the prediction of the color was shown to the brewing operator and he manually changed the recipe. But in a more ideal situation you would use the machine learning model to change the recipe automatically. Of course, with a monitor to check that whenever something changes in the environment and a model is no longer valid you will have safeguards in place. But it takes a long time to take something like this in production in a brewery because you need to take a lot into account. Not only IT, but also operational considerations. It is always a challenge when you move from the experiment to implementation to connect the product to actual IT systems. The experiment might take 4 months and the implementation might take a year or even longer. But we had an added challenge because we are dealing with the operational side of the Heineken business, so you must take OT (operational technology) into account and you have to deal with the added security that comes with it. Added to that there have been several system changes in the brewery in Zoeterwoude, and that is why for the last year it has been almost in production, but it is not really live yet.”
You mentioned change management earlier, I can imagine that a brewing operator now suddenly gets presented an algorithm that starts to advise them what they should do. How do they respond to it? To what extend is there an adoption barrier and what can you do to overcome it?
Sandra: “In my experience, people respond very differently. And their reaction also depends on their background and the kind of person they are. We had one brewing operator involved in the program that was very enthusiastic from the start. Basically, he convinced all his colleagues on the brewing floor to start working with this. We always involved the end users, the people that are going to use the results of the model, from the start of the project. From the minute we start, they are at the table discussing what the idea is and how we could use machine learning to improve the business processes.”
Jasper, is that different in finance, because people are more used to numbers?
Jasper: Yes, they are used to numbers, but they are also not embracing an algorithm to replace their work. And I do not think they should. I think it should be an addition. For instance, in my work, if there is an acquisition or disposal, an algorithm does not know about it. It will discover it over time if you acquire a business, because the revenue increases. At first the algorithm is off, then it detects the change, and it starts being accurate again. A human should always interpret what comes out of the algorithm and use it as an addition. That is what I hope in the future use, and that is the direction we are heading into. That we can combine algorithms with human knowledge and make it better. But you need to have a certain background. Now, we are all familiar with data, but some of us are nerdier than others. And people need to experience it, like in the example of Sandra. If someone who is skeptical sees the outcomes and that they are accurate, maybe this person might be persuaded that they can learn from it. That usually is the best way to get support. And you need to find out the application together. Like in my example, we are not going to replace how Opco’s do their production planning, forecasting, and stock management. I do not think we are even going to replace how they are doing their rolling forecast. But if we are in a central position and we need to make sense of the consolidated forecast, if the forecast is not what was expected from it, it would make sense to get a second opinion for the overlay that you want to do centrally. That is something we need to discuss with Business Control, to see how they see use for it and then we can adapt to that in our developments. In my opinion, you should first create a first version of a product, and then you get the thinking going. Because upfront it is difficult to formulate what a customer wants or needs. For instance, in the car business, people did not know they wanted APS or Airbags, but when it was available you wanted that option in your car. This is a great example of the dynamic between technology and the customer to refine the product.”
You mentioned the interplay and collaboration between algorithms and humans. As the number of algorithms increase within Heineken in different areas reaching a certain level of scale, how would that change people’s roles and how people work?
Jasper: “Data is more and more important. Not just in finance but in many roles we need to be data savvy. In the training I did, there were a lot of commercial guys, I did not expect that. They also work with a lot of data, which I did not realize. Just like we have English as our business language, data should also be one of the languages we speak. Of course, not everyone has to be an analytics translator, but if your role is really dominated by data, I think you cannot survive without certain translator skills. It is the business knowledge on the one hand, but also data knowledge on the other hand that creates value. We need to evolve here.”
That goes broader than perhaps the people that are intensely involved in developing these new solutions. It also concerns the people that eventually are the main end users I suspect. Or is it something that a central team could develop and ship?
Jasper: “I think we are learning that now. We are co-developing with Opco’s and with regions and we are finding out that it is certainly not something you can only do centrally. You need the feedback and refinement from users, partially because the relevance of the product immediately improves, but also to gain a bit of trust and faith in what is predicted. If you are only shipping the finished product, you are missing the point of why people would use it. However, if people are part of the development and see that an algorithm can predict more accurately, they can start thinking about how to use it.”
Sandra, do you have another example where earlier in the experimentation process business impact was demonstrated?
Sandra: “A good case example of a case where we demonstrated business impact is predicting churn. Churn refers to customers who stop ordering from us in the future. It is important because if we have a significant percentage of customers leaving us every year, sales needs to put in a lot of effort to acquire new customers to replace them. It is much better to try and retain the current customers that are already delivering value. That means that you need to know which customers to focus on. The idea behind this use case was that if we could find early warning signals in customer behaviour that tell us that this customer is thinking about leaving us, then we can feed that information to our sales managers, who can get in touch with their customers. They might be able to find out what is going on and what we can do to retain them. Developing a model was not an easy task because we needed data from different data sources and good knowledge of the business process. We did this in two iterations with the UK, in an agile way of working involving sales managers from the beginning of the project. In finance, people might have an analytical background, but it is different when working with commerce, because people do not have the same analytical background. In this case, you have more explaining to do about how this kind of modelling works, and why exactly we need them to be involved, and to help them understand what the value could be for their daily work. Like Jasper said, with these models it is not about replacing people, which is what people sometimes fear when you talk about AI, but it is about helping them to do a better job and reaching their goals. If there is one thing that sales managers trigger on, it is to reach their objectives. We could work together to help them achieving their goals in an easier way because they got relevant information from the model.
What are some of the surprising insights that the model came up with?
Sandra: “When we first started talking with the sales managers, they had some conceptions about what was impacting customer churn. They thought that rural customers are more loyal than urban customer because in a city there is more competition and pressure. That was one of the things that we found that was not true. We experience the same level of churn across rural and urban areas. We dove into a number of parameters to try and explore these relationships, and this helped us to better understand the domain and it helped the sales manager to better understand what we were trying to achieve.”
Were there translators involved from the commercial side that also went to the training and already got used to this new way of working and developing AI solutions?
Sandra: “No, because there were none in the UK at that time, but we were lucky to be working with a product owner from the commerce domain who was very open to these new technologies and helped in a translator kind of role to convince sales teams to work with us. It also helped that we had two sales managers in our project team, because they could tell their colleagues what was going on and how this was going to help them.”
How do you both see the role of AI evolving within the company and what kind of changes will that bring?
Sandra: “We will see AI in more business domains and more business processes than we see it now and in a more mature way. In the past few years, we were mostly experimenting and now we are at the stage where more of these successful experiments are being transformed into actual products, analytical solutions, that can be used by multiple operating companies within Heineken. The ultimate goal we are working towards is that we get a Heineken app store with all analytical solutions. And if on the data side we achieve sufficient harmonization and standardization, we could take such an analytical solution from the app store, plug it in to your own operating company, and immediately see the benefits from having an improved business process. It will never be this ideal world, of course, but we are taking big steps to get there.”
If you would give one piece of advice to other companies on the same journey, with respect to training their staff and people to accelerate the adoption of AI, what would that be?
Jasper: “For one, if you see this as your core or it being critical to your success, make sure to work with internal people, because of the knowledge you build during this journey. This is not like buying a book. Also, having harmonized data is one of the most important things, including good availability, security, and governance. Lastly, enthusiasm at the top, for instance we have our IT board member on this topic that really helps, because to get it moving in the beginning, you need support.”
Sandra, what about you. If you think about an approach for reskilling in the era of AI, what are your main learnings?
Sandra: “Jasper stressed the most important points, the importance of data and harmonization, the importance of management support at all levels. And indeed, what we need in the end is that basically everyone at Heineken becomes data savvy and that they know enough about this digital world to know when and where to apply it in their daily work and that they have tools and data at their fingertips to make it happen. So, we need a major upskilling effort across Heineken to make people aware of these tools and technologies out there and how it can help them.”
As a final thought, what might the consumer in the future notice from all these changes.
Sandra: “They will only see relevant ads when they go to the web, they will get Heineken beers that are of perfect quality and at a lower cost. In the end the consumers will notice it in our contribution to a world we can all live in, to do what we do in a sustainable way.”
Thank you, Jasper and Sandra. Super interesting to hear what a global beverage company is doing with AI and thanks a lot of sharing your learnings, experiences, and visions.
Sandra: Thanks for inviting us!