Using Generative AI like a natural: Allnatura’s data-driven approach to customer retention

Hello and good morning to our webinar between hmmh and Emarsys with the topic Using Generative AI Like A Natural. Thank you for joining. We waited a couple of minutes longer to get everyone in the room, but I'm really excited and thank you for everyone to be here and looking forward to a really interactive session and lots of questions hopefully at the end of the presentation. Maybe just quickly to introduce ourselves, my name is Philip. I'm from Emarsys and the Managing Director in Germany, and I am joined today with Amir from hmmh. Amir, maybe introduce yourself. Thank you, Philip, and thank you for having me. I'm Amir. I'm Unit Director with hmmh. We are pioneers and connected commerce. We're an agency and do all things surrounding e-commerce and communication. And I'm very glad to be here today with you, Philip and be able to share some insights regarding AI. Awesome. Really great. And I'm really looking forward to the case that you brought with our customer AllNatura. Before we lead into that, we just wanted to maybe set a bit of scene and talk about, a bit more general about AI, the trends in the industry and after Amir's showcase and the life example, we'll then also give you an exclusive sneak preview on what we have planned on the AI roadmap from Emarsys. So stay tuned and, let's get going. I think everyone who is here, we all know that generative AI is a massive topic right now. But also just if we look at the market opportunity and the business value, it is really a huge, huge potential for everyone to participate. I think especially also, from Germany or Europe, we've seen the recent news from Microsoft, investing heavily also into AI development in Germany. And, if you also look at the widespread adoption, despite the early days, we already see that a lot of organizations are using generative AI already, as well as planning to adopt and invest into that market. So I think generative AI is not just a buzzword and trend. It is there. It is there to stay. And it's now up to everyone to adopt and participate, because everyone who is now leaning in will have also a competitive advantage over their competition and in the market. But before we go into that, obviously everyone that is here is, sort of interested in AI, but we would like to see, how much are you using in your daily business already? So if you can take a couple of seconds and answer one of those options, we would be really appreciated and get started. All right, so most of you have implemented some basic AI. Not what you're using it bar adding it to our tech which I think is good because that's the audience that we aim to have for this webinar. So not people that are completely new and unfamiliar with AI, but also not the experts that they are able to build their large language model all by themselves. So it's an in between and I'm really excited about the case with Amir and the two of us to see how a marketer can actually then, utilize the AI functionality already here. But just also to frame it a bit more, where do we use AI in marketing? And currently I think everyone that works in marketing, e-comm, and retail knows predictive AI. So that's basically what has been state of the art for a long time, especially around product recommendation, system optimization, affinity lifecycle management and all those things. But predictive AI is basically using historic data and then predicting in the future. It's not what we are currently experiencing which is the generative AI. So we're generating content and supporting in the daily routine. And those are the two areas that we are predominantly looking at when we talk about AI in marketing, how can we support the marketer in his creative or her creative work in terms of creating content, creating, images, creating text, and then also be supportive in their daily routine in terms of becoming more efficient and using their time more efficiently through AI so that they can spend this time saved for more complex tasks. And we will also talk about how we are going to do that in the future, as well as then also exploring the next level, which is the conversational AI. So how to interact with an AI to create a new level of user experience and also making it more efficient and effective to communicate both with consumers but also internally, through conversational AI. Just maybe also looking at how we then look at AI use cases, I think it's really important that we're not just there to use AI as a feature but it really needs to be connected to real life use cases and challenges that marketers have. So let's take a very, generic example, where a marketer is tasked to reduce inventory. We have too much on stock. I don't care how you do it but just get this stuff out and sell it. Okay. So right now, if the campaign manager would then need to look at all the tasks that she had need to perform for finding the products that have a high inventory and low sale, creating the campaigns, checking all the channels that need to be connected, localize those campaigns in all the different languages, find the right context to match to those products, because we want to be relevant in the campaigns. Then also look at optimizing response rates so that the content that goes out actually also returns click and purchases. There are a lot of really, really, complex subtasks that a campaign manager need to perform. And we are looking how AI and all those elements can actually support to make her life easier. So from selecting products, writing content, finding the best channel, helping in the language translation, generating launch business, sending campaigns, optimizing that on the fly. Those are all steps where the AI can already help. And the time saved is really tremendously and Amir will then also talk about that in his showcase how much time you can actually save, especially if you're looking at localizing and multi-language, campaigns that you are planning. So, that's just a sneak peek where you would be coming from when we talk about AI. And now let's get to the really exciting part, Amir. Really looking forward for you to present our customer case. So over to you. Thank you Philip. Truly, we are living in such an exciting time with so many opportunities regarding AI. However, it is crucial for us at hmmh not only to pioneer in e-commerce and new technologies, but also to ensure our efforts are sustainable, pragmatic and yield the anticipated business results. This is especially true in the context of AI as a hired topic, where so many over promised right now and set unrealistic expectations. So with our more than 30 active Emarsys clients and being the leading Emarsys partner worldwide, we are always driven to pave the way in navigating the future, particularly in Emarsys, but also beyond. This commitment led us to the development of the current customer case, where we were able to employ AI in a real life environment through a proper proof of concept project. We are so happy that one of our hidden champion clients, AIlnatura, was open to come with us on this journey. And just allow me to give you some quick facts about Allnatura to begin. AIlnatura is synonymous with sustainable, ecological and environmentally friendly home and living products. Over the past several years, they were repeatedly awarded for customer satisfaction and sustainability. At the same time, they were able to register the most significant growth in their 40 year history, just in the last few years alone. And an important contributor to this growth was certainly the marketing with Emarsys. Selina has been in the center of this effort and is convinced that personalization is key to success. We just love how Selina is always open to new ideas and how she values AI as a tool for achieving the next level. So it didn't even take us any real convincing to start into this POC. The main idea was to utilize AI and large language models to enhance recurring campaigns, ensuring clients receive personalized and diverse content always. Our aim was to achieve this ideally with reduced effort or at the very least, with the same effort currently required for creating campaign content. And we were open to experimentation and A/B testing. But above all, we aim to ensure that business outcomes were met and that they ideally even exceeded expectations. Our objectives included significantly reducing editing time, ensuring thousands of different variations, maximizing abandoned cart recovery, and achieving a sensible return on investment. To accomplish this, we needed to set up the campaign and automation, define a content structure, create prompts for each content segment to feed the AI, generate content variations, and produce translations for the multilingual setup. Once we are ready, we invited Selina and her market colleagues to conduct the final editorial review and to optimize the generated content. In this process, we received positive feedback from Selina, who found the generated text to be a time saving source of inspiration. She noted how the versatility made the content more interesting, and she appreciated the ease of producing high quality translations for various markets. Yet, she kind of realized how crucial it was to really review everything to give their final touches through some word repetitions and some other missteps could still be found in the AI results. Now, you might be wondering how we managed. Essentially, we opted for dedicated and specialized tools to organize the content in addition to our beloved marketing platform, Emarsys. Emarsys, of course, served as the foundation of our approach, and by utilizing the predefined abandoned cart tactic, we were able to swiftly set up the basics. And then by enriching the campaign with external content. We found a seamless method to technically connect Emarsys with the text engine. For the text engine. We involved another value party, Retresco. We've collaborated with Retresco over the past several years, particularly in the product data field, where we managed to generate product descriptions for thousands of different products based on attribution models. Applying with Retresco in the context of marketing texts was kind of a novel idea, but it totally made sense since we needed to organize complex text structures, consume personalization in the attribution model, and utilize natural language generation by large language models to produce dynamic content. And this hybrid model enabled dynamic content to be saved and optimized by an editor. Lastly, we could generate outputs for different languages with the simple click and finalize everything with the colleagues in the different markets for the final content editing. All right, enough theory. Let's jump into the demo and show you hands on how all this looks and works. We will start on Emarsys, but the process for handling abandoned carts couldn't be simpler. Utilizing the Emarsys tactics, here we have ready to use scenarios of various strategies. In this instance, let us focus on the abandoned carts. Of course, we can set up this tactic for the email channel or in combination with multiple channels. For today, we'll really focus on email, although the mechanics would work just as well for any other channel. Okay, so here we are in the abandoned cart automation that allows us to respond to a user's e-commerce behavior. When an abandoned cart is registered. We filter out contacts who made a purchase within the last three days to only target the most relevant customers. Then we send out the campaign, followed by a reminder after two days for those who did not make a purchase. Simple enough right? Let's jump into the email template real quick. So here we have added custom blocks containing placeholders for the external content integration between Emarsys and Retresco. This functionality provides an API connector, enabling us to send campaign and contact data to Retresco at the moment, the campaign is about to be sent out and receiving personalized message parts that can be incorporated into the campaign and subject line before the composed message is actually dispatched to the contact. This might sound a little technical, but for the marketeer and their day to day operations, making use of these mechanics is as simple as placing the block into the campaign and setting the configuration for the desired cartridge. In this case, we have a multi-language campaign and we can switch between the relevant language variants. Let's switch to English and choose the appropriate cartridge from Retresco with the correct language. As you can see, we can have cartridges per use case and language. Within the English version, you can also see a dynamic version, which we'll get back to in a second. We can also define additional individual configurations like this language type options. Currently Allnatura uses formal language in all their communications. That is very relevant for German, Dutch and French, where the formal and informal language differs quite a bit. But we also wanted to pave the way to have age based dynamic option. So now you can choose whether the language should remain formal, be informal in general, or automatically adjust based on the user's age using informal language for anyone younger than 40. All right, now that you've seen the set up in Emarsys, let's move into what's happening behind the scenes and how the text is actually generated. Now here in Retresco, the Text Engine, you can see that we have cartridges for the different variations you just saw in the drop down in the Emarsys template. Starting with a baseline, we began generating dynamic content. In this particular cartridge you can observe different message parts, and for each message part we can set up multiple templates based on the defined rules. The actual content is prompted by instructing the AI to write a personalized message, choosing the language style, informal or formal based on the rule, and utilizing the data fields provided by the external content call from Emarsys for gender, first name, last name. With this information, the greeting can be tailored to the appropriate language type. And now, looking at the introduction, it becomes even more intriguing. If specified categories are present, the prompt will consider these and compose a five line introductory text for an abandoned cart email, emphasizing engagement and motivation to increase the likelihood of completing the purchase. We also specified that the message represents sustainable furniture, and explicitly instructed that no discounts should be promised to the customer, as the AI at some point started promising discounts we didn't ask for in some variations. Okay, so let's quickly return to Emarsys and preview the capabilities of the engine so far. Back in our campaign for the English language, we will switch to the Dynamic English cartridge we just saw and preview the campaign for an English speaking client. Here you can see the polite address to Mr. Baldwin acknowledging their abandoned shopping basket with the context of the product category home textiles taken into account while sustainability is emphasized. The call to action button text is also dynamically generated with buy now. Now if we request a new preview. Let's see. We kind of have the same text. But now if you look at the call to action, you will notice that the label now says order now. Okay, let's request yet another version and see what happens. So now we have a totally new text, and we see that there is a new CTA and the text is more focused on the actual category and the cozy home. So, yeah, very interesting variations. However, we also recognize that while the accuracy and the quality of this content is quite high, it's not really sufficient to completely trust the AI to always produce the perfect content. For instance, looking back at the hallucinated discounts, you really saw that there are limitations. And this is where Retresco's hybrid model really shines. By jumping into the non dynamic cartridge, now you can see how we've set a more granular rule set up here explicitly for the formal and informal differentiation and major relevant categories, as well as the general rule for other categories. Within each message part and each template, we applied the same prompts as in the dynamic cartridge, but with more detail, as the content parts now have to be more accurately targeted. Looking into the greeting, we can quickly look into how the template is set up. Now, while we can set this template to be dynamic, in this instance, we just want to use the AI to generate text variants. And now I can either pick a type of text from the presets, or write my own prompt as a custom instruction. I can add data fields that I can take from Emarsys and lastly I can provide additional keywords. This is really what made creating the templates for the intro and the individual categories really easy. By altering the keyword for each category, we could generate multiple variations per category with just a click. In this case, looking into the living category, you can see how we generated content with high variance. But now we could add our own finishing touches to the saved content and optimize language, correct some inaccuracies, and once the editing was completed, the messages and templates are revised and in the ideal desired quality, we can go ahead and take these and easily translate them into any additional language like we did here for French and Dutch. Just for the sake of it. Let me show you how to set up the cartridge in Chinese. So let's see what happens if we create the translation for that complicated language. Now, while the translation is being generated, let's revisit where we really, really started from. Because as you may know, we and Allnatura are based in Germany. Hence our primary working language is of course, German. The original cartridges we wrote were actually in German, and these prompts were sufficient to create a dynamic English cartridge as well, based on the German prompts. So now if you see here, we still have the German prompts in the English language version, but the generated output is actually in English. And this greatly simplifies the editing process, allowing editors to prompt in their native language and still receive translated content in the correct language, ready for further editing and adjustments if necessary. Okay, so while the Chinese version has been generated. So let's jump into that. Great! So the whole cartridge is now translated and I can forward it to my Chinese editors for quality assurance over there. Perfect. So now let us stimulate an actual send out for multiple customers in different languages, demonstrating the versatility of the results received in the inbox. Okay, so let's click here. And now an abandonment cart event is being triggered for 19 exemplary customers activating the Emarsys automation. It pulls individual language content via external content from Retresco, providing the subtext of each customer. And then we can see the results coming into our inbox in all relevant languages, with individual subject line specific personalization based on formal or informal language, product categories, cart content, and even the abandoned products in the cart are visible here. The ease with which we created such an engaging scenario, with such a high variance and quality content is really impressive. Let us move to the conclusion here. And I mean, you remember we talked about we want to create a sustainable, KPI driven scenario. We, have established goals and we want to learn from the send outs, and we want to see what works well and what didn't. Our future vision even includes feeding the send out variances and the conversion rates back into the AI, and adding that to make better decisions and target, the contexts more efficiently in the future. However, we have to recognize that the current technical limitations are just there, and we cannot expect to achieve everything immediately. So being pragmatic as we are, we have set up an analytics structure to at least be able to monitor the different message paths generated through Retresco, and to see which combinations have been sent out, what revenue was generated with each campaign, with each contact in the campaign. And this allows us to identify the best performing template variations for each message. Ultimately, this information will enable editors to further optimize content or to create additional variations based on the most successful ones, or maybe even learn from the ones that didn't work well and not focus on them too much. So that concludes the demonstration. Let us return to the slides and see, where Emarsys is and what they have planned for the near future in terms of AI capabilities because, Philip, I know that you have lots in store. So. Yeah. Thank you. Yeah. So, like I said, we wanted to give you a glimpse on what's on, on our roadmap. Again, to recap the things that we are working on our generative AI and then also looking into conversational channel. Just a disclaimer, some of those things are currently available. Some are in pilot and also in development. All the things that I'm going to show you now are aimed to be released to Emarsys customers throughout this year. But again, with all the things in roadmaps and product prioritization, none of that can be set in stone. But we are always looking for more pilots and customers willing to try out things. So if you are here as an Emarsys customer and interested to work with us, let us know. Again when I said what are the benefits for marketers using AI, we always want to look into what marketers need and how AI can then offer something that can really bring additional value to it, and then also do it in a very smart way. So what Amir also showed is there are some nuances into what AI does and what it doesn't. But it's very important to understand that the AI in general will not make mistakes. The mistakes are coming from the human implementation that an AI made. So we've talked about prompt engineering in setting the right prompts. So it's very important that, while we still want to be efficient and focused on the ROI and giving, insights to the data, we also want to make sure that we are limiting the possibilities of mistakes in using the AI, so that the outcome then isn't what the marketer wants to see. And if we look overall into our AI vision, we really want to enable marketers to create a end to end intelligent customer journey. And this is focusing on two areas. First, again, the marketer productivity become faster, quicker, more efficient, but also, focusing on the hyper personalization that we at Emarsys really are striving for when it comes to the different channels that we are using and the personalization capabilities that we either bring out of the box or can work with partners, like we've seen in this example. But really from start to finish from using an AI assistance on the analytics part, then using, segmentation, optimizing the channels, optimizing the content from product recommendation to subject line to text to products to literally everything. Also then, working with predictive scoring and then an AI based preview like we've seen in the example from here, we really want to make sure that we are using AI where it makes sense. So the marketer can really have a complex use case and complex customer journey, build up in a very sophisticated but easy way without too complex steps in between. So just looking at some example and this is already part of the product. When we look at subject line generation, which seems like a very basic topic, it is important that we are also looking how can we make subject line generation, long term, relevant and efficient. So there is always a baseline instruction, but that will always be used from the smart AI within Emarsys for also repeated prompts. So you don't need to start fresh every time. You can actually reuse prompts that have been successful automatically in the next iteration of your next campaign to generate subject lines. And when we talk about subject lines, that's obviously text for emails, but it can also be applied to short text messages, SMS, push messages and the likes to use that as an omnichannel tool already. The other part is using the data from the product catalog for a more compelling personalization and text. So we've seen, what Amir showed outside of the Emarsys platform. We are also trying to work to get more of those capabilities into the platform in terms of the text generation, because the product catalog and the sales data are already here. So using that in terms of them creating more personalized content and text is the next step that we are working on and, want to make it easy for marketers to use. So one example, that we can see here how that could work is, if I want to, for instance, generate a specific text, I can get the AI to do that and then select the ones that I want. But I also can make changes based on the different things that I want to do, as well as also give text instructions. We will also provide those instructions and prompt examples as part of that solution in order to reduce the margin of human error, as part of prompt engineering and setting up the AI results. Another very exciting example is if you look at the initial use case that we had as a challenge, how can I quickly find the right products for a campaign based on the use case that I want to deliver? So in this case, if I want to find the top five selling products for women from the last month, I obviously then get a result from this example prompt, but also the AI is intelligent enough to already work with some suggestions on how that could be improved. So do I actually mean the numbers of purchases? Do I mean the month of January, or do I want to make it generic in terms of 30 to 31 one days? So those prompts can already be improved for the future use. And all I need to do is then click the products that I want, automatically put them into the template and be ready for execution. So the next step is from text. Then having full fledged content blocks from the content library available to be able to create emails in a couple of seconds and minutes. So, that just as a short sneak preview on things that we have been working on and are working on, more of that obviously to come. I've rushed a bit through it because we are really short on time and we want to spend enough time to get the questions answered, but maybe, Amir, just for you quickly, if I am a marketer right now and if I want to apply AI to my business, how can I get started right now? Maybe if you can, spend a couple of minutes to explain what you would recommend, and then we use the remaining time to go through the questions? Perfect. Yes, gladly Philip. Yeah. I think you've seen, I hope the technical issues were not too big and you were able to follow and understand me properly. But, I hope you saw how our approach actually works and how realistic it is that AI is actually usable in day to day operations. That's kind of the danger. It's a hype topic and people maybe misuse it and go the not ideal way and basically endanger their business. And we found, I think, a very, very good middle ground where we, on the one hand, actually use the, technical capabilities that are there right now. That we understand the limitations that we cannot maybe trust blindly into what is being generated. And that the quality is always 100% there, but we can, save quite some time. And, with that, we value the AI capabilities of Emarsys. Also, the things you just showed are so so great and are going to make things easier down the road. And yet, we see the value in having external integrations and using external content with Retresco, in this case to have a powerful engine to really manage the content, to really have rules to also allow the editor to understand what they are doing to be able to build complexity without it being overwhelming. And then, within the engine to make use of the large language model to make use of the AI. So, that's really valuable. But, at the same time, it's important we have a strong partner on your side that actually does the conceptual work. It's not easy to begin with. This whole set up to just come up with the scenarios to bring together the systems. It took quite some engineering to get there but together with our client, with our partners at Emarsys and Retresco, we were able to really build a very strong foundation and to really bring this to production in a reliable way and a future proof way. So I think, it all has to come together. And if you do it right, if you really, make conscious decisions, you can use AI right now, save time, bring in value and move forward. And with that, we are a bit over time. So I'd like to close that session with a big thank you. Thank you for dialing in. Thank you for the really interactive questions and making this not just a one way street. We will provide you the content and also the demo that we have done, maybe in a better quality way as a recap, but thank you really for leaning in and if you are interested in learning more about the AI capabilities or moving forward either to Amir or with Emarsys, let us know. Reach out to us. We will share the contact details. And would be really excited to hear from you and continue to work with you. With that, we all wish you a good day. Successful week and have a good day. Bye bye everyone. Thank you very much. Bye.